6
$\begingroup$

Let $V$ be a finite dimensional Euclidean space and let $W_1,W_2$ be two subspaces of $V$. Let $P_1,P_2$ denote the projections onto $W_1,W_2$ respectively. Is it true that the composition $P_1\circ P_2$ is always diagonalizable?

$\endgroup$
7
$\begingroup$

Orthogonal projections are Hermitian positive semi-definite (with eigenvalues $0$ and/or $1$). The product of two positive semi-definite matrices (operators) is diagonalisable (see, e.g., Corollary 7.6.2 here -- unfortunately, it is not part of the preview). Therefore, the product of two orthogonal projections is diagonalisable.

The proof is based on the fact that given two positive semi-definite matrices $A$ and $B$ such that $r=\mathrm{rank}(A)$, there is a nonsingular matrix $X$ such that $$\tag{$*$} A=X\begin{bmatrix}I_r & 0 \\ 0 & 0\end{bmatrix}X^*, \quad B=X^{-*}DX^{-1}, $$ where $D$ is diagonal (and has nonnegative diagonal entries). Then it follows that $$ AB=X\left(\underbrace{\begin{bmatrix}I_r & 0 \\ 0 & 0\end{bmatrix}D}_{\text{diagonal matrix}}\right)X^{-1}. $$


The proof of ($*$) is not hard but rather technical.

Case $A$ is definite: It is easy to show that $AB$ is diagonalisable if $A$ and $B$ are Hermitian positive semi-definite and at least one of them (say, $A$) is positive definite. To show that we can consider the Cholesky factorisation of $A$: $A=LL^*$ for some nonsingular triangular matrix $L$. Then, since $L^*BL$ is positive semi-definite, we can unitarily diagonalise it as $$ L^*BL=UDU^*, $$ where $U$ is unitary and $D$ has non-negative diagonal. Now set $X=LU$ and see that $$\tag{1} A=LL^*=LUU^*L=XX^*, \quad B=L^{-*}UDU^*L^{-1}=X^{-*}DX^{-1}. $$

Both $A$ and $B$ semi-definite: The case that both $A$ and $B$ are semi-definite is a bit more complicated but proceeds essentially in the same fashion. Assume that $r=\mathrm{rank}(A)$. There exists a nonsingular matrix $L$ (does not need to be triangular) such that $$\tag{2} A=L\begin{bmatrix}I_r&0\\0&0\end{bmatrix}L^*. $$ This decomposition can be obtained, e.g., using the eigen-decomposition of $A$ and by "absorbing" the diagonal matrix containing the eigenvalues of $A$ to the unitary factors. Note that we cannot proceed directly as in the previous case because for a unitary $U$, we do not have in general $$ U\begin{bmatrix}I_r&0\\0&0\end{bmatrix}U^*\neq \begin{bmatrix}I_r&0\\0&0\end{bmatrix}, $$ which was used to show that $A=XX^*$ before; see (1). So we must find a bit more fancy diagonalisation of $L^*BL$ which will leave the diagonal matrix in (2) untouched by multiplication from both sides.

So, consider the matrix $$ L^*BL=\left[\begin{array}{cc}B_{11}&B_{12}\\B_{21}&B_{22}\end{array}\right]\begin{array}{l}\}r\\\}n-r\end{array}. $$ We can use the fact that (since $L^*BL$ is positive semi-definite) $\mathcal{N}(B_{11})\subset\mathcal{N}(B_{12}^*)$ (which is easy to show using the fact that if $x^*Cx=0$ for a semi-definite $C$ and a vector $x$, then $Cx=0$; it is in fact a generalisation of the property of semi-definite matrices which states that a zero on the diagonal implies zeros in the corresponding row and column), which implies (due to the relation between fundamental spaces) that $\mathcal{R}(B_{12})\subset \mathcal{R}(B_{11})$, and hence there is a matrix $Z\in\mathbb{C}^{r\times(n-r)}$ such that $B_{12}=B_{11}Z$. It follows that $$ L^*BL=\begin{bmatrix}B_{11}&B_{11}Z\\Z^*B_{11}&B_{22}\end{bmatrix} =\begin{bmatrix}I&0\\Z^*&I\end{bmatrix}\begin{bmatrix}B_{11}&0\\0&B_{22}-Z^*B_{11}Z\end{bmatrix}\begin{bmatrix}I&Z\\0&I\end{bmatrix}. $$ Now both diagonal blocks on the right-hand side are semi-definite (since we have a congruent factorisation) so there are unitary matrices $U_1$ and $U_2$ and diagonal positive semi-definite matrices $D_1$ and $D_2$ such that $$ B_{11}=U_1D_1U_1^*, \quad B_{22}-Z^*B_{11}Z=U_2D_2U_2^*. $$ The transformation matrix $X$ we look for is then given by $$ X=L\begin{bmatrix}I&-Z\\0&I\end{bmatrix}\begin{bmatrix}U_1&0\\0&U_2\end{bmatrix}. $$ With $D=\mathrm{diag}(D_1,D_2)$, it is easy to see that $B=X^{-*}DX^{-1}$. Indeed, $$ \begin{split} B &= L^{-*} \begin{bmatrix} I & 0 \\ Z^* & I \end{bmatrix} \begin{bmatrix} B_{11} & 0 \\ 0 & B_{22}-Z^*B_{11}Z \end{bmatrix} \begin{bmatrix} I & Z \\ 0 & I \end{bmatrix} L^{-1}\\ &= \underbrace{L^{-*} \begin{bmatrix} I & 0 \\ Z^* & I \end{bmatrix} \begin{bmatrix} U_1 & 0 \\ 0 & U_2 \end{bmatrix}}_{X^{-*}} \begin{bmatrix} D_1 & 0 \\ 0 & D_2 \end{bmatrix} \underbrace{\begin{bmatrix} U_1 & 0 \\ 0 & U_2 \end{bmatrix}^* \begin{bmatrix} I & Z \\ 0 & I \end{bmatrix} L^{-1}}_{X^{-1}}\\ &=X^{-*}DX^{-1}. \end{split} $$ Also, $$ \begin{split} X\begin{bmatrix}I_r&0\\0&0\end{bmatrix}X^* &= L\begin{bmatrix}I&-Z\\0&I\end{bmatrix}\begin{bmatrix}U_1&0\\0&U_2\end{bmatrix} \begin{bmatrix}I_r&0\\0&0\end{bmatrix} \begin{bmatrix}U_1&0\\0&U_2\end{bmatrix}\begin{bmatrix}I&0\\-Z^*&I\end{bmatrix}L^*\\ &= L \begin{bmatrix}I_r&0\\0&0\end{bmatrix} L^*\\ &=A. \end{split} $$


Note that the product of two oblique projections need not to be diagonalisable, since for $$ P_1 = \begin{bmatrix} 1 & 1 \\ 0 & 0 \end{bmatrix} \quad\text{and}\quad P_2 = \begin{bmatrix} 0 &1 \\ 0&1 \end{bmatrix}, $$ the product $$ P_1P_2=\begin{bmatrix} 0 & 2 \\ 0 & 0 \end{bmatrix} $$ cannot be diagonalised.

| cite | improve this answer | |
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.