Bases, dimension, and structure

Coordinates feel absolute until you change them. Graphics frames, sensor arrays, and data tables all confront the same struggle: which independent directions are enough to describe everything you care about, and what stays true when you rewrite the labels?

Emmy Noether (1882-1935) rebuilt algebra around structure and homomorphism, often while institutions blocked her path. Students flocked to her anyway. MacTutor is the biographical stop. Steal the ethic carefully (and without flattening Grassmann's earlier geometric ambitions into the same story): fierce clarity about what stays invariant when representations change, and generosity in teaching that clarity. A basis is a minimal wardrobe that still dresses every vector uniquely. Dimension counts how many independent costume pieces you need.

For a modern rereader, this chapter is about structure that survives rewriting - why PCA, graphics frames, and change-of-basis drills keep returning. Without bases, coordinates feel absolute. With bases, you see the costume for what it is.

Vectors $\mathbf{v}_1,\ldots,\mathbf{v}_k$ are linearly independent if the only solution to

is all $a_i = 0$. No vector in the set is redundant.

A basis of a space is an independent spanning set: every vector has unique coordinates relative to that basis.

Dimension is the number of vectors in any basis (a theorem says they all agree).

The column space dimension of a matrix is its rank: how many independent output directions it can hit.

The same arrow can wear the costume $(1,0)$ in one basis and a different pair in another. A plane is two-dimensional because two independent directions span it - not because someone drew an $x$ and $y$ axis in a textbook.

This is why change of basis matters in graphics, PCA, and physics: choose axes that make the phenomenon simple. PCA seeks informative orthogonal directions of variance. CAD software lets you pick convenient frames for a part. Sensors that duplicate each other are dependent directions wearing hardware clothes.

Algebrica's vector spaces and change-of-basis pages reward patience. Dimension is not "number of coordinates on the page." It is the size of a basis.

Three mutually perpendicular unit axes basis ordinary space. Four legs of a table that impose coplanar constraints show dependence in physical disguise: more sticks than independent motions.

Dimension equals ambient list length. A line through the origin in $\mathbb{R}^3$ is still one-dimensional even though each point wears three ambient coordinates. Dimension is about the subspace, not the ambient costume size.

Any spanning set is a basis. Spanning sets can be redundant. Bases are spanning and independent. Throw out extras until uniqueness of coordinates appears.

Rank as a mysterious matrix property. Rank is dimension of the column space (equivalently row space). It answers: how many independent output directions can this map produce?

Coordinates as the object. Change basis and coordinates change, but geometric relationships (when properly expressed) persist. Confusing the costume for the actor is the root bug.

Ambient dimension worship. More coordinates are not automatically more truth. Effective dimension of data may be far smaller.

Confusing orthogonal with independent in casual speech. Orthogonal is an inner-product story. Independence is a linear-combination story. Related in Euclidean settings with care; not synonyms everywhere.

Emmy Noether rebuilt algebra around structure and homomorphism, often while institutions blocked her path. Students flocked to her anyway. MacTutor's Emmy Noether biography is essential reading. The ethic is fierce clarity about what stays invariant when representations change - and generosity in teaching that clarity despite the room's politics.

Change-of-basis thinking is why the same physical arrow can wear different number costumes in different frames. Twentieth-century geometry made that flexibility physical; linear algebra makes it computational.

Choosing axes for a CAD model, or picking independent sensors that do not duplicate information, is choosing a useful basis. Robotics and computer vision change frames constantly: camera, robot, world.

Compression and dimensionality reduction are bets that data approximately live in a lower-dimensional structure inside a larger ambient space.