NLA Visualizations

Conditioning of Matrix II

Geometry of matrix perturbations

Formulation

κ(x)AAx=AxAx\kappa(\mathbf{x})_{A \mapsto A\mathbf{x}} = \frac{||A|| \cdot ||\mathbf{x}||}{||A\mathbf{x}||}
κ(A)=AA1\le \kappa(A) = ||A|| \cdot ||A^{-1}||

Core Idea

We fix x\mathbf{x} and explore perturbations to the matrix itself: AA+δAA \to A + \delta A.

  • We visualize a "worst-case" δA\delta A as a solid red ellipse tilted out of Span(A)\text{Span}(A), centered at the tip of AxA\mathbf{x}.
  • By implicitly choosing δA\delta A such that x\mathbf{x} aligns perfectly with its right singular vector, the perturbation δAx\delta A\mathbf{x} assumes its maximum possible magnitude and maps along the long axis of the local ellipse.
  • The dotted yellow ellipse shows the fully perturbed global mapping (A+δA)(A + \delta A) applied to the entire domain circle.

Geometric Controls

45°

Fixes the input vector x\mathbf{x} in the domain.

0.40
0 (Rank-1)0.5 (Full)
||Ax||
1.458
κ_{A ↦ Ax}(x)
1.372
Click & drag to rotate camera