15.2.3 Tensor Symmetrization Area

Tensor Symmetrization Area is the branch of application in which the symmetrization operator itself, rather than the resulting symmetric tensors alone, is the tool of direct interest, used to extract a symmetric approximation from an imperfectly symmetric array, to enforce symmetry as a modeling or numerical constraint, and to build symmetric functions systematically in combinatorics and statistics.

Extracting a Symmetric Part From Imperfect Data

Correcting Numerical Asymmetry

A matrix produced by a numerical computation that should, in principle, be symmetric, such as a covariance estimate computed from finite sample data or a stiffness matrix assembled by a finite element method, often fails to be exactly symmetric due to rounding error; applying the symmetrization operator, replacing $A$ by $\frac{1}{2}(A+A^T)$, produces the nearest exactly symmetric matrix in the sense of minimizing the distance to $A$, a routine and widely used correction step before further processing.

Separating Physical From Numerical Antisymmetry

In continuum mechanics, a computed stress or strain array is symmetrized as a matter of course, since the underlying physical quantity is symmetric by the governing equations, and any antisymmetric component present in a numerical result is understood to be a computational artifact rather than physically meaningful; the symmetrization area here overlaps directly with, but is conceptually distinct from, the form area's treatment of stress tensors as already-symmetric objects.

Symmetrization as a Modeling Constraint

Enforcing Symmetry in Constructed Models

When building a model, such as a proposed similarity or covariance structure, from raw pairwise scores that need not themselves be symmetric, applying the symmetrization operator to the raw score matrix is a standard way of imposing the required symmetry before the result is used as an input to an algorithm, such as spectral clustering, that specifically requires a symmetric matrix.

Symmetrized Eigenvalue Problems

Numerical eigenvalue algorithms specialized for symmetric matrices are typically faster and more numerically stable than general eigenvalue algorithms; symmetrizing an approximately symmetric matrix before applying such an algorithm trades a small, controlled perturbation of the input for access to this more efficient and more stable computational machinery.

Symmetric Functions in Combinatorics

Elementary Symmetric Polynomials

The elementary symmetric polynomials, $e_k(x_1,\dots ,x_n)=\sum _{i_1<\dots <i_k}{x}_{i_1}\cdots x_{i_k}$, are themselves symmetric tensors under the identification of homogeneous polynomials with symmetric tensors, and Newton's identities relate them to the power sums $\sum _i{x}_i^k$, another family of symmetric polynomials built directly by summing, rather than averaging, over permutations of the indices.

Symmetrization as a Generating Technique

More generally, any function of several variables can be turned into a symmetric function by explicitly averaging over all permutations of its arguments, exactly the symmetrization operator applied at the level of functions rather than of tensors, a technique used throughout enumerative combinatorics to construct symmetric generating functions from asymmetric building blocks.

Statistics and U-Statistics

Symmetric Kernels and U-Statistics

A U-statistic is built from a kernel function of several sample points by averaging that kernel over every ordered selection of distinct sample points, which is precisely an application of the symmetrization operator to the kernel viewed as a tensor-like object on the sample; symmetrizing the kernel in advance, when it is not already symmetric, simplifies this construction and is standard practice in the theory of U-statistics.

Unbiasedness Considerations

Symmetrizing an estimator built from an asymmetric formula is also used to remove spurious dependence on the arbitrary order in which sample points are listed, ensuring the resulting statistic is a genuine function of the unordered sample rather than of a particular labeling of it, a property required for many standard asymptotic results in statistics to apply.