Brownian Bridge™

Setup

Why linear algebra in quant finance?

Three operations dominate production quant code:

Portfolio aggregation: multiply a weight vector $w \in \mathbb{R}^n$ by a covariance matrix $\Sigma \in \mathbb{R}^{n \times n}$ to get portfolio variance $w^\top \Sigma w$ .
System solving: bootstrap a yield curve by solving a linear system $Ax = b$ where $A$ encodes instrument sensitivities to discount factors.
Factor decomposition: decompose a risk matrix to identify which linear combinations of assets drive most of the variance.

All three reduce to linear algebra. This module builds the precise foundations: what a vector space is, what a matrix does geometrically, and what the rank-nullity theorem tells you about the structure of solutions to linear systems.

INSIGHT

Why this matters on a derivatives desk. When bootstrapping a swap curve, the system $Ax = b$ has $A \in \mathbb{R}^{n \times n}$ encoding instrument-to-discount-factor relationships and $b$ containing observed par swap rates. If two instruments are economically equivalent (redundant), $A$ is singular. The rank-nullity theorem tells you exactly by how many degrees the solution is underdetermined — and therefore how many regularisation constraints you need to add to recover uniqueness.

Conventions

All vectors are column vectors in $\mathbb{R}^n$ unless stated otherwise.
$\mathbb{R}^{m \times n}$ denotes the set of real $m \times n$ matrices.
The standard inner product on $\mathbb{R}^n$ is $\langle x, y \rangle = x^\top y = \sum_{i=1}^n x_i y_i$ .
Notation: $\ker(A) = \{x \in \mathbb{R}^n : Ax = 0\}$ (null space); $\operatorname{im}(A) = \{Ax : x \in \mathbb{R}^n\}$ (column space / image).
Superscript $\top$ denotes transpose: $(A^\top)_{ij} = A_{ji}$ .

Theory

1. Vector Spaces

DEFINITION

Definition 1.1 (Vector Space). A vector space over $\mathbb{R}$ is a non-empty set $V$ equipped with two operations — addition $V \times V \to V$ and scalar multiplication $\mathbb{R} \times V \to V$ — satisfying eight axioms:

(Closure) $u + v \in V$ and $\alpha v \in V$ for all $u, v \in V$ , $\alpha \in \mathbb{R}$ .
(Commutativity) $u + v = v + u$ .
(Associativity) $(u + v) + w = u + (v + w)$ and $(\alpha\beta)v = \alpha(\beta v)$ .
(Zero vector) There exists $\mathbf{0} \in V$ with $v + \mathbf{0} = v$ for all $v$ .
(Additive inverses) For each $v \in V$ there exists $-v \in V$ with $v + (-v) = \mathbf{0}$ .
(Unit scalar) $1 \cdot v = v$ .
(Scalar distributivity) $\alpha(u + v) = \alpha u + \alpha v$ .
(Vector distributivity) $(\alpha + \beta)v = \alpha v + \beta v$ .

The canonical example is $\mathbb{R}^n$ : column vectors with componentwise addition and scaling. Other examples relevant to quant finance:

$\mathbb{R}^{m \times n}$ : the space of $m \times n$ matrices, used as parameter spaces in calibration.
$C[0, T]$ : continuous functions on $[0, T]$ , the natural home for asset price paths.
$L^2(\Omega, \mathcal{F}, \mathbb{P})$ : square-integrable random variables; the space in which conditional expectations and martingales live (covered in the Probability Theory course).

DEFINITION

Definition 1.2 (Subspace). A non-empty subset $W \subseteq V$ is a subspace if it is closed under addition and scalar multiplication — equivalently, if $\alpha u + \beta v \in W$ for all $u, v \in W$ and $\alpha, \beta \in \mathbb{R}$ .

The subspace test: $W$ is a subspace iff (i) $\mathbf{0} \in W$ and (ii) $W$ is closed under linear combinations.

EXAMPLE

Example 1.3 (Dollar-neutral portfolios). Let $V = \mathbb{R}^n$ be the space of portfolio weight vectors. The set $W = \{w \in \mathbb{R}^n : \mathbf{1}^\top w = 0\}$ of dollar-neutral portfolios (zero net exposure) is a subspace of dimension $n - 1$ . The sum of two dollar-neutral portfolios remains dollar-neutral; scaling preserves the constraint. This subspace arises in statistical arbitrage and market-neutral strategies.

2. Linear Spans and Bases

DEFINITION

Definition 2.1 (Linear span). The span of vectors $v_1, \ldots, v_k \in V$ is $\operatorname{span}(v_1, \ldots, v_k) = \left\{ \sum_{i=1}^k \alpha_i v_i : \alpha_i \in \mathbb{R} \right\}.$ This is the smallest subspace of $V$ containing all $v_i$ .

DEFINITION

Definition 2.2 (Basis and dimension). A basis of $V$ is a set $\{e_1, \ldots, e_n\}$ that is simultaneously:

Linearly independent: $\sum_i \alpha_i e_i = \mathbf{0}$ implies $\alpha_i = 0$ for all $i$ .
Spanning: $\operatorname{span}(e_1, \ldots, e_n) = V$ .

The number of elements in any basis is the dimension $\dim V$ — this is well-defined (any two bases of a finite-dimensional space have the same cardinality, proved via the Steinitz exchange lemma).

REMARK

Coordinate representation. Once a basis $\{e_1, \ldots, e_n\}$ is fixed, every $v \in V$ has a unique representation $v = \sum_i \alpha_i e_i$ . The coefficients $(\alpha_1, \ldots, \alpha_n)$ are the coordinates of $v$ in that basis. A change of basis is itself a linear map (an invertible matrix).

3. Matrices as Linear Maps

A matrix $A \in \mathbb{R}^{m \times n}$ defines a linear map $T_A : \mathbb{R}^n \to \mathbb{R}^m$ by $T_A(x) = Ax$ . Linearity is immediate: $T_A(\alpha x + \beta y) = A(\alpha x + \beta y) = \alpha Ax + \beta Ay = \alpha T_A(x) + \beta T_A(y).$

Conversely, every linear map between finite-dimensional spaces has a unique matrix representation once bases are fixed — the $j$ -th column of $A$ is the image of the $j$ -th basis vector of the domain.

DEFINITION

Definition 3.1 (Kernel and image). For $A \in \mathbb{R}^{m \times n}$ :

The null space (kernel): $\ker(A) = \{x \in \mathbb{R}^n : Ax = 0\}$ — a subspace of $\mathbb{R}^n$ .
The column space (image): $\operatorname{im}(A) = \{Ax : x \in \mathbb{R}^n\} = \operatorname{span}(\text{columns of } A)$ — a subspace of $\mathbb{R}^m$ .

Solvability: the system $Ax = b$ has a solution iff $b \in \operatorname{im}(A)$ . It has a unique solution iff additionally $\ker(A) = \{0\}$ .

THEOREM

Theorem 3.2 (Rank-Nullity). For any $A \in \mathbb{R}^{m \times n}$ : $\underbrace{\dim \operatorname{im}(A)}_{\operatorname{rank}(A)} + \underbrace{\dim \ker(A)}_{\operatorname{nullity}(A)} = n.$

Proof sketch. Let $r = \operatorname{rank}(A)$ . Choose a basis $\{u_1, \ldots, u_{n-r}\}$ for $\ker(A)$ , then extend to a basis $\{u_1, \ldots, u_{n-r}, v_1, \ldots, v_r\}$ of $\mathbb{R}^n$ . For any $x = \sum \beta_i u_i + \sum \gamma_j v_j$ , we have $Ax = \sum \gamma_j Av_j$ . It follows that $\{Av_1, \ldots, Av_r\}$ spans $\operatorname{im}(A)$ and is linearly independent (if a combination were zero, it would imply a dependence contradicting the extended basis). So $\dim \operatorname{im}(A) = r$ , giving $r + (n - r) = n$ . $\square$

EXAMPLE

Example 3.3 (Redundant swap instruments). Suppose five market instruments each price as a linear combination of six discount factors. The pricing equations are $Ax = b$ with $A \in \mathbb{R}^{5 \times 6}$ . By rank-nullity, $\operatorname{rank}(A) + \operatorname{nullity}(A) = 6$ . Since $A$ has only 5 rows, $\operatorname{rank}(A) \leq 5$ , so $\operatorname{nullity}(A) \geq 1$ : there is at least a one-dimensional family of solutions. The solution is not unique — additional constraints (e.g., smoothness of the forward curve) are required to select one.

4. The Four Fundamental Subspaces

For $A \in \mathbb{R}^{m \times n}$ with $\operatorname{rank}(A) = r$ , there are four canonical subspaces:

Subspace	Lives in	Dimension
Column space $\operatorname{im}(A)$	$\mathbb{R}^m$	$r$
Left null space $\ker(A^\top)$	$\mathbb{R}^m$	$m - r$
Row space $\operatorname{im}(A^\top)$	$\mathbb{R}^n$	$r$
Null space $\ker(A)$	$\mathbb{R}^n$	$n - r$

THEOREM

Theorem 4.1 (Orthogonality of fundamental subspaces). $\operatorname{im}(A) \perp \ker(A^\top) \quad \text{and} \quad \operatorname{im}(A^\top) \perp \ker(A).$ Furthermore, $\operatorname{im}(A) \oplus \ker(A^\top) = \mathbb{R}^m$ and $\operatorname{im}(A^\top) \oplus \ker(A) = \mathbb{R}^n$ (orthogonal direct sum decompositions).

Proof. Let $y \in \ker(A^\top)$ and $b = Ax \in \operatorname{im}(A)$ . Then $\langle b, y \rangle = (Ax)^\top y = x^\top A^\top y = x^\top \mathbf{0} = 0$ . The direct sum decomposition follows from dimension counting and orthogonality. $\square$

REMARK

Geometric picture. Think of $A$ as a map from $\mathbb{R}^n$ to $\mathbb{R}^m$ . The row space $\operatorname{im}(A^\top)$ is the "effective" part of $\mathbb{R}^n$ — the directions $A$ actually acts on. The null space $\ker(A)$ is the "invisible" part — directions $A$ crushes to zero. In $\mathbb{R}^m$ , the column space $\operatorname{im}(A)$ is the range and the left null space $\ker(A^\top)$ is the "unachievable" part.

5. Symmetric and Positive (Semi-)Definite Matrices

DEFINITION

Definition 5.1. A matrix $A \in \mathbb{R}^{n \times n}$ is:

Symmetric if $A^\top = A$ .
Positive semi-definite (SPSD) if $A^\top = A$ and $x^\top A x \geq 0$ for all $x \in \mathbb{R}^n$ .
Positive definite (SPD) if $A^\top = A$ and $x^\top A x > 0$ for all $x \neq 0$ .

Notation: $A \succeq 0$ (SPSD), $A \succ 0$ (SPD).

The quadratic form $x^\top A x$ has a geometric interpretation: it measures how much $A$ stretches $x$ in the direction of $x$ itself. Positive definiteness says $A$ never "collapses" any direction to zero or below.

THEOREM

Proposition 5.2 (Covariance matrices are SPSD). Let $X \in \mathbb{R}^{T \times n}$ be a returns matrix ( $T$ observations, $n$ assets) with mean-zero columns. The sample covariance matrix $\Sigma = \frac{1}{T} X^\top X \in \mathbb{R}^{n \times n}$ satisfies $\Sigma \succeq 0$ . It is positive definite iff $X$ has full column rank (no asset is a perfect linear combination of others).

Proof. $\Sigma^\top = (X^\top X)^\top = X^\top X = \Sigma$ (symmetric). For any $w \in \mathbb{R}^n$ : $w^\top \Sigma w = \frac{1}{T} w^\top X^\top X w = \frac{1}{T} \|Xw\|^2 \geq 0.$ The form equals zero iff $Xw = 0$ , i.e., $w \in \ker(X)$ . Full column rank of $X$ means $\ker(X) = \{0\}$ , giving positive definiteness. $\square$

WARNING

Near-singular covariance matrices. In practice, assets are highly correlated (e.g., equity index options at adjacent strikes), and $\Sigma$ estimated from finite data is frequently near-singular: it is theoretically SPD but has eigenvalues close to zero. Inverting $\Sigma$ directly (as required for minimum-variance weights $w^* \propto \Sigma^{-1}\mu$ ) amplifies these small eigenvalues and produces wildly unstable portfolio weights. The fix — regularisation via SVD or adding a diagonal ridge — is treated in Module 4.

Validation

The companion notebook verifies:

Vector space axioms — checks all eight axioms for $\mathbb{R}^3$ with explicit counterexamples showing what fails when each axiom is dropped.
Subspace test — confirms the dollar-neutral subspace is closed under linear combinations.
Rank-nullity — for a $4 \times 6$ matrix, computes $\operatorname{rank}(A)$ and $\operatorname{nullity}(A)$ via row reduction and verifies their sum equals 6.
Orthogonality of fundamental subspaces — constructs a basis for each of the four subspaces and verifies the pairwise orthogonality via dot products.
Covariance SPSD — builds a sample covariance from synthetic returns and confirms $w^\top \Sigma w \geq 0$ for 1000 random weight vectors.

PRACTICE

By hand before opening the notebook. Let $A = \begin{pmatrix} 1 & 2 & 3 \\ 4 & 5 & 6 \\ 7 & 8 & 9 \end{pmatrix}.$

Are the rows of $A$ linearly independent? What is $\operatorname{rank}(A)$ ?
By the rank-nullity theorem, what is $\operatorname{nullity}(A)$ ?
What is the dimension of $\ker(A^\top)$ (left null space)?
Does the system $Ax = (1, 2, 3)^\top$ have a solution? Does $Ax = (1, 1, 1)^\top$ ?

(Answers: $\operatorname{rank}(A) = 2$ since row 3 = 2 × row 2 − row 1; nullity = 1; $\dim \ker(A^\top) = 1$ ; $(1,2,3)^\top$ is column 1 of $A$ so yes; check whether $(1,1,1)^\top$ is in the column space.)

Limitations

WARNING

Numerical rank is not algebraic rank. The definitions above are exact, but in floating-point arithmetic, a matrix that is "theoretically" rank-2 may appear rank-3 due to rounding errors — or vice versa. Production code determines rank using singular values with a tolerance: count the number of singular values exceeding $\varepsilon \cdot \sigma_{\max}(A)$ for a problem-specific threshold $\varepsilon$ . This is treated rigorously in Module 4 (SVD and condition numbers).

WARNING

Invertibility is not conditioning. A matrix can be theoretically invertible but numerically catastrophic if its columns are nearly linearly dependent. The condition number $\kappa(A) = \|A\| \cdot \|A^{-1}\|$ quantifies this. A yield curve system with $\kappa \approx 10^{12}$ means computed solutions are meaningless below 12 significant digits — which is the entire double-precision range on a bad day. Always check condition numbers before trusting a linear solve in calibration.

Scope limitations:

This module assumes finite-dimensional spaces over $\mathbb{R}$ . Infinite-dimensional Hilbert spaces ( $L^2$ , Sobolev spaces) require additional structure (completeness, bounded operators) and arise in the functional analytic foundations of interest rate models.
Results extend to complex vector spaces (replacing transpose with conjugate transpose $A^*$ ), which are needed for Fourier-transform pricing (Module in fourier-fft-pricing course).

Interview Angle

PRACTICE

L1 (Junior quant / developer).

"State the rank-nullity theorem and give an example with a $3 \times 4$ matrix." — Expected answer: $\operatorname{rank}(A) + \operatorname{nullity}(A) = 4$ (number of columns). Example: rank-2 matrix has a 2-dimensional null space.
"When does the system $Ax = b$ have exactly one solution?" — Expected answer: when $b \in \operatorname{im}(A)$ (solvability) and $\ker(A) = \{0\}$ (uniqueness), i.e., $A$ is square and invertible.
"What does it mean for a covariance matrix to be positive definite, and why does it matter for portfolio optimisation?" — Expected answer: $w^\top \Sigma w > 0$ for all $w \neq 0$ , so portfolio variance is always strictly positive (no costless zero-variance portfolio). Positive definiteness guarantees $\Sigma$ is invertible, which is needed to compute minimum-variance weights $w^* \propto \Sigma^{-1}\mu$ .

PRACTICE

L2 (Senior quant).

"You bootstrap a 5-instrument swap curve and get infinitely many solutions. What does linear algebra tell you?" — Expected answer: the system matrix $A$ is rank-deficient. By rank-nullity, $\operatorname{nullity}(A) \geq 1$ , so the solution set is an affine subspace of dimension $\geq 1$ . The instruments are redundant or the system is underdetermined — either remove a redundant instrument or add a regularisation constraint (smoothness prior on the forward curve).
"Explain the four fundamental subspaces and their orthogonality relations for an $m \times n$ matrix of rank $r$ ." — Expected answer: column space and left null space are orthogonal complements in $\mathbb{R}^m$ with dimensions $r$ and $m - r$ ; row space and null space are orthogonal complements in $\mathbb{R}^n$ with dimensions $r$ and $n - r$ .
"When is a covariance matrix estimated from 50 days of returns on 100 assets positive definite?" — Expected answer: never — the data matrix $X \in \mathbb{R}^{50 \times 100}$ has rank at most 50, so $\Sigma = X^\top X / T$ has rank at most 50 < 100. At least 50 eigenvalues are zero. The matrix is SPSD, not SPD.

PRACTICE

L3 (Researcher).

"In a factor risk model, the factor covariance $F \in \mathbb{R}^{k \times k}$ is full rank, but the full asset covariance $\Sigma = B F B^\top + D$ (where $B \in \mathbb{R}^{n \times k}$ with $n \gg k$ ) is used in a Markowitz optimisation. How does the structure of $BFB^\top$ affect the optimisation?" — Expected answer: $BFB^\top$ is rank $k \ll n$ (SPSD, not SPD). It lies in an $n$ -dimensional space but has a $(n-k)$ -dimensional null space. Portfolio optimisation using $(BFB^\top)^{-1}$ directly is impossible; one must use the full $\Sigma = BFB^\top + D$ (where $D$ is a diagonal idiosyncratic matrix, making $\Sigma$ SPD), or use the Woodbury identity for efficient inversion.
"How does the null space of a linear operator relate to regularisation in calibration?" — Expected answer: if $\ker(A) \neq \{0\}$ , the calibration objective has infinitely many minimisers differing by elements of $\ker(A)$ . Tikhonov regularisation (penalising $\|x\|^2$ ) is equivalent to selecting the unique minimum-norm solution, which is the orthogonal projection of any solution onto the row space $\operatorname{im}(A^\top)$ .