Member eigenvalues?

Given a real symmetric matrix $H \in \mathbb{R}^{n\times n}$, we can show that:

\[\forall \mathbf{v} \in \mathbb{R}^n, \|\mathbf{v}\|=1 \Rightarrow \mathbf{v}^\top H \mathbf{v} \leq \lambda_{max}\]

where $\lambda_{max}$ denotes the largest eigenvalue of $H$.

Proof

First of all, we use the eigendecomposition of a real symmetric value, e.g. $H = Q \Lambda Q^\top $ and introduce $\mathbf{y} = Q^\top \mathbf{v}$. Thus, we have:

\[\begin{align*} \mathbf{v}^\top H \mathbf{v} & = \mathbf{y}^\top \Lambda \mathbf{y} \\ & = \lambda_{max} y_1^2 + \lambda_2 y_2^2 + \ldots + \lambda_n y_n^2 && \Lambda \text{ is a diagonal matrix} \\ & \leq \lambda_{max} (y_1^2 + y_2^2 + \ldots + y_n^2) \\ & = \lambda_{max} \mathbf{y}^\top \mathbf{y} \\ & = \lambda_{max} \mathbf{v}^\top Q Q^\top \mathbf{v} \\ & = \lambda_{max} \mathbf{v}^\top \mathbf{v} && Q Q^\top=I \\ & = \lambda_{max} \|\mathbf{v} \| \\ & = \lambda_{max} && \| \mathbf{v} \| = 1 \end{align*}\]

Yayyy, we are done!

Note: inverse of a symmetric matrix

What shown above is true also for the inverse of a symmetric matrix, since it is also symmetric. Indeed, given A nonsingular and symmetric:

\[\begin{align*} I & = I^\top \\ AA^{-1} &= (AA^{-1})^\top && AA^{-1}=I \\ AA^{-1} &= (A^{-1})^\top A^\top && (AB)^\top = B^\top A^\top \\ A^{-1} A &= (A^{-1})^\top A^\top && AA^{-1} = A^{-1} A = I \\ A^{-1} A &= (A^{-1})^\top A && A^\top = A \\ A^{-1} A A^{-1} &= (A^{-1})^\top A A^{-1} && \text{multiply both sides by } A^{-1} \\ A^{-1} &= (A^{-1})^\top \end{align*}\]

Yayyy, we are done again!

Written on May 27, 2018