NLA Visualizations

Conditioning of LSE II

Under Construction

Sensitivity to matrix perturbations δA\delta A

Formulation

AxbA\mathbf{x} \approx \mathbf{b}
y^=A(AA)1Ab\mathbf{\hat{y}} = A(A^*A)^{-1}A^*\mathbf{b}
κAy^rcosθ\kappa_{A \mapsto \mathbf{\hat{y}}} \sim \frac{\|\mathbf{r}\|}{\cos \theta}

Core Idea

We fix b\mathbf{b} along the minor axis σn\sigma_n of AA to visualize the worst-case relative error.

  • No matter how the column space is tilted, the line by^\mathbf{b} - \mathbf{\hat{y}}' remains at a right angle to 0y^0 - \mathbf{\hat{y}}'.
  • Therefore, as Span(A)\text{Span}(A) is tilted towards b\mathbf{b}, y^\mathbf{\hat{y}}' moves strictly along a sphere of radius b/2\|\mathbf{b}\|/2 centered at b/2\mathbf{b}/2.

Geometric Controls

45°
0.40
0 (Rank-1)0.5 (Full)
||ŷ||
1.000
||r||
1.000
Click & drag to rotate camera