Chapter 13 — Key Takeaways

The big ideas

  • The column space $C(A)$ is everything $A\mathbf{x}$ can reach — and it decides solvability. It is the span of the columns, equivalently the set of all outputs $\{A\mathbf{x}\}$, and it lives in the output space $\mathbb{R}^m$. The one criterion to remember: $A\mathbf{x} = \mathbf{b}$ is solvable if and only if $\mathbf{b} \in C(A)$. This is the column picture of Chapter 3 promoted to a named subspace. (Theme: geometry and algebra are two views of one object — $C(A)$ is at once "the span of the columns" and "the flat the transformation reaches.")
  • The null space $N(A)$ is everything $A$ collapses to zero — and it decides uniqueness. It is the solution set of $A\mathbf{x} = \mathbf{0}$, equivalently the directions the transformation crushes, and it lives in the input space $\mathbb{R}^n$. $N(A) = \{\mathbf{0}\}$ means the columns are independent and the map is one-to-one; a bigger null space means the matrix glues inputs together. (Also called the kernel, $\ker A$.)
  • Both come from one row reduction. Reduce $A$ to RREF. A basis for $C(A)$ is the pivot columns of the original $A$ (never the columns of the RREF). A basis for $N(A)$ is the special solutions — one per free variable, found by setting that free variable to $1$ and the rest to $0$. The number of pivots is the rank; the number of free variables is the nullity.
  • The complete solution of $A\mathbf{x} = \mathbf{b}$ is a particular solution plus the entire null space: $\mathbf{x} = \mathbf{x}_p + N(A)$. Geometrically, the solution set is the flat $N(A)$ slid over to pass through $\mathbf{x}_p$. This unifies the none/one/infinite trichotomy of Chapter 3: reachability ($\mathbf{b}\in C(A)$?) gives existence, and the null space gives uniqueness.
  • Each special solution is a dependence relation among the columns. A vector in $N(A)$ is a recipe $x_1\mathbf{a}_1 + \cdots + x_n\mathbf{a}_n = \mathbf{0}$ — exactly the linear dependence of Chapter 6. This is why the column-space basis is the pivot columns: the free columns are the redundant ones the special solutions expose.
  • Shape constrains the subspaces before you compute. Wide ($n>m$): always a nontrivial null space, never a unique solution. Tall ($m>n$): column space is a proper flat, most $\mathbf{b}$ unreachable (the least-squares regime, Chapter 17). Square: invertible $\iff$ $C(A)=\mathbb{R}^n$ and $N(A)=\{\mathbf{0}\}$ together. (Theme: computation validates theory — every claim here was checked against np.linalg.matrix_rank and scipy.linalg.null_space, and you built both basis-finders from scratch.)

Skills you gained

  • Finding a basis for $C(A)$ (pivot columns of the original matrix) and stating its dimension as the rank.
  • Finding a basis for $N(A)$ via special solutions, one per free variable, and stating its dimension as the nullity.
  • Testing whether $\mathbf{b} \in C(A)$ two ways: looking for a contradiction row in $[A\mid\mathbf{b}]$, and the rank test $\operatorname{rank}([A\mid\mathbf{b}]) = \operatorname{rank}(A)$ (Rouché–Capelli).
  • Writing the complete solution of a consistent system as $\mathbf{x}_p + c_1\mathbf{s}_1 + \cdots$.
  • Reading the dependence relations among the columns directly off the null-space basis.
  • Predicting existence and uniqueness from the matrix's shape alone (wide / tall / square).
  • Implementing column_space_basis(A) and null_space_basis(A) from scratch on top of Chapter 4's row_reduce, and verifying dimensions against numpy/scipy.

Terms to know

column space $C(A)$ (span of the columns; reachable outputs; in $\mathbb{R}^m$) · null space $N(A)$ (solutions of $A\mathbf{x}=\mathbf{0}$; collapsed directions; in $\mathbb{R}^n$) · kernel ($\ker A = N(A)$) · image / range ($C(A)$) · special solution (a null-space basis vector from one free variable) · pivot column / free column · particular solution $\mathbf{x}_p$ · homogeneous solution · complete solution ($\mathbf{x}_p + N(A)$) · rank (number of pivots = $\dim C(A)$) · nullity ($\dim N(A)$ = number of free variables) · consistent / inconsistent · reachable / unreachable · four fundamental subspaces · controllability matrix (its column space = reachable states) · stoichiometric / conservation matrix (its null space = balanced reactions).

Notation reinforced

  • The two locked glyphs: $C(A)$ (column space) and $N(A)$ (null space). The transpose's two subspaces, coming in Chapter 14, are $C(A^{\mathsf{T}})$ (row space) and $N(A^{\mathsf{T}})$ (left null space).
  • Bases written as $\operatorname{span}\{\dots\}$; complete solution as $\mathbf{x}_p + c_1\mathbf{s}_1 + c_2\mathbf{s}_2$.
  • numpy: np.linalg.matrix_rank(A), scipy.linalg.null_space(A), np.column_stack([A, b]); remember math is 1-indexed ($x_1$, "column 1"), numpy 0-indexed (A[:, 0]).

How this connects forward

  • Chapter 14 (Row Space, Left Null Space, and Rank–Nullity) is the immediate next step: it adds the other two fundamental subspaces and proves the rank–nullity theorem, the conservation law $\operatorname{rank}(A) + \dim N(A) = n$ that we kept seeing in the bookkeeping ($2 + 2 = 4$). Every "count the pivots, count the free variables, they sum to the columns" observation in this chapter is that theorem, foreshadowed.
  • Chapter 15 (Dimension, Basis, and Coordinates) makes "dimension" rigorous, proving every basis of a subspace has the same size — so "$\dim C(A)$" and "$\dim N(A)$" are well-defined numbers, not artifacts of how you reduced.
  • Chapter 16 (Change of Basis) shows the same subspace described in different coordinate systems — useful once the SVD gives you orthonormal bases for all four subspaces.
  • Chapter 17 (Linear Regression) is the payoff for tall matrices: when $\mathbf{b} \notin C(A)$ (noisy, overdetermined data), least squares finds the point of $C(A)$ closest to $\mathbf{b}$. The reachable/unreachable distinction you learned here is what makes regression meaningful.
  • Chapter 19 (Orthogonal Projection) reveals that $N(A)$ is perpendicular to the row space and $N(A^{\mathsf{T}})$ to the column space — the four subspaces pair into two right-angled couples.
  • Chapter 30 (SVD) hands you orthonormal bases for all four fundamental subspaces at once, and the rank (the number you computed here) becomes the count of nonzero singular values.

The recurring themes, revisited

This chapter advanced two of the book's six themes. The four fundamental subspaces are the organizing framework for all of linear algebra: you built the first two, $C(A)$ and $N(A)$, and met Strang's diagram that the rest of the book hangs from. Geometry and algebra are two views of one object: the rank is simultaneously a count of pivots, the dimension of an image, the number of reachable directions in a control system, and the number of independent conservation laws in a reaction — one number, many faces. Keep the funnel picture — a matrix pinches the null space to a point and spreads the rest across the column space — in front of you; Chapter 14 is about to make its dimensions add up exactly.