Are row rank and column rank equal?

Yes, for every finite matrix over the same field.

Does transposing a matrix change its rank?

No. Transposition exchanges row and column viewpoints without changing the rank.

Can a rectangular matrix be full rank?

Yes. A two-by-five matrix of rank two is full row rank, although it cannot have five independent columns.

AI Glossary / Mathematics and optimization / Intermediate

Rank

Definition

The rank of a matrix is the number of linearly independent columns it contains, equivalently the number of linearly independent rows. It measures the dimension of the matrix's output space and identifies how many independent linear directions the matrix represents.

What it means

A table can contain many columns without containing that many independent linear patterns. If one column is the sum of two others, it adds no new direction to their span. Matrix rank counts directions rather than labels or stored entries.

For an m-by-n matrix, rank lies between zero and min(m,n). The zero matrix has rank zero. A matrix is full rank when it reaches the maximum allowed by its shape, although full column rank and full row rank make different claims for rectangular matrices.

This meaning differs from a search-result ranking or the number of axes of a tensor in a programming library. It also describes linear relationships only. A dataset with nonlinear structure can still have high matrix rank.

How it works

Elementary row operations expose pivot positions. The number of pivots is the rank. An alternative view uses singular value decomposition: in exact arithmetic, rank equals the number of nonzero singular values.

For a transformation A from n-dimensional inputs into m-dimensional outputs, rank is the dimension of all attainable outputs. The null space contains inputs mapped to zero. Rank plus null-space dimension equals n, the number of input coordinates.

Rank is important in solving Ax=b. Full column rank permits at most one exact solution for a given b, while full row rank makes every m-dimensional b attainable. A square matrix has an ordinary inverse exactly when its rank equals its size.

Numerical software uses tolerances because rounding and measurement noise obscure exact zeros. A tiny singular value may be treated as zero for a particular purpose. Numerical rank is therefore a decision tied to scale, tolerance and the reliability of the data, rather than a context-free integer obtained without assumptions.

Three practical examples

A regression analyst checks whether explanatory columns are exact combinations of others before interpreting separate coefficients.
A data-compression researcher uses low-rank approximations to represent a large matrix with fewer independent factors.
An engineer examines whether a measurement system can distinguish every unknown direction or leaves some combinations unobservable.

Think of it this way

Think of receiving several directions for navigating a flat map. “Go east,” “go north” and “go northeast” are three instructions, but the third direction can be formed from the first two.

Rank counts the independent linear directions. The analogy does not mean that one business variable is unimportant whenever it is redundant in a particular data matrix; usefulness and measurement design require additional judgment.

Worked example

Three columns, two independent directions

Let A have rows (1,0,1), (0,1,1) and (1,1,2). Its third column is the sum of its first two. Those first two columns are independent, so the rank is exactly two.

For x=(u,v,w), the output is (u+w, v+w, u+v+2w). Every output therefore has third coordinate equal to the sum of the first two.

To solve Ax=(3,4,7), choose any real t and set w=t, u=3−t, v=4−t. All these inputs produce the same output. The null space is spanned by (−1,−1,1), and its dimension is one. Rank two plus nullity one equals the three input coordinates.

By contrast, Ax=(3,4,8) has no exact solution because 8 is not 3+4. If the final matrix entry changes from 2 to 2+ε for nonzero ε, its determinant becomes ε and its exact rank becomes three. A very small ε nevertheless signals a fragile distinction when data are noisy.

Illustrative case study

A fictional retailer builds a sales model with separate columns for two product groups and their combined total. The software returns coefficients, but the team notices that different coefficient allocations produce the same predictions.

The analyst identifies the exact sum relationship and explains the rank deficiency. The team can remove the redundant total, choose a different parameterization, or use a documented regularization rule. It cannot interpret all three coefficients as independently identified by those observations.

The retailer also checks whether the relationship holds consistently after data cleaning. An accidental inconsistency may make the numerical matrix full rank while still leaving the business variables nearly redundant.

Common misconceptions and mistakes

Rank equals the number of columns. Only when all columns are linearly independent.
Full rank means well conditioned. A full-rank matrix can be extremely sensitive to small perturbations.
Duplicate columns are the only cause of deficiency. Any exact linear combination can create it.
Rank zero means no rows. A nonempty matrix filled with zeros has rank zero.
Low rank proves low predictive usefulness. A small number of directions may contain valuable information.

Business use

Rank checks are useful when a model relies on separating effects or recovering unknown quantities from measurements. They can reveal redundant features, ambiguous coefficient estimates and information that the available data cannot identify.

For a practical review, ask whether the rank claim is exact or numerical, which preprocessing was applied and what tolerance was used. Do not interpret a tiny departure from exact dependence as strong evidence of an additional reliable signal. Measurement uncertainty can be much larger than arithmetic rounding error.

Check your understanding

Why can the example matrix produce (3,4,7) but not (3,4,8)?

Answer: Every output lies in the two-dimensional plane whose third coordinate equals the sum of the first two. The first target satisfies this relationship; the second does not.

Related terms

Learn next

Read null space and pseudoinverse. Multiply the worked matrix by (−1,−1,1), then explain why adding any multiple of that vector to an existing solution leaves its output unchanged.

Pseudoinverse

Sources

NumPy, numpy.linalg.matrix_rank: https://numpy.org/doc/stable/reference/generated/numpy.linalg.matrix_rank.html. Official documentation explains singular-value counting and tolerance choices.

Aalto Dictionary of Machine Learning: https://aaltodictionaryofml.github.io/ADictML_Main.pdf. Matrix-rank terminology. The matrix relationships, perturbation calculation and fictional retail scenario are original illustrations.

Previous: Range restrictions Next: Rank (ordinality)