Brownian Bridge™

Setup

Every matrix — whether square or rectangular, rank-deficient or full-rank — admits a singular value decomposition (SVD). This is a strictly stronger result than eigendecomposition, which applies only to square matrices and fails for non-diagonalisable ones. The SVD is the canonical tool for understanding the geometry of a linear map and the sensitivity of the linear system $Ax = b$ .

Where this lives on a desk. The SVD appears in three distinct quant workflows:

Calibration and least-squares fitting. When you fit a volatility surface or calibrate a yield curve model, you solve an overdetermined system $\min_\theta \|F(\theta) - \text{market}\|^2$ . The normal equations $A^\top A\theta = A^\top b$ are solved via the pseudoinverse $A^+ = V\Sigma^+ U^\top$ — the SVD-based answer that minimises the residual with minimum-norm $\theta$ .
Risk factor models. Returns matrices $X \in \mathbb{R}^{T \times n}$ are decomposed into orthogonal factors. SVD of $X$ gives the principal components directly without forming $X^\top X$ — numerically more stable when $T$ is small relative to $n$ .
Numerical stability diagnostics. The condition number $\kappa_2(A) = \sigma_1 / \sigma_n$ quantifies how much the solution $x$ of $Ax = b$ amplifies errors in $b$ . An ill-conditioned calibration problem signals that the model has redundant parameters or the data is insufficient to identify them separately.

Mathematical setting. Let $A \in \mathbb{R}^{m \times n}$ with $m \geq n$ (the overdetermined case common in calibration). No symmetry assumption. $r = \text{rank}(A) \leq \min(m, n)$ .

Notation. $\sigma_1 \geq \sigma_2 \geq \cdots \geq \sigma_r > 0 = \sigma_{r+1} = \cdots = \sigma_n$ are the singular values of $A$ . Subscripts follow the usual convention: $\sigma_1$ is the largest (spectral norm of $A$ ).

INSIGHT

Financial Insight. SVD makes the geometry of the calibration problem explicit: $U$ rotates the output space (market observables), $V$ rotates the input space (model parameters), and $\Sigma$ stretches/compresses each independent direction by its singular value. Directions with $\sigma_i \approx 0$ are directions in parameter space that barely affect market observables — the model is near-unidentifiable in those directions.

Theory

1. The Singular Value Decomposition

THEOREM

Theorem 4.1 (SVD Existence). For any $A \in \mathbb{R}^{m \times n}$ there exist orthogonal matrices $U \in \mathbb{R}^{m \times m}$ and $V \in \mathbb{R}^{n \times n}$ , and a matrix $\Sigma \in \mathbb{R}^{m \times n}$ with $\Sigma_{ii} = \sigma_i \geq 0$ and $\Sigma_{ij} = 0$ for $i \neq j$ , such that $A = U \Sigma V^\top.$ The diagonal entries of $\Sigma$ , taken in non-increasing order $\sigma_1 \geq \sigma_2 \geq \cdots \geq 0$ , are unique and called the singular values of $A$ .

Derivation. The key is to relate singular values to eigenvalues of the symmetric matrices $A^\top A$ and $A A^\top$ .

Since $A^\top A \in \mathbb{R}^{n \times n}$ is symmetric positive semi-definite (SPSD), the Spectral Theorem (Module 3) guarantees $A^\top A = V \Lambda V^\top, \quad \Lambda = \text{diag}(\lambda_1, \ldots, \lambda_n), \quad \lambda_i \geq 0, \quad V \text{ orthogonal.}$

Define $\sigma_i = \sqrt{\lambda_i}$ for $i = 1, \ldots, r$ (the non-zero eigenvalues). For $i \leq r$ , set $u_i = \frac{1}{\sigma_i} A v_i \in \mathbb{R}^m.$

These are orthonormal: for $i \neq j$ , $\langle u_i, u_j \rangle = \frac{1}{\sigma_i \sigma_j} v_i^\top A^\top A v_j = \frac{1}{\sigma_i \sigma_j} v_i^\top V \Lambda V^\top v_j = \frac{\lambda_j}{\sigma_i \sigma_j} \delta_{ij} = 0.$

Extend $\{u_1, \ldots, u_r\}$ to an orthonormal basis $\{u_1, \ldots, u_m\}$ for $\mathbb{R}^m$ . Then $A = U \Sigma V^\top$ holds by construction.

Note: $\sigma_i^2 = \lambda_i(A^\top A) = \lambda_i(A A^\top)$ for $i \leq \min(m,n)$ — the non-zero eigenvalues of $A^\top A$ and $A A^\top$ coincide.

DEFINITION

Definition 4.1 (Thin SVD). The thin (economy) SVD retains only the $r$ non-zero singular values: $A = U_r \Sigma_r V_r^\top,$ where $U_r \in \mathbb{R}^{m \times r}$ , $\Sigma_r = \text{diag}(\sigma_1, \ldots, \sigma_r) \in \mathbb{R}^{r \times r}$ , $V_r \in \mathbb{R}^{n \times r}$ . For $m \gg r$ this is far cheaper to compute and store than the full SVD.

2. Geometric Interpretation

The decomposition $A = U \Sigma V^\top$ factors the action of $A$ into three steps:

$V^\top$ : rotate the input vector $x$ into the right singular vector basis.
$\Sigma$ : independently scale each coordinate by $\sigma_i$ (and discard the null space).
$U$ : rotate the stretched vector into the output (column) space.

INSIGHT

Financial Insight. Think of $V^\top$ as transforming raw model parameters into uncorrelated "eigen-parameters". $\Sigma$ tells you how sensitively observable prices respond to each eigen-parameter. Eigen-parameters with near-zero singular values are unobservable from the data — the calibration is degenerate along those directions.

3. The Four Fundamental Subspaces via SVD

The SVD gives explicit orthonormal bases for all four subspaces first seen in Module 1.

THEOREM

Theorem 4.2 (Fundamental Subspaces). Let $A = U \Sigma V^\top$ with rank $r$ . Then:

Column space (image): $\text{col}(A) = \text{span}(u_1, \ldots, u_r)$ — first $r$ left singular vectors.
Left null space: $\text{null}(A^\top) = \text{span}(u_{r+1}, \ldots, u_m)$ — last $m - r$ left singular vectors.
Row space: $\text{row}(A) = \text{span}(v_1, \ldots, v_r)$ — first $r$ right singular vectors.
Null space: $\text{null}(A) = \text{span}(v_{r+1}, \ldots, v_n)$ — last $n - r$ right singular vectors.

Proof sketch. $Av_i = \sigma_i u_i$ for $i \leq r$ (so $v_i \in \text{row}(A)$ , $u_i \in \text{col}(A)$ ), and $Av_i = 0$ for $i > r$ (so $v_i \in \text{null}(A)$ ). The orthogonality of $U$ and $V$ gives the subspace dimensions. $\square$

4. The Moore-Penrose Pseudoinverse

When $Ax = b$ is overdetermined ( $m > n$ , more equations than unknowns) the system generally has no exact solution. The least-squares solution minimising $\|Ax - b\|_2$ is: