Chapter 6 — Key Takeaways

The big ideas

  • A subspace is a subset closed under linear combination — and it always contains the origin. Take any vectors inside it, scale and add them however you like, and you never leave. The three-part test is: contains $\mathbf{0}$, closed under addition, closed under scalar multiplication. The fastest disqualifier in linear algebra: if it misses the origin, it is not a subspace. In $\mathbb{R}^2$ the subspaces are exactly the origin, lines through the origin, and the whole plane; in $\mathbb{R}^3$, add planes through the origin. (Theme: geometry and algebra are two views of one object — a subspace is a flat through the origin AND a set closed under an algebraic operation.)
  • The span of a set of vectors is everything you can reach by scaling and adding them. Geometrically it is a line (one direction), a plane (two independent directions), or all of space (three, in $\mathbb{R}^3$). The span of any set is always a subspace — proved in §6.8 — so span is the machine that manufactures subspaces. Adding a redundant vector does not enlarge the span.
  • Linear independence means no vector is redundant. Precisely: the only combination reaching $\mathbf{0}$ is the trivial all-zeros one. Equivalently (§6.9): no vector lies in the span of the others. Equivalently: row reduction gives a pivot in every column. Dependence is the existence of a nontrivial combination equal to $\mathbf{0}$ — a redundancy you can read off as "$\mathbf{v}_j = $ a combination of the rest."
  • "Spans $\mathbb{R}^n$" and "is independent" are different — often opposite — conditions. With $k$ vectors in $\mathbb{R}^n$ stacked as columns ($\operatorname{rank} = r$): independent $\iff r = k$ (pivot in every column); spans $\iff r = n$ (pivot in every row). A set can span without being independent (too many vectors) or be independent without spanning (too few). (Common pitfall: do not assume one from the other.)
  • Independence forces uniqueness of coordinates. If a set is independent, every vector in its span has exactly one recipe. This is what makes a basis usable as a coordinate system — and dependence destroys it.
  • Basis = independence + spanning, and a basis of $\mathbb{R}^n$ needs exactly $n$ vectors. This is the sweet spot the whole chapter aims at: big enough to reach everything, small enough to waste nothing. (Theme: computation validates theory — every claim here was checked with np.linalg.matrix_rank, and you built independent from scratch to confirm it.)

Skills you gained

  • Testing a subset for the three subspace conditions, and disqualifying non-subspaces instantly by checking the origin.
  • Computing and visualizing the span of vectors as a line, plane, or all of space (matplotlib, 2D and 3D).
  • Deciding span membership by setting up and solving the system $A\mathbf{c} = \mathbf{b}$.
  • Testing linear independence three equivalent ways: the definition, row reduction to a pivot in every column, and np.linalg.matrix_rank.
  • Reading off the dependency relation among dependent vectors (via the null space) and identifying which vector is redundant.
  • Distinguishing "spans $\mathbb{R}^n$" from "independent" using the rank-versus-rows and rank-versus-columns comparisons.
  • Implementing independent(vectors) from scratch and verifying it against numpy.

Terms to know

subspace (closed under linear combination; contains $\mathbf{0}$) · closure (under addition and scalar multiplication) · line / plane through the origin · linear combination · span · spanning set · linear independence · linear dependence · trivial combination (all coefficients zero) · nontrivial combination · redundant vector · pivot (a pivot in every column $\Leftrightarrow$ independent) · rank (number of pivots = number of independent vectors) · basis (independent + spanning) · dimension (number of basis vectors) · null space (nontrivial solutions of $A\mathbf{c}=\mathbf{0}$; certifies dependence) · multicollinearity (dependent feature columns) · singular configuration (Jacobian columns become dependent).

Notation reinforced

  • Span: $\operatorname{span}\{\mathbf{v}_1, \dots, \mathbf{v}_k\}$ (roman "span").
  • Independence as a rank statement: $\operatorname{rank}(A) = k$ for $k$ vectors stacked as columns.
  • Vectors bold lowercase ($\mathbf{v}, \mathbf{w}, \mathbf{c}$); the coefficient/coordinate vector is $\mathbf{c}$.
  • numpy: np.linalg.matrix_rank(A), np.column_stack, scipy.linalg.null_space; remember math is 1-indexed, numpy 0-indexed.

How this connects forward

  • Chapter 7 makes the book's central turn — a matrix is a function that transforms space — and the span and independence you learned here become a matrix's column space and rank.
  • Chapters 13–14 (the four fundamental subspaces) are the immediate payoff: the column space $C(A)$ is the span of $A$'s columns (this chapter!), the null space $N(A)$ is the solution set of $A\mathbf{x} = \mathbf{0}$ (the dependence relations!), and the row space and left null space complete Strang's organizing picture. Rank–nullity ties their dimensions together.
  • Chapter 15 (Dimension, Basis, and Coordinates) builds the basis properly: it proves every basis of a subspace has the same size (so dimension is well-defined) and that coordinates in a basis are unique — the uniqueness you glimpsed in §6.10.
  • Chapter 16 (Change of Basis) explores the same vector having different coordinates in different bases — exactly the RGB-vs-YCbCr story of Case Study 2.
  • Chapter 17 (Linear Regression) and Chapter 19 (Orthogonal Projection) revisit Case Study 1: least squares finds the best approximation when $\mathbf{b}$ is not in the column span, and collinearity (dependent columns) is why some regressions misbehave.
  • Chapter 38 (Numerical Linear Algebra) formalizes the condition number from Case Study 1 — the measure of how close to dependent a set of columns is.

The recurring themes, revisited

This chapter advanced two of the book's six themes. Geometry and algebra are two views of one object: a span is simultaneously a flat through the origin (geometry) and the set of solutions to a system / the set of all combinations (algebra); independence is both "no redundant arrow" and "a pivot in every column." Computation validates theory: you proved span is a subspace and that dependence means redundancy, then confirmed every example numerically and built independent from scratch to check against matrix_rank. Keep the picture of the spanning plane in your head — when Chapters 13–14 attach four subspaces to every matrix, you will recognize each one as a span or a solution set you already understand.