Rank
The rank–nullity theorem, the inequalities for rank(AB) and rank(A+B), rank-one matrices, and what AB=0 tells you.
What to know by heart
For A∈Mm×n(F), the rank is rankA=dimImA, and it equals the rank of A⊤.
- Rank–nullity: rankA+dimkerA=n.
- Product: rank(AB)≤min(rankA,rankB). Multiplying by an invertible matrix does not change the rank.
- Sylvester: for A∈Mm×n, B∈Mn×p, rankA+rankB−n≤rank(AB).
- Sum: rank(A+B)≤rankA+rankB, because Im(A+B)⊆ImA+ImB.
- Rank one: rankK=1 if and only if K=uv⊤ with u,v=0. Then trK=v⊤u and K2=(trK)K.
- Over R: rank(A⊤A)=rankA, because A⊤Ax=0⇒x⊤A⊤Ax=∥Ax∥2=0⇒Ax=0.
When to use it
- AB=0 with A,B∈Mn: by Sylvester, rankA+rankB≤n. Equivalently, ImB⊆kerA.
- Idempotent matrices (A2=A): rankA=trA. The eigenvalues are 0 or 1, A is diagonalisable, and trA counts the ones. Also rankA+rank(I−A)=n.
- A rank-one difference, e.g. rank(A−B)=1: write A−B=uv⊤, so that the whole problem goes through one vector. See the move "Rank one".
- An image inside a kernel: after a relation AB=0 or A2=0, the goal becomes a comparison of dimensions.
Pitfalls
- rank(A⊤A)=rankA fails over C with plain transposition. For A=(10i0) we have AA⊤=0, while rankA=1. Over C it holds with the conjugate transpose A∗.
- rank(A2) is not determined by rankA. (0010) and (1000) both have rank 1, with squares of rank 0 and 1.
- Rank is not the number of non-zero eigenvalues, unless the matrix is diagonalisable. A nilpotent matrix =0 has positive rank and only zero eigenvalues.
No problems with these tools yet.