Abstract:In this paper we propose a new generalized cyclic symmetric structure in the factor matrices of polyadic decompositions of matrix multiplication tensors for non-square matrix multiplication to reduce the number of variables in the optimization problem and in this way improve the convergence.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is to reduce the number of variables in the multilinear optimization problem of non - square matrix multiplication tensors, thereby improving convergence. Specifically, the authors propose a new Generalized Cyclic Symmetric (GCS) structure to be applied in the multilinear decomposition of matrix multiplication tensors. By introducing this structure, the number of variables in the optimization problem can be reduced, thereby increasing the solution efficiency and convergence speed.
### Problem Background
In the Fast Matrix Multiplication (FMM) problem, matrix multiplication can be represented by a tensor equation:
\[ C = \sum_{i,j,k = 1}^{m,p,n} E_{ik}^{m\times p}\otimes E_{kj}^{p\times n}\otimes E_{ij}^{m\times n}\cdot F(A,B) \]
where \( E_{ij}^{m\times p} \) is a basis matrix in \( R^{m\times p} \), satisfying \( E_{ij}^{m\times p}(i_1,i_2) = 1 \) if and only if \( (i_1,i_2)=(i,j) \), and 0 otherwise. Through this method, matrix multiplication can be represented as the inner product form of a third - order tensor.
### Proposed Method
To reduce the number of variables in the optimization problem, the authors propose the Generalized Cyclic Symmetric (GCS) structure. This structure is still applicable in the case of non - square matrices and can significantly reduce the number of variables in the optimization problem. Specifically, the GCS structure is defined in the following ways:
1. **Generalized Cyclic Symmetry**: For any \( i_1,i_2,j_1,j_2,k_1,k_2\leq m \), satisfy:
\[ T_{mpn}(i_1+(i_2 - 1)m,j_1+(j_2 - 1)p,k_1+(k_2 - 1)n)=T_{mpn}(k_1+(k_2 - 1)m,i_1+(i_2 - 1)p,j_1+(j_2 - 1)n)=T_{mpn}(j_1+(j_2 - 1)m,k_1+(k_2 - 1)p,k_1+(k_2 - 1)n) \]
2. **Decomposition Structure**: Decompose the factor matrices \( U,V,W \) into symmetric and asymmetric parts, as follows:
\[
U = [A,B,C,D],\quad V = [A,D,B,C],\quad W = [A,C,D,B]
\]
where \( A,B,C,D \) represent the symmetric and asymmetric parts respectively.
### Experimental Results
Through numerical experiments, the authors show that using the Generalized Cyclic Symmetric (GCS) structure can effectively reduce the number of variables and improve the solution efficiency. For example, in the process of finding \( PD_{11}(T_{223}) \) and \( PD_{14}(T_{224}) \), after using the GCS structure, more accurate and practical solutions are obtained.
### Summary
The main contribution of this paper is to propose a new Generalized Cyclic Symmetric (GCS) structure for reducing the number of variables in the multilinear optimization problem of non - square matrix multiplication tensors, thereby improving the solution efficiency and convergence speed. The effectiveness of this method is verified through numerical experiments.