Cauchy-Schwarz inequality for positive semidefinite matrices

cauchy-schwarz-inequalitylinear algebramatrices

I have a vector space $V$ of finite dimension $n$ over the field $\mathbb R$. I define the following product
$$u\otimes v=\sum_{i=0}^\infty b_iu_iv_i^T$$
where $b_i\in\mathbb R^+$ and all $u_i$ and $v_i$ are elements of the vector space $V=\mathbb R^n$. In the context of this formula, we use $u$ (and $v$) to mean the infinite list of $u_i$ (and $v_i$). This product defines a matrix in $\mathbb R^{n\times n}$, given a choice of basis in $V$. In particular, $u\otimes u$ will be (symmetrical) positive semidefinite.

We can assume that $\sum b_i$ converges and that all $u_i$ and $v_i$ are bounded (the $L^2$ norm of each vector is bounded). That ensures that $u\otimes v$ converges.

If we choose $n=1$, the product $u\otimes v$ simplifies to a scalar product and we can easily show using Cauchy-Schwarz that:
$$(u\otimes u)-(u\otimes v)(v\otimes v)^{-1}(v\otimes u)\ge0$$

How would I prove a similar inequality in general, for any finite $n$? As $b_i>0$ for all $i$, we have both $u\otimes u$ and $v\otimes v$ (symmetrical) positive semidefinite matrices so the inequality should be understood as meaning that $(u\otimes u)-(u\otimes v)(v\otimes v)^{-1}(v\otimes u)$ is also (symmetrical) positive semidefinite.

Best Answer

This is true even if $v\otimes v$ is singular, provided that you replace $(v\otimes v)^{-1}$ by the Moore-Penrose pseudo-inverse $(v\otimes v)^+$. For each integer $k$, let $U_k=\pmatrix{\sqrt{b_1}u_1&\sqrt{b_2}u_2&\cdots&\sqrt{b_k}u_k}$ and define $V_k$ analogously. Then $$ U_kU_k^T-U_kV_k^T(V_kV_k^T)^+V_kU_k^T =U_k\left[I_k-V_k^T(V_kV_k^T)^+V_k\right]U_k^T.\tag{1} $$ Since $I_k-V_k^T(V_kV_k^T)^+V_k$ is an orthogonal projection, both sides on $(1)$ are positive semidefinite. Pass $k$ to the limit, we obtain $u\otimes u-(u\otimes v)(v\otimes v)^+(v\otimes u)\succeq0$.