How a matrix moves when the basis moves
Fix an operator T: V -> V and two bases of V. Let A be the matrix of T in the first basis, A' in the second, and let S be the change of basis matrix between them. The two matrices are related by A' = S^-1 A S. Reading right to left: translate new coordinates to old (S), apply T in old coordinates (A), translate the answer back to new (S^-1).
Two square matrices A and A' are called similar when A' = S^-1 A S for some invertible S. The whole content of similarity is this: A and A' are similar precisely when they represent the same operator in different bases. Similarity is the matrix-level shadow of 'it's the same map underneath.'
A = [2, 0; S = [1, 1; S^-1 = [ 1, -1;
0, 3] 0, 1] 0, 1]
A' = S^-1 A S
= [1,-1; 0,1] * [2,0; 0,3] * [1,1; 0,1]
= [1,-1; 0,1] * [2,2; 0,3]
= [2, -1;
0, 3]
A and A' are similar: different numbers, SAME operator.
Shared fingerprints: trace 2+3 = 5, determinant 2*3 = 6.Invariants: what survives every disguise
The operation A -> S^-1 A S is called conjugation of operators. Because similar matrices are the same operator in disguise, any quantity that does NOT change under conjugation is a genuine property of the operator. These are the similarity invariants — and you already know several.
- Trace is invariant: tr(S^-1 A S) = tr(A), because trace ignores conjugation (tr(XY) = tr(YX)).
- Determinant is invariant: det(S^-1 A S) = det(S)^-1 det(A) det(S) = det(A).
- The characteristic polynomial is invariant, and with it the entire spectrum of eigenvalues.
- Rank and nullity are invariant — conjugation cannot change how much the operator collapses.
Conjugation as moving the viewpoint
There is a slogan worth memorizing: conjugation is a change of viewpoint, not a change of object. S^-1 A S means 'do A, but seen from the coordinate system S sets up.' This reframes a search for the best matrix as a search for the best basis — diagonalize when you can, and when an operator cannot be diagonalized, find the cleanest near-diagonal form instead.
This viewpoint also explains a powerful move: if a subspace U is invariant under T (meaning T(U) ⊆ U), then choosing a basis adapted to U makes the matrix block-triangular. The top-left block is the restriction of T to U, and the bottom-right block is the induced map on the quotient V/U. Similarity lets us choose the basis that exposes such structure.