A Row of Zeros in Gaussian Elimination Is an Answer, Not a Mistake

Three equations, three unknowns.

$$\begin{aligned} x_1+2x_2+x_3 &= 4\\ 2x_1+4x_2+3x_3 &= 11\\ 3x_1+6x_2+4x_3 &= 15 \end{aligned}$$

Adding the first equation to the second gives $3x_1+6x_2+4x_3$, the third exactly. So the third says nothing new about the variables; the only question is whether its constant agrees.

Subtract twice the first row from the second, three times the first from the third, then the new second from the new third.

$$\left[\begin{array}{ccc|c}1&2&1&4\\ 2&4&3&11\\ 3&6&4&15\end{array}\right] \longrightarrow \left[\begin{array}{ccc|c}1&2&1&4\\ 0&0&1&3\\ 0&0&1&3\end{array}\right] \longrightarrow \left[\begin{array}{ccc|c}1&2&1&4\\ 0&0&1&3\\ 0&0&0&0\end{array}\right]$$

The bottom row reads $0 = 0$, contradicting nothing. The second row's leading entry sits in the third column, so column two has no pivot and $x_2$ is free. That row gives $x_3 = 3$, and the first then reads $x_1+2x_2+3 = 4$, so $x_1 = 1-2x_2$. Writing $x_2 = t$:

$$(x_1,\,x_2,\,x_3) = (1-2t,\ t,\ 3)$$

Substitute back into the originals: $(1-2t)+2t+3 = 4$, then $2-4t+4t+9 = 11$, then $3-6t+6t+12 = 15$. Every $t$ works, so there are infinitely many solutions, a whole line of them.

Change one number: the third constant becomes $16$.

$$\left[\begin{array}{ccc|c}1&2&1&4\\ 2&4&3&11\\ 3&6&4&16\end{array}\right] \longrightarrow \left[\begin{array}{ccc|c}1&2&1&4\\ 0&0&1&3\\ 0&0&1&4\end{array}\right] \longrightarrow \left[\begin{array}{ccc|c}1&2&1&4\\ 0&0&1&3\\ 0&0&0&1\end{array}\right]$$

Same steps, same order. The bottom row now reads $0 = 1$: no solution. Geometrically the first two planes still meet in a line, and the third runs parallel to that line without containing it, so no point lies on all three. The left-hand sides still force the third equation to be the sum of the first two, so its constant had to be $4+11 = 15$. It is $16$.

In echelon form, count the pivots — the leading nonzero entries, one per nonzero row — and check the right-hand side of every all-zero row.

• Every variable column has a pivot, and no row reads $0 = c$ with $c$ nonzero: exactly one solution.
• A variable column has no pivot, and still no such contradictory row: that column is a free variable, and there are infinitely many solutions.
• Any row reads $0 = c$ with $c$ nonzero: that row is false, so there is no solution, whatever the rows above say.

A zero in the pivot position is different, and the fix is cheap. In $0x_1+2x_2 = 4$, $3x_1+x_2 = 8$, the first row has nothing to eliminate with in column one, so swap.

$$\left[\begin{array}{cc|c}0&2&4\\ 3&1&8\end{array}\right] \longrightarrow \left[\begin{array}{cc|c}3&1&8\\ 0&2&4\end{array}\right]$$

Now $x_2 = 2$, and $3x_1+2 = 8$ gives $x_1 = 2$. The cue is that the entry you need is zero, not just awkward.

Three corrections belong here. Echelon form does not require leading entries equal to $1$; that normalization belongs to the reduced form, and a row reading $2x_3 = 6$ is fine. Swapping rows to bring the largest coefficient to the top keeps rounding errors small in computer arithmetic, but it is no algebraic requirement. Landing cleanly on $0 = 1$ is not a slip: that line proves the system inconsistent, which is a complete answer.

One study followed a single well-taught tenth-grade geometry class for a year: the students grew proficient at the procedures and did well on the usual tests, yet came away with a patchy sense of the subject, understood few of the connections between the methods they had learned, and picked up inaccurate ideas about what mathematics is^[1].

Very little research watches undergraduates actually row-reduce; the closest looked at the final-exam work of 68 students in an introductory undergraduate linear algebra course at one large public research university in the United States^[2]. Row-reducing there involved at least five moments where the student writes something new: two went the same way for nearly everyone, while the other three varied widely, in how students renamed variables, invented new ones, and reasoned about parameters^[2].

Here is what makes a collapsed row unsurprising. Adding a multiple of one equation to another, swapping two equations, or multiplying an equation by a nonzero number gives a different-looking system with exactly the same solutions. Nothing is lost. So if $x_2$ has vanished from your rows, it has not gone anywhere; the system never pinned it down, which is what free means.

Students who think that way appear to judge such questions better. In one study, 98 eighth- and ninth-graders judged whether pairs of equations had the same solutions; some solved both and compared answers, others looked at what the step between the two equations had done, and that second group was right more often^[3].

Setting a wrong solution next to a right one can help, but the order matters. In two experiments on learning to solve probability problems, mixing wrong solutions in with correct ones helped learners who already knew something about the topic, and only on the hardest problems, beyond what they had studied; learners who knew little did better when shown correct solutions only^[4]. So meet the correct reading of a zero row first, then hold the wrong reading against it.

Closer to this material, two classroom experiments followed students through two-step linear equations in an Algebra I tutoring program: those who explained worked examples while they practiced understood the ideas better than those who practiced alone, and the second experiment pointed to wrong examples, alone or with correct ones, as especially helpful for that understanding^[5].

None of these studies is about this error. The one that watched undergraduates row-reduce describes how 68 students wrote their work on a single exam, not anyone misreading a collapsed row^[2]. The equivalence study used eighth- and ninth-graders on single equations, not systems, and compared students who already differed rather than assigning them a way of thinking^[3]. The warning that fluency can outrun understanding is a decades-old case study of one geometry class^[1]. That a first-year student reads $0 = 0$ as failure, and that naming the invariant repairs it, is my reasoning and nobody's finding.

The example research came from two-step equations and probability problems, not Gaussian elimination, and the algebra authors offered their wrong-example result as a suggestion rather than a settled one^[5][4]. The review of self-explanation found a small to moderate gain in the steps and the ideas when students were tested right afterwards, with much thinner evidence that it holds in a real classroom or after a delay^[6]. Asking learners to explain why a common mistake is wrong is those authors' advice, not something the review measured^[6].

Row-reduce this by hand.

$$\begin{aligned} x_1+x_2+2x_3 &= 5\\ 2x_1+3x_2+5x_3 &= 13\\ x_1+2x_2+3x_3 &= 8 \end{aligned}$$

The third row will collapse. Before deciding what that means, explain it to yourself, a habit that review recommends, though its clearer gains showed up on immediate tests^[6]. For each operation, say why the solutions have not changed. Then name which of the three cases you are in, and why the other two are ruled out.

The answer: the third equation is the second minus the first, so the bottom row comes out all zeros against zero. Two pivots, three variables, so $x_3$ is free, $x_2 = 3-x_3$ and $x_1 = 2-x_3$. With $x_3 = t$ the solutions are $(2-t,\ 3-t,\ t)$, and putting that triple into the originals returns $5$, $13$ and $8$.

A collapsed row is the elimination finishing its job. Reading it is the rest of the question, and worth practicing on its own.