Hristo S. Sendov
Abstract: In this work we propose a generalization of the Hadamard product between two matrices to a tensor-valued, multi-linear product between k matrices for any $k \ge 1$. A multi-linear dual operator to the generalized Hadamard product is presented. It is a natural generalization of the Diag x operator, that maps a vector $x \in \R^n$ into the diagonal matrix with x on its main diagonal. Defining an action of the $n \times n$ orthogonal matrices on the space of k-dimensional tensors, we investigate its interactions with the generalized Hadamard product and its dual. The research is motivated, as illustrated throughout the paper, by the apparent suitability of this language to describe the higher-order derivatives of spectral functions and the tools needed to compute them. For more on the later we refer the reader to [14] and [15], where we use the language and properties developed here to study the higher-order derivatives of spectral functions.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is the research on the high - order derivative calculation and properties of spectral functions. Specifically, the author proposes the concept of a generalized Hadamard product, generalizes it from an operation between matrices to a tensor - valued multilinear operation among multiple matrices, and studies the mutual relationship between this generalized product and the action of the orthogonal group. These studies aim to provide a new language and tool for describing and calculating the high - order derivatives of spectral functions.
### Specific background of the problem
A spectral function refers to a function \(F: S_n \to \mathbb{R}\) that is defined on the space of symmetric matrices \(S_n\) and is invariant under the action of the orthogonal group. Such a function can be expressed as the composition of a symmetric function \(f: \mathbb{R}^n \to \mathbb{R}\) and an eigenvalue mapping \(\lambda: S_n \to \mathbb{R}^n\):
\[F(X) = f(\lambda(X))\]
where \(\lambda(X) = (\lambda_1(X), \dots, \lambda_n(X))\) is the eigenvalue vector of the matrix \(X\), and satisfies \(\lambda_1(X) \geq \cdots \geq \lambda_n(X)\).
### Research motivation
The research on spectral functions has a wide range of applications, including matrix perturbation theory, eigenvalue optimization, and quantum mechanics. In particular, the smoothness and differentiability properties of spectral functions are crucial in these applications. Existing research shows that a spectral function \(F\) has a certain property (such as continuous differentiability, strict differentiability, etc.) at a point \(X\) if and only if the corresponding symmetric function \(f\) has the same property at \(\lambda(X)\). Therefore, studying the high - order derivatives of spectral functions is very important for understanding their properties and practical applications.
### Main contributions of the paper
In order to better describe and calculate the high - order derivatives of spectral functions, the author introduces the concept of a generalized Hadamard product and defines a multilinear operator dual to it. In this way, the author is able to express the high - order derivative of a spectral function as a tensor - valued mapping \(A_\sigma\) and the form of a generalized Hadamard product:
\[ \nabla^k (f \circ \lambda)(X)[E_1, \dots, E_k] = \sum_{\sigma \in P_k} \langle A_\sigma(\lambda(X)), \tilde{E}_1 \circ_\sigma \cdots \circ_\sigma \tilde{E}_k \rangle \]
where \(\tilde{E}_i = V^T E_i V\), and \(V\) is an orthogonal matrix such that \(X = V \text{Diag}(\lambda(X)) V^T\).
### Conclusion
By introducing the generalized Hadamard product and the corresponding multilinear algebraic tools, this paper provides a new description method for the high - order derivatives of spectral functions. This not only simplifies the calculation of high - order derivatives but also provides a theoretical basis for further research on the properties of spectral functions.