Distance invariant method for normalization of indexed differentials☆
Introduction
Symbolic manipulation of indexed expressions, e.g. tensor expressions, is one of the oldest research topics in computer algebra (Fulling et al., 1992; Christensen and Parker, 1994; Ilyin and Kryukov, 1996; Portugal, 1998, Portugal, 1999, Portugal, 2000; Manssur et al., 2002; Martín-García et al., 2007, Martín-García et al., 2008; Liu et al., 2013). It remains a challenging problem for the following reasons: In order to compute the canonical form of an indexed polynomial, we need to find a finite Gröbner basis for the ideal generated by the basic syzygies. Unfortunately, the ideal is not finitely generated, mainly due to the property of renaming dummy index. Various efforts have been made to develop algorithms for simplifying tensor expressions. Portugal, 1998, Portugal, 1999 presented algorithms to normalize tensor expressions with respect to monoterm symmetries and cyclic symmetry. However, those algorithms are unable to deal with indexed differential expressions, because index elimination is indispensable for simplifying them.
A natural problem in symbolic computation is: Given two indexed differential expressions, how can we judge whether they are equal or not (what is the normalization algorithm)?
In differential geometry, massive calculation involving indexed differential expressions arises from tensor verification problem and the problem of finding transformation rules of indexed functions under coordinate transformation. Therefore, normalization algorithms are indispensable, as evinced by the following problems.
Problem 1.1 The curvature tensor determined by the connection on differentiable manifold is an indexed function defined by Show that R is a tensor, i.e., it satisfies the following transformation rule of ()-typed tensor:
Problem 1.2 A Riemannian manifold admits an almost complex structure . Let be 1/8 times the Nijenhuis tensor of , i.e., The “torsional derivative” of a tensor field on is defined in (Willmore, 1993) as follows: where has the following “rather strange” definition: What are the transformation rules of and under coordinate transformations?
Liu et al. (2009) presented a normalization algorithm for polynomials in , which consists of two parts. In the first part, a polynomial is rewritten modulo monoterm symmetries. In the second part, we compute the canonical form of a polynomial by collecting equivalent terms with respect to monoterm symmetries. Then, two polynomials are equal if and only if they have the same canonical forms. By using this algorithm, one can prove that and in Problem 1.1, Problem 1.2 are tensors, and can derive that has the following transformation rule:
However, the algorithm with respect to monoterm symmetries in (Liu et al., 2009) has a factorial complexity. More precisely, suppose that f is a monomial indexed with i, j, , , where i is the number of functions, j is the number of pairs of dummy indices, is the number of functions of the same name without free indices, and is the number of functions with the same free indices and the same name. Since we need to compare all the monomials equivalent to f, the complexity is at least , where k is the number of groups of pairwise interchangeable indices, and () is the number of indices in each group. Other representative methods for computing canonical forms with respect to monoterm symmetries are due to Portugal (1999) and Manssur et al. (2002). The algorithm in (Portugal, 1999) has four steps. First, list all possible permutations according to commutativity of multiplication (symmetry under the interchange of functions). Second, for each permutation, replace the indices with integers according to the name of tensors, the index positions and index classes. Third, sort the indices inside each function. At last, find the smallest element in the equivalence class. So the complexity is at least . Computational group theory is applied in (Manssur et al., 2002). Its complexity for some special Riemann indexed monomials (with all indices contracted) is at least , where is the number of free indices.
The reason for such factorial complexity in some of these algorithms is that dummy indices within an indexed polynomial f are described either by original letters or by integers according to the order of their appearance. Neither of the descriptions is an invariant. In other words, they vary when we rewrite f. Consequently, to find the canonical form, we need to list and compare all the elements in the equivalence class of f.
At this point two questions arise: (1) What is an invariant?, and (2) how can we provide an efficient normalization algorithm with respect to monoterm symmetries of the invariant?
Besides, the normalization algorithm in (Liu et al., 2009) depends on a skillful and tricky classification of 2nd-order partial differential functions according to their connections (e.g. circle, chain, maximum lower tree and so on). But the classification is no longer valid for higher order. It appears infeasible to classify partial differential functions for each order.
Then another question arises: How can we provide a normalization algorithm that is independent of function classifications?
In (Liu, 2017), we only partly answered these questions based on function distances, as the proofs of validness of some algorithms are not given. Besides, the complexity of the normalization algorithm with respect to monoterm symmetries in (Liu, 2017) is at least cubic, higher than the one in this paper, as we will see below.
In this paper, we first define the distance from free indices to dummy indices, and prove that it is an invariant (with respect to certain symmetries). Then, we present an index replacement algorithm which is uniquely determined by the distance, and use it to develop a normalization algorithm with respect to monoterm symmetries for polynomials in . The complexity is at most , lower than known algorithms, where m is the sum of the number of free indices and the number of pairs of dummy indices. Finally, by the method of index replacement, and by choosing the monomial associated with the smallest numerical list as the canonical form, a normalization algorithm is provided for polynomials in , which is independent of function classifications.
Section snippets
Indexed differential polynomial ring
In this section we briefly review some notions presented in (Liu et al., 2009).
An indexed function is composed of four parts: a function name, a sequence of upper indices, a sequence of lower indices, and variables. For example, the Christoffel symbol has the function name Γ, upper index i, lower indices , and variables x. In this paper, all indexed functions are scalar-valued, and the base field is either or . An indexed monomial is the product of indexed functions. The
Distance and index-structure directed graph
In this section, we define the distance from a free index to a dummy index, and prove that it is an invariant with respect to monoterm symmetries and BS.
Definition 3.1 Suppose . Take free indices with identical names as different elements, and let V be the set of indices of f. Let be a directed graph, where A is the set of directed arcs. Let with . There is an arc in A from to if and only if is a lower index, is an upper index, and both indices belong to the same
Index replacement and normalization with respect to monoterm symmetries
Since (1), (2), (3) and (4) are all binomials, the problem of normalization of polynomials in is equivalent to the one in (as mentioned in (Liu et al., 2009)). In Sections 4 and 5, we only need to consider monomials in .
In this section, we first present an index replacement algorithm using the distance invariant. Then, by index replacement, we develop a normalization algorithm with respect to monoterm symmetries. The complexity is shown to be at most .
Notation 4.1 We will denote the
Normalization
In this section, we present a normalization algorithm for monomials in by using the method of index replacement.
According to Proposition 4.1 in (Liu et al., 2009), suppose and are equivalent, if we rewrite as () successively by elimination, mixed rewriting of unmixed interior circle, and coordinate system unification of self-restrained dummy indices, then . Hence, in order to find the canonical form of any monomial, we only need to develop a normalization
Conclusions
We have presented an index distance method for normalization of indexed differentials, and developed an efficient algorithm. The algorithm enables us to determine the equivalence of indexed differential polynomials in . One of its core parts is a normalization algorithm with respect to monoterm symmetries, whose complexity is shown to be lower than known algorithms.
Although the index distance in this paper is defined for indexed differentials, the idea of distance invariant can be used to
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
References (15)
- et al.
ATENSOR–REDUCE program for tensor simplification
Comput. Phys. Commun.
(1996) - et al.
The Invar tensor package: differential invariants of Riemann
Comput. Phys. Commun.
(2008) - et al.
The Invar tensor package
Comput. Phys. Commun.
(2007) An algorithm to simplify tensor expressions
Comput. Phys. Commun.
(1998)The Riegeom package: abstract tensor calculation
Comput. Phys. Commun.
(2000)- et al.
MathTensor, A System for Performing Tensor Analysis by Computer
(1994) Graph Theory
(1997)
Cited by (1)
Normalization of Indexed Differentials by Extending Gröbner Basis Theory
2022, Journal of Systems Science and Complexity