15.2.4 Tensor Symmetric Basis Area

Tensor Symmetric Basis Area is the branch of application in which the explicit choice of a basis for a space of symmetric tensors, built from the monomial-style basis of symmetric products introduced earlier, is the practically decisive tool, underlying compact storage of symmetric data, multi-index bookkeeping in numerical approximation, and the parametrization of feasible sets in optimization over symmetric matrices.

Compact Storage Exploiting the Reduced Basis Count

Storing Only Independent Components

Because the basis of ${\text{Sym}}^k\left(V\right)$ consists of symmetric products indexed by nondecreasing sequences, a numerical implementation representing a symmetric tensor need only store one value per basis element, rather than one value per entry of the full tensor power; for a degree-two symmetric tensor on an $n$-dimensional space, this reduces storage from $n^2$ entries to $\frac{n(n+1)}{2}$ entries, exactly the number given by the dimension formula for $k=2$.

Index Maps Between Full and Reduced Representations

Practical implementations maintain an explicit bijection between the nondecreasing index sequences of the reduced basis and a single linear storage index, allowing entries of the symmetric tensor to be looked up or updated in constant time without reconstructing the full, redundant array; this is the direct computational analogue of the basis and dimension facts established at the level of the symmetric power's own basis.

Multi-Index Notation in Numerical Approximation

Taylor Expansions as Symmetric Tensors

The $k$-th order term of a multivariate Taylor expansion is built from the $k$-th derivative tensor of a function, which is symmetric because mixed partial derivatives commute under standard smoothness assumptions; representing this derivative tensor in the basis of symmetric products, rather than as a full array of partial derivatives, avoids redundantly storing or recomputing derivatives that are guaranteed equal by symmetry.

Multi-Index Shorthand

The standard multi-index notation $\alpha$, writing a monomial ${x_1}^{{\alpha }_1}\cdots {x_n}^{{\alpha }_n}$ using a single index $\alpha$ recording how many times each variable occurs, is precisely a compressed label for a basis element of ${\text{Sym}}^k\left(V\right)$, with $k=\left|\alpha \right|$ the total degree, and this notation is used pervasively in numerical analysis and in the theory of partial differential equations precisely because it matches the reduced symmetric basis rather than the full tensor index set.

Parametrization in Semidefinite Optimization

The Basis of Symmetric Matrices as a Search Space

In semidefinite programming, the decision variable is a symmetric matrix constrained to be positive semidefinite, and the underlying vector space of symmetric matrices, of dimension $\frac{n(n+1)}{2}$, is parametrized using exactly the basis of symmetric products of the standard basis vectors, giving each independent entry of the matrix its own coordinate in the optimization problem.

Consequence for Problem Size

The reduced dimension of this basis, rather than the full $n^2$-dimensional space of all matrices, directly determines the number of scalar variables an interior-point or other numerical solver must track, so the symmetric basis area has a direct, quantitative impact on the computational cost of solving semidefinite programs of a given matrix size.

Basis Choice and Numerical Conditioning

Orthonormal Symmetric Bases

When $V$ carries an inner product, the basis of symmetric products of an orthonormal basis is not itself orthonormal without an additional scaling factor, since a symmetric product such as $e_1\odot e_2$ has a different induced norm than $e_1\odot e_1$; correcting for this discrepancy with an explicit normalizing constant, depending on how many repeated indices occur, is a standard practical step whenever a genuinely orthonormal basis for a symmetric power is needed, such as in orthogonal polynomial constructions built from symmetric tensor bases.