Rank and the rank-nullity theorem
Rank measures how many independent directions a matrix produces. The rank-nullity theorem balances that against the directions it destroys.
Rank
The rank of A is the dimension of its column space, equal to its number of pivot columns. For an m x n matrix, rank A is at most the smaller of m and n. A matrix has full column rank if rank A = n (independent columns) and full row rank if rank A = m (columns span \(\mathbb{R}^m\)).
Row rank equals column rank
The number of pivots counts both the independent columns and the non-zero rows of an echelon form, so the dimension of the row space equals the dimension of the column space. Hence rank A = rank A^T, a surprising fact: a 5 x 100 matrix's 100 columns span at most a 5-dimensional space, because its rows do.
The rank-nullity theorem
For an m x n matrix A, \(\operatorname{rank} A + \dim \operatorname{Nul} A = n\), the number of columns. Each column is either a pivot column, counted by rank, or corresponds to a free variable, counted by nullity. In transformation terms, the dimension of the input space splits into dimensions that survive and dimensions that are sent to zero.
Using rank-nullity
If a 6 x 9 matrix has rank 4, its null space has dimension 9 - 4 = 5. If a 7 x 7 matrix's null space is a line, its rank is 6, so it is singular. If the solutions of a system of 10 equations in 12 unknowns form a 2-dimensional space, the rank must be 10, so every right-hand side gives a consistent system.
Rank and the invertible matrix theorem
For an n x n matrix, rank A = n is equivalent to invertibility, to Nul A = \(\{\mathbf{0}\}\), and to the columns forming a basis of \(\mathbb{R}^n\). Rank therefore joins determinants and pivots as a way of testing invertibility.
Rank in applications
In data analysis, a matrix of measurements with low rank means the data really depends on only a few underlying factors. Recommendation systems model the matrix of users and products as approximately low rank. In control engineering, rank tests decide whether a system can be steered to any state (controllability).
Numerical rank
With real measurements and rounding, matrices are rarely exactly rank-deficient; tiny pivots may be noise. Numerical software estimates rank by counting singular values above a tolerance, using the singular value decomposition from the final unit, which is more reliable than counting pivots in floating-point arithmetic.
Worked example
In the earlier example with rows (1, 2, 0, 1), (2, 4, 1, 4), (3, 6, 1, 5), there were 2 pivots, so the rank is 2 and the nullity is 4 - 2 = 2, matching the two null-space basis vectors. The row space has basis (1, 2, 0, 1) and (0, 0, 1, 2), also of dimension 2. The column space is a plane in \(\mathbb{R}^3\), so some systems \(A\mathbf{x} = \mathbf{b}\) have no solution.
Key takeaways
Rank is the dimension of the column space, the number of pivots, at most min(m, n); full column rank means independent columns and full row rank means spanning columns. Row rank equals column rank, so rank A = rank A^T. The rank-nullity theorem: rank A + nullity A = n, the number of columns. It quickly gives null-space dimensions and consistency facts. For square matrices, rank n is equivalent to invertibility. Low rank reveals few underlying factors in data, and rank tests controllability. Numerical rank counts singular values above a tolerance.
Flashcard
Dimension of the column space
Check your understanding
A 5 x 8 matrix has rank 5. What is the dimension of its null space?
