On Product and Difference of Orthogonal Projectors

03 Jun 2020

I spent some time studying the conditioning of the linear least squares problem. As part of my efforts I learned a few facts about the products of orthogonal projectors and their difference. This post sums up the facts I learned and gives proofs for them. It was not easy for me to extract the proofs from the literature so once I had them in a satisfactory shape, I decided to write them up and post them. Hopefully someone else finds it useful.

The theorems were taken from “On the Perturbation of Pseudo-Inverses, Projections and Linear Least Squares Problems” by Stewart and the lemma was taken from one of its references “Perturbation Theory for Pseudo-Inverses” by Wedin. The proofs were put together from the references in the Stewart’s work [1] and supported by classical Linear Algebra text books. Full set of references I used can be found at the end of my write-up.

Here I will just put the main theorems proven in the document. For the proofs click here:

Theorem 1. (Theorem 2.3, part 1 in [1]) For any matrix $A$ and $B$ the following statement is true:

Let $P_A$ be orthogonal projector onto the range of $A$ and let $P_B$ be orthogonal projector onto the range of $B$. If $rank(A) = rank(B)$:

1. The singular values of $P_AP_B^\bot$ and $P_BP_A^\bot$ are the same.

2. The nonzero singular values $\sigma_i$ of $P_AP_B^\bot$ correspond to pairs $\pm\sigma_i$ of eigenvalues of $P_B-P_A$, so that: \(\begin{eqnarray} \nonumber \\ \|P_B-P_A\|_2=\|P_AP_B^\bot\|_2=\|P_BP_A^\bot\|_2 \end{eqnarray}\)

Theorem 2. (Theorem 2.3, part 2 in [1]) For any two matrices $A_{m \times n}$ and $B_{m \times l}$ and two orthogonal projectors $P_A$ and $P_B$ onto the ranges of $A$ and $B$ respectively we have: \(\begin{eqnarray} \nonumber \\ \|P_B-P_A\|_2<1\Rightarrow rank(A)=rank(B) \end{eqnarray}\)

Lemma 1. (Lemma 7.1 in [2]) Let $X$ and $Y$ be subspaces of $\mathbb C^n$ and $P_X$, $P_Y$ orthogonal projections onto $X$ and $Y$, respectively. $\sigma_i$ is a singular value of $P_XP_Y$ if and only if $\sigma_i^2$ is an eigenvalue of $P_XP_Y$.

References

[1] Stewart, G.W. On the Perturbation of Pseudo-Inverses, Projections and Linear Least Squares Problems, SIAM Review, Vol. 19, No. 4, 634-662 (1977).

[2] Wedin, P. Perturbation theory for pseudo-inverses. BIT 13, 217–232 (1973).