What a Pivot Argument Must Keep Track Of
Pivot positions are powerful, but their meaning depends on what matrix is being reduced and what question is being asked. A pivot column tells us an index: it identifies a column of the original matrix. It does not say that the corresponding column of the reduced matrix is the same vector as the original column. Similarly, a free variable in a linear system does not by itself guarantee solutions: an inconsistent system has no solutions, whether or not some variable column is free. Dimension counts also have precise directions. A list may be independent without spanning, or span without being independent.
Throughout, vectors are columns over \(\mathbb{R}\). For a matrix \(\mathbf{A}\in\mathbb{R}^{m\times p}\), row reduction produces a row-equivalent matrix \(\mathbf{R}=\operatorname{rref}(\mathbf{A})\). Each elementary row operation is reversible, so it preserves the solution set of a system when applied to both sides. But applying row operations to a matrix generally changes its column vectors. Keeping these facts separate prevents several common errors.
Original Pivot Columns and the Column Space
It is useful to know exactly what the pivot indices do tell us. The original columns at those indices form a basis for the column space. The proof explains both why those columns are independent and why they span; it also pinpoints why taking the pivot columns from the RREF instead can give the wrong vectors.
Proof. Write \(\mathbf{R}=\operatorname{rref}(\mathbf{A})\). Each row operation is reversible, so the sequence of row operations taking \(\mathbf{A}\) to \(\mathbf{R}\) is multiplication on the left by an invertible matrix \(\mathbf{P}\); thus \(\mathbf{R}=\mathbf{P}\mathbf{A}\). In \(\mathbf{R}\), each pivot column is a standard basis vector, so the pivot columns are linearly independent. If a linear combination of the original pivot columns is zero, applying \(\mathbf{P}\) gives the same combination of the pivot columns of \(\mathbf{R}\) equal to zero. Their independence forces every coefficient to be zero. Thus the original pivot columns are independent. For any nonpivot column of \(\mathbf{R}\), its entries in the pivot rows give coefficients expressing it as a combination of the pivot columns of \(\mathbf{R}\); its other entries are zero. Applying \(\mathbf{P}^{-1}\) to this equality expresses the corresponding original column of \(\mathbf{A}\) as the same combination of the original pivot columns. Therefore the original pivot columns span every column of \(\mathbf{A}\), and hence span \(\operatorname{Col}\mathbf{A}\). They are independent and spanning, so they form a basis. \(\square\)
Worked Example: The RREF Columns Are Not the Original Columns
Consider the columns \(\mathbf{a}_1=(1,0,1)\), \(\mathbf{a}_2=(0,1,1)\), and \(\mathbf{a}_3=(1,1,2)\). Reduce the matrix with these columns:
The pivot indices are \(1\) and \(2\). The pivot columns of the RREF are \((1,0,0)^{\top}\) and \((0,1,0)^{\top}\), but the original pivot columns are \(\mathbf{a}_1=(1,0,1)^{\top}\) and \(\mathbf{a}_2=(0,1,1)^{\top}\). The latter form a basis for the original column space. In fact, \(\mathbf{a}_3=\mathbf{a}_1+\mathbf{a}_2\), so the original three columns span exactly the plane of vectors \((x,y,x+y)^{\top}\). The RREF columns instead span the plane of vectors \((x,y,0)^{\top}\), which is a different plane.
Check: \(\mathbf{a}_1+\mathbf{a}_2=(1,0,1)^{\top}+(0,1,1)^{\top}=(1,1,2)^{\top}=\mathbf{a}_3\). Also, every combination \(c_1\mathbf{a}_1+c_2\mathbf{a}_2=(c_1,c_2,c_1+c_2)^{\top}\), confirming the stated original column space.
The row operations in this example preserve the equations’ solutions, but they change the actual column vectors and their span. A common incorrect step is to read the RREF’s pivot columns as a basis for the original column space. The pivot indices locate the original columns to use; the reduced columns themselves usually belong to a transformed column space.
Free Variables Require a Consistent System
For a homogeneous system \(\mathbf{A}\mathbf{x}=\mathbf{0}\), a nonpivot variable can be assigned freely, and assigning a nonzero value produces a nonzero solution after the pivot variables are determined. For an inhomogeneous system \(\mathbf{A}\mathbf{x}=\mathbf{b}\), first check consistency. A row of the form \([0\ \cdots\ 0\mid c]\) with \(c\ne0\) represents the impossible equation \(0=c\). In that case, the system has no solutions, not infinitely many.
Worked Example: A Free Variable Does Not Rescue an Inconsistent System
Consider the system \(x+y=2\), \(2x+2y=5\). Its augmented matrix reduces as follows:
There is a pivot in column \(1\), the column for \(x\), and in column \(3\), the augmented column. Column \(2\), for \(y\), has no pivot. But the last row says \(0=1\), so the system is inconsistent and has no solutions. The free-variable parameter procedure applies only after consistency has been established.
Check: The first original equation gives \(x+y=2\), so the left side of the second equation would be \(2x+2y=4\), not \(5\). Thus no pair \((x,y)\) can satisfy both equations.
Another mistaken reading is to count a pivot in the augmented column as if it were a pivot for a variable. The augmented column is not a variable column. Its pivot signals inconsistency when it occurs in a row whose coefficient entries are all zero.
Dimension Counts: Check Both Directions
Let \(V\) be a finite-dimensional vector space over \(\mathbb{R}\), with \(\dim V=d\), and let \(\mathcal{B}\) be a basis. For a list of \(p\) vectors in \(V\), placing their coordinate vectors \([\mathbf{v}_1]_{\mathcal{B}},\dots,[\mathbf{v}_p]_{\mathcal{B}}\) into a \(d\times p\) matrix turns independence and spanning into column questions in \(\mathbb{R}^d\). The coordinate representation is unique, so a relation among the vectors holds exactly when the same relation holds among their coordinate columns.
Proof. Form the \(d\times p\) coordinate matrix \(\mathbf{C}\) of the list in a basis of \(V\), and let \(r\) be its number of pivots. By the pivot-column theorem, the pivot columns form a basis for \(\operatorname{Col}\mathbf{C}\). The list is independent exactly when every column is a pivot column: if a column is nonpivot, it is a combination of pivot columns and gives a nontrivial dependence relation; if all are pivot columns, they are independent by the theorem. Thus independence implies \(r=p\). Since there can be at most \(d\) pivots in a matrix with \(d\) rows, \(p\le d\). The list spans \(V\) exactly when its coordinate columns span \(\mathbb{R}^d\). By the pivot-column theorem, this holds exactly when \(r=d\), so \(p\ge d\). If \(p=d\), independence is equivalent to \(r=p=d\), and spanning is also equivalent to \(r=d=p\). Therefore, for a list of exactly \(d\) vectors, independence and spanning are equivalent. \(\square\)
The count does not say that every list with fewer than \(d\) vectors is independent, or that every list with more than \(d\) vectors spans. It gives necessary bounds: independence requires \(p\le d\), while spanning requires \(p\ge d\). When \(p=d\), either property forces the other; when \(p\ne d\), one must not infer the other from the count alone.
Worked Example: Three Vectors Can Span \(\mathbb{R}^3\) and Still Be Dependent
Consider four vectors in \(\mathbb{R}^3\): \(\mathbf{v}_1=(1,0,0)\), \(\mathbf{v}_2=(0,1,0)\), \(\mathbf{v}_3=(0,0,1)\), and \(\mathbf{v}_4=(1,1,0)\). Their column matrix is already in RREF:
The first three columns are pivot columns, so the list spans \(\mathbb{R}^3\): they are the standard basis vectors. The fourth column is nonpivot and is a combination of the first two. Thus the list is not independent, despite spanning the whole space. This is consistent with the count: there are \(p=4>d=3\) vectors, so independence is impossible, but spanning remains possible.
Check: \(\mathbf{v}_1+\mathbf{v}_2-\mathbf{v}_4=(1,0,0)+(0,1,0)-(1,1,0)=(0,0,0)\), a nontrivial dependence relation. The first three vectors themselves span because \(a\mathbf{v}_1+b\mathbf{v}_2+c\mathbf{v}_3=(a,b,c)\) for every \((a,b,c)\in\mathbb{R}^3\).
Common Mistakes and Their Corrections
Mistake: “The pivot columns of the RREF are a basis for the original column space.” In the first example, this would claim that \((1,0,0)^{\top}\) and \((0,1,0)^{\top}\) span the original column space. But every original column has third coordinate equal to the sum of its first two coordinates, while \((1,0,0)^{\top}\) does not. Correction: use the pivot indices from the RREF to select columns of the original matrix.
Mistake: “There is a free variable, so there are infinitely many solutions.” The augmented example has a free variable column and nevertheless has no solutions. Correction: inspect for an inconsistent row first. A consistent system with at least one free variable has infinitely many solutions: choose the free parameter in infinitely many ways, and the pivot variables are then determined. Without consistency, there are zero solutions.
Mistake: “An independent list is a basis, even if it has too few vectors.” The two vectors \((1,0,0)^{\top}\) and \((0,1,0)^{\top}\) are independent, but they do not span \(\mathbb{R}^3\), since their combinations always have third coordinate \(0\). Correction: a basis must be both independent and spanning. Dimension counts show that a list of fewer than three vectors cannot span \(\mathbb{R}^3\).
| Observation | Valid conclusion | Do not conclude |
|---|---|---|
| A column is nonpivot | The corresponding original column is a combination of original pivot columns | The RREF pivot columns are the original basis vectors |
| A variable column is free | There is a parameter if the system is consistent | The system necessarily has solutions |
| An independent list has \(p\) vectors in dimension \(d\) | \(p\le d\) | The list necessarily spans when \(p<d\) |
| A spanning list has \(p\) vectors in dimension \(d\) | \(p\ge d\) | The list necessarily is independent when \(p>d\) |
Before accepting a pivot or dimension argument, identify the matrix being reduced, distinguish variable columns from any augmented column, and state whether the system is consistent. For a column-space claim, select original columns by pivot index. For a dimension claim, check whether the argument concerns independence, spanning, or both; only a list whose size equals the dimension links those two properties directly.
Check Your Understanding
For each question, state the relevant condition before drawing a conclusion.
- Why do pivot indices identify original columns, but the columns in the RREF need not form a basis for the original column space?
- A consistent system has a free variable. What can be concluded about its solution set, and why does consistency matter?
- Can four vectors in \(\mathbb{R}^3\) span \(\mathbb{R}^3\)? Can they be independent? Explain both answers using the dimension count.
- Give a reason that two independent vectors in \(\mathbb{R}^3\) need not be a basis for \(\mathbb{R}^3\).
- In an augmented RREF, what does a pivot in the augmented column indicate when all entries to its left in that row are zero?
