A matrix has two different kinds of vectors around it: input vectors in and output vectors in . Keeping these spaces separate makes a useful test for whether a system can be solved. This unit builds on rank–nullity and Euclidean orthogonality.
Two pairs of perpendicular spaces
Definition 1.2.1
(The four fundamental subspaces)
Let , with .
- The column space is the set of outputs .
- The row space is the span of the rows of , each regarded as a column vector.
- The null space lives in the input space.
- The left null space lives in the output space.
The null space and the left null space describe two different kinds of orthogonality, each inside its own ambient space.
Theorem 1.2.2
(Orthogonality of the four subspaces)
For every ,
and consequently and . If , the four spaces have dimensions , , and for , , and .
Proof
(Dot products against rows and columns)
Let be the rows of . Then holds exactly when for every , that is, when is perpendicular to every row. A vector is perpendicular to every row exactly when it is perpendicular to every linear combination of rows, because the dot product is linear in each slot. Hence .
Applying this to , whose row space is , gives .
For dimensions, row rank equals column rank, so . Rank–nullity for gives , and for (also of rank , with columns) gives .
Finally, in every subspace satisfies and . Both sides have the same dimension, so . Taking complements in the two equalities above gives and .
Proof X-ray
Which step uses which fact
The first equality is only an unfolding of the definition of matrix–vector multiplication: row of is the dot product . No hypothesis on enters, which is why a square or full-rank matrix is not needed. The one structural move is the second equality, obtained by applying the first to : transposing exchanges the roles of rows and columns, and with them the input and output spaces.
The dimension count is where a genuine theorem is used, namely that row rank equals column rank; without it could not be read off from . The argument closes through plus equal dimensions, which is a finite-dimensional step. When the two perpendicular pairs sit in different ambient spaces, so a statement such as has no meaning.
Consistency is a statement about every left-null vector
Proposition 1.2.3
(Solvability certificate)
The system is solvable if and only if
In particular, a single with and shows that has no solution.
Proof
(Pairing the equation with a left-null vector)
If and , then . This proves the forward direction and the statement about a single certificate. Conversely, suppose for all . Then by the orthogonality theorem, so for some .
A single left-null vector certifies impossibility, but it certifies solvability only when it spans the entire left null space. When the left null space is zero, every is in the column space.
Counterexample mode
One passing test may miss an obstruction
Claim: if some nonzero has , then is solvable. This is false. Let be the zero matrix, so and . For , the nonzero vector lies in and satisfies , yet for every . The independent left-null vector gives and detects the obstruction. The repaired statement requires for a whole basis of .
Worked examples
Worked example 1.2.4
(Computing all four subspaces)
Take
The third row is twice the first plus the second; the first two are independent, so . Solving the two independent equations , gives
The second vector is in because ; the dimension shows it spans. Hence
The column space is a plane in , as expected for rank two.
Worked example 1.2.5
(Testing consistency with the certificate)
Continue with the matrix above. For we get , so the system is solvable. Solving gives all solutions , ; indeed and . The free parameter is the multiple of the null vector.
For the left-null vector has dot product , so is impossible. The witness explains the obstruction without attempting every possible input.
Worked example 1.2.6
(A full-row-rank edge case)
Let , so , and . Then has dimension , hence it is and the certificate imposes no condition: every is attainable. Explicitly gives . The null space has dimension : from and we get , which is perpendicular to both rows, since and . So solutions exist for every right-hand side but are never unique.
Common mistakes and subtle points
Common mistake
(Row operations change the column space)
Elimination preserves the row space and the null space, but not the column space or the left null space. For , the column space is ; after the row operation the matrix is with column space . Testing consistency of must use the original matrix, or an augmented matrix reduced with the right-hand side.
Common mistake
(Comparing spaces from different ambient spaces)
The row space and null space live in ; the column space and left null space live in . For a matrix the row space is a subspace of while the column space is a subspace of , so they cannot be equal or perpendicular. Only the pairs in the theorem are orthogonal complements.
Exercises
Quick check
(In the first worked example, verify that is perpendicular to each row of .)
Compute the three dot products.
Solution · Answer
The rows are , and . The dot products with are , and .
Quick check
(In the same example, does lie in the column space?)
Use the equation of the column space.
Solution · Answer
Yes: . For example maps to the first column, which is .
Quick check
(If is with rank , what is ?)
The left null space lives in .
Solution · Answer
It is , so and every is attainable.
Quick check
(If and , what can be concluded about ?)
Pair the equation with .
Solution · Answer
The system is inconsistent: a solution would give . The vector is a certificate of impossibility.