NLA MT25, Weyl's inequality


Flashcards

@State Weyl’s inequality.

Suppose:

  • $A$ is a matrix
  • $E$ is another matrix (a perturbation)

Then for all $i$, $ \vert \sigma _ i(A + E) - \sigma _ i(A) \vert \le \|E\| _ 2$. As other cases,

  • $ \vert \vert A \vert \vert _ 2 - \vert \vert E \vert \vert _ 2 \le \vert \vert A + E \vert \vert _ 2 \le \vert \vert A \vert \vert _ 2 + \vert \vert E \vert \vert _ 2$,
  • For the eigenvalues of a symmetric matrix $A$, for all $i$, $ \vert \lambda _ i (A + E) - \lambda _ i (A) \vert \le \|E\| _ 2$.

@Prove ∆weyls-inequality, i.e. that if

  • $A$ is a matrix
  • $E$ is another matrix (a perturbation)

Then for all $i$, $ \vert \sigma _ i(A + E) - \sigma _ i(A) \vert \le \|E\| _ 2$. As other cases,

  • $ \vert \vert A \vert \vert _ 2 - \vert \vert E \vert \vert _ 2 \le \vert \vert A + E \vert \vert _ 2 \le \vert \vert A \vert \vert _ 2 + \vert \vert E \vert \vert _ 2$
  • For the eigenvalues of a symmetric matrix $A$, for all $i$, $\lambda _ i (A + E) \in \lambda _ i (A) + [- \vert \vert E \vert \vert _ 2, \vert \vert E \vert \vert _ 2]$

We apply ∆courant-fischer-minmax-theorem to $A + E$. For every fixed subspace $\mathcal S$ with $\dim \mathcal S = i$ and every $x \in \mathcal S \setminus \{0\}$, the reverse triangle inequality gives the bound

\[\frac{\|(A + E)x\| _ 2}{\|x\| _ 2} \le \frac{\|Ax\| _ 2}{\|x\| _ 2} + \frac{\|Ex\| _ 2}{\|x\| _ 2} \le \frac{\|Ax\| _ 2}{\|x\| _ 2} + \|E\| _ 2\]

Then, taking $\min _ {x \in \mathcal S}$ on both sides,

\[\min _ {x \in \mathcal S \setminus \{0\}} \frac{\|(A + E)x\| _ 2}{\|x\| _ 2} \le \min _ {x \in \mathcal S \setminus \{0\}} \frac{\|Ax\| _ 2}{\|x\| _ 2} + \|E\| _ 2\]

This holds for every $\mathcal S$, so we may take $\max _ {\dim \mathcal S = i}$ on both sides to yield

\[\begin{aligned} \sigma _ i(A + E) &= \max _ {\dim \mathcal S = i} \min _ {x \in \mathcal S \setminus \{0\}} \frac{\|(A + E)x\| _ 2}{\|x\| _ 2} \\ &\le \max _ {\dim \mathcal S = i} \min _ {x \in \mathcal S \setminus \{0\}} \frac{\|Ax\| _ 2}{\|x\| _ 2} + \|E\| _ 2 \\ &= \sigma _ i(A) + \|E\| _ 2. \end{aligned}\]

Swapping the roles of $A$ and $A + E$, we have the matching lower bound

\[\sigma _ i(A + E) \ge \sigma _ i(A) - \|E\| _ 2,\]

so

\[ \vert \sigma _ i(A + E) - \sigma _ i(A) \vert \le \|E\| _ 2\]

For the first special case, take $i = 1$. Then $\sigma _ 1(M) = \|M\| _ 2$, so

\[\big \vert \|A + E\| _ 2 - \|A\| _ 2\big \vert \le \|E\| _ 2\]

which rearranges to

\[\|A\| _ 2 - \|E\| _ 2 \le \|A + E\| _ 2 \le \|A\| _ 2 + \|E\| _ 2.\]

For the second special case, we replace the singular value Courant-Fischer bound with the eigenvalue one, to yield

\[\frac{x^\top (A + E) x}{x^\top x} \le \frac{x^\top A x}{x^\top x} + \frac{ \vert x^\top E x \vert }{x^\top x} \le \frac{x^\top A x}{x^\top x} + \|E\| _ 2\]

and then follow the rest of the proof.

Bite-sized

What does Weyl’s inequality tell us about conditioning?

Singular values of any matrix and eigenvalues of symmetric matrices are well-conditioned under perturbations: a perturbation of size $\|E\| _ 2$ moves each $\sigma _ i$ or symmetric $\lambda _ i$ by at most $\|E\| _ 2$ — additive, not multiplicative. So a backward-stable algorithm can compute these quantities to essentially working precision $O(u)$, with no $\kappa$-amplification.

Source Lecture 4, §4.1 of the lecture notes (post-Theorem-4.2 discussion).

@bite~

Weyl’s inequality follows directly from the Courant-Fischer minimax characterisation combined with the reverse triangle inequality $\ \vert (A+E)x\ \vert _ 2 \le \ \vert Ax\ \vert _ 2 + \ \vert Ex\ \vert _ 2$, applied inside the min-over-subspace.

Source: Lecture 4, §4.1 of the lecture notes (the proof of Theorem 4.2 is left as an exercise via this strategy).

@bite~

Contrast Weyl’s bound (additive in $\ \vert E\ \vert _ 2$ for symmetric eigenvalues and singular values) with what happens for non-symmetric eigenvalues: the latter can be sensitive by a factor of $\epsilon^{1/n}$ for the worst Jordan-block case, demonstrating that non-symmetric eigenvalues are not well-conditioned in general.

Source: Lecture 4, §4.1.1 of the lecture notes (“Eigenvalues of nonsymmetric matrices are sensitive to perturbation”).

@bite~