Common Errors in Span and Column-Space Problems

Three Questions That Must Not Be Confused

Span problems often begin with the same objects—a list of vectors and a matrix—but ask different questions. Is one particular vector generated by the list? Do the vectors generate every vector in the ambient space? Or what vectors are the columns of a row-reduced matrix generating? Confusing these questions can produce a plausible computation and an incorrect conclusion.

Throughout, scalars are real, and vectors in \(\mathbb{R}^m\) are columns. For vectors \(\mathbf{a}_1,\dots,\mathbf{a}_n\), their span is the set of all linear combinations \(x_1\mathbf{a}_1+\cdots+x_n\mathbf{a}_n\), where \(x_1,\dots,x_n\in\mathbb{R}\). If \(\mathbf{A}=[\mathbf{a}_1\ \cdots\ \mathbf{a}_n]\), its column space \(\operatorname{Col}\mathbf{A}\) is precisely the span of its columns. The product \(\mathbf{A}\mathbf{x}\) is the combination of those columns with coefficients from \(\mathbf{x}\).

These definitions suggest two different tests. For one target, solve one system, \(\mathbf{A}\mathbf{x}=\mathbf{b}\). For a claim that the columns span the whole space, the claim concerns every possible target, not just one chosen vector.

Test the Target You Were Given

Proof. Suppose \(\mathbf{b}\in\operatorname{Col}\mathbf{A}\). By the definition of column space, \(\mathbf{b}\) is a linear combination of the columns \(\mathbf{a}_1,\dots,\mathbf{a}_n\). Thus there are scalars \(x_1,\dots,x_n\) such that \(\mathbf{b}=x_1\mathbf{a}_1+\cdots+x_n\mathbf{a}_n\). With \(\mathbf{x}=(x_1,\dots,x_n)^\top\), this equation is \(\mathbf{A}\mathbf{x}=\mathbf{b}\), so the system is consistent. Conversely, if the system is consistent, it has a solution \(\mathbf{x}\). The matrix product \(\mathbf{A}\mathbf{x}\) is a linear combination of the columns of \(\mathbf{A}\), and \(\mathbf{A}\mathbf{x}=\mathbf{b}\); therefore \(\mathbf{b}\in\operatorname{Col}\mathbf{A}\). \(\square\)

To test membership, place the target in the augmented column: row-reduce \([\mathbf{A}\mid\mathbf{b}]\) and check whether the resulting system is consistent. Row operations on an augmented matrix preserve its solution set, provided they are applied to the whole augmented matrix. Omitting \(\mathbf{b}\) removes the information that the membership question is about.

One Successful Target Is Not a Whole-Space Test

If a system is consistent for one target, that establishes membership for that target only. To establish that columns span \(\mathbb{R}^m\), every target in \(\mathbb{R}^m\) must be obtainable. It is enough to check the standard basis vectors, because every vector in \(\mathbb{R}^m\) is a linear combination of them.

Proof. If the columns span \(\mathbb{R}^m\), then every vector in \(\mathbb{R}^m\), including each \(\mathbf{e}_i\), belongs to their span. For the reverse direction, suppose each \(\mathbf{e}_i\) belongs to \(\operatorname{Col}\mathbf{A}\). Take any \(\mathbf{b}=(b_1,\dots,b_m)^\top\in\mathbb{R}^m\). By the definition of the standard basis, \(\mathbf{b}=b_1\mathbf{e}_1+\cdots+b_m\mathbf{e}_m\). Each \(\mathbf{e}_i\) is a combination of the columns of \(\mathbf{A}\); substituting those combinations into this expression shows that \(\mathbf{b}\) is also a combination of those columns. Since \(\mathbf{b}\) was arbitrary, the columns span \(\mathbb{R}^m\). \(\square\)

This criterion does not say that a single successful test suffices. It gives \(m\) particular membership tests that together certify all targets. Alternatively, a general row reduction can expose a row of zeros in the coefficient matrix, which can prevent some targets from being reached. The next example shows the distinction directly.

Row-Reduced Columns Are Not the Original Generators

Row operations are useful for solving a system, but they change the entries of the matrix. In particular, the columns after row reduction are not automatically the original vectors whose span is being tested. When reducing \([\mathbf{A}\mid\mathbf{b}]\), the question is whether the transformed system has the same solutions as the original system. When someone instead reads the columns of the reduced coefficient matrix as new original generators, they have changed the question.

A row operation acts on every column of the matrix. So it can move the generators, and therefore change the set of vectors they generate. The following small example makes the distinction visible: the original column lies on a diagonal line, whereas its row-reduced version lies on a coordinate axis.

A Reliable Routine and Common Pitfalls

Before computing, identify whether the question concerns one target or all of the ambient space. Then keep the target visible throughout a membership calculation. A short routine helps prevent the three errors discussed above:

Common mistake: “There is a pivot, so the target is in the span.” A pivot in the coefficient matrix alone says nothing about whether the particular right-hand side is compatible. In the first example, the coefficient matrix has a pivot, but the augmented row \(0=-1\) proves that the target is not generated. The correction is to include the target column and inspect the complete system.

Common mistake: “One target worked, so the columns span the space.” In the second example, \((5,0)^\top\) is generated, but \((0,1)^\top\) is not. The correction is to state the limited conclusion—membership of the tested target—or perform the whole-space test.

Common mistake: “The reduced matrix has these columns, so they are the generators.” In the third example, row reduction changes \((1,1)^\top\) into \((1,0)^\top\). The correction is to use row operations to solve the augmented system, while checking any proposed coefficients against the original matrix.

In short, keep three objects distinct: the original list of generators, the particular target, and the matrix produced during row reduction. The original columns define the span; the augmented system tests membership; and row reduction is a way to solve that system without changing its solution set.