Feed a matrix a vector and it does something to it. Some directions come out the other side alive, stretched or squished but still there in the output. Other directions get flattened to nothing — the matrix sends them straight to the zero vector, and they vanish. Here's the beautiful part: every single input direction has exactly one of these two fates. It either survives into the output, or it collapses to zero. Nothing does both; nothing does neither.

The rank–nullity theorem is just the accountant's version of that sentence. For an m \times n matrix A — a machine that takes vectors from \mathbb{R}^n and returns vectors in \mathbb{R}^m — it says:

The number of surviving directions plus the number of collapsing directions equals the number of directions you started with. The books always balance. And that n on the right is the number of columns — the dimension of the input space — a detail that trips up almost everyone the first time, so keep an eye on it.

Meet the two characters

Both terms in the equation you already know from the two pages before this one; the theorem just insists they add up.

• Rank is the dimension of the column space — also called the image. It counts the surviving directions: how many independent directions the output actually fills. It equals the number of pivot columns once you row-reduce.
• Nullity is the dimension of the null space. It counts the collapsing directions: the independent directions A crushes to zero. It equals the number of free columns (the non-pivot columns) once you row-reduce.

And there is the whole proof in one breath. Row-reduce A. Every one of its n columns is either a pivot column or a free column — there is no third kind. Pivot columns are counted by the rank, free columns are counted by the nullity, and every column falls into exactly one bucket. So the two counts must add to n. The theorem isn't a coincidence you verify with examples; it's the statement that you can't be in two buckets at once.

See the books balance

Picture the input space \mathbb{R}^n as a bar of height n. Drag the sliders: n sets how many input dimensions there are, r sets the rank. The bar splits into a surviving part (height = rank, which flows across to the image) and a collapsing part (height = nullity, which folds down onto the single zero vector). The two pieces always stack back up to the full height n — that's the theorem, made of coloured blocks.

Notice what the extremes mean. Push r all the way up to n and the collapsing block disappears — nullity 0, nothing crushed, the map is injective (one-to-one). Pull r down to 0 and the whole space collapses to a point — that's the zero matrix, which sends everything to \mathbf{0}.

Worked example: a 3×4 matrix

Let A be the 3 \times 4 matrix A = \begin{bmatrix} 1 & 2 & 0 & 3 \\ 0 & 0 & 1 & 4 \\ 2 & 4 & 1 & 10 \end{bmatrix}. Row-reduce. Row 3 minus twice row 1 gives [0,0,1,4], which is identical to row 2, so subtracting row 2 wipes it out: \begin{bmatrix} 1 & 2 & 0 & 3 \\ 0 & 0 & 1 & 4 \\ 0 & 0 & 0 & 0 \end{bmatrix}. Two pivots — in columns 1 and 3. So \operatorname{rank}(A) = 2.

Now the theorem does the rest without any more arithmetic. There are n = 4 columns, so \operatorname{nullity}(A) = n - \operatorname{rank}(A) = 4 - 2 = 2. The null space is two-dimensional — a whole plane of vectors that A sends to zero. Columns 2 and 4 are the free columns, one free parameter each, which is exactly the two collapsing directions the count promised. You never had to solve A\mathbf{x} = \mathbf{0} to know its solution set is a plane; the theorem told you the moment you counted the pivots.

Worked example: a wide system must have hidden solutions

Here's the theorem earning its keep as a guarantee. Take any homogeneous system with more unknowns than equations — say 2 equations in 3 unknowns, A\mathbf{x} = \mathbf{0} with A a 2 \times 3 matrix. Does it have a solution other than the obvious \mathbf{x} = \mathbf{0}?

Rank is bounded by the number of rows, so \operatorname{rank}(A) \le 2. There are n = 3 columns, so \operatorname{nullity}(A) = 3 - \operatorname{rank}(A) \ge 3 - 2 = 1. A nullity of at least 1 means the null space contains more than just the zero vector — there is a whole line (at least) of nonzero solutions. So every homogeneous system with more unknowns than equations has infinitely many solutions. Not "usually" — always, provably, with no computation at all. More generally, a map \mathbb{R}^n \to \mathbb{R}^m with n > m can never be injective: the rank can't exceed m, so the nullity is forced to be positive, and something must get crushed.

Why this ties everything together

Once you have this one equation, a pile of facts about matrices stop being separate rules and become the same rule seen from different sides. Take a square n \times n matrix. The following are all just \operatorname{rank} + \operatorname{nullity} = n read out loud:

• A is invertible \iff \operatorname{nullity}(A) = 0 \iff \operatorname{rank}(A) = n. Full rank leaves no room for a nontrivial null space, so the map is one-to-one and onto at once.
• If any direction collapses (\operatorname{nullity} > 0), the rank falls short of n, the columns can't span \mathbb{R}^n, and the matrix is singular.
• Injective and surjective become the same condition for a square matrix — lose a dimension to the null space and you lose exactly a dimension of the image. That's why "it has an inverse," "its determinant is nonzero," and "it has no free variables" are all one idea.

For rectangular matrices the two halves separate again — a tall matrix can be injective without being surjective — and rank–nullity is precisely the ledger that tracks the trade-off. It's the hinge the rest of linear algebra swings on, and it sits right alongside rank and solvability, which reads the same numbers off to decide how many solutions a system has.