You can solve a system of three equations with three variables by eliminating one variable at a time until a single equation with one unknown remains, then back-substituting upward. The method that survives every case is Gaussian elimination: pick a pivot, clear that variable from the other two equations, repeat, then substitute back. It works on 3Γ3 systems with a unique solution, exposes the ones with none, and flags the ones with infinitely many.
What follows is the full procedure with exact arithmetic moves, two worked examples, a comparison of elimination against substitution and Cramer's rule, and the failure modes β inconsistent rows, dependent planes, near-parallel coefficients β that textbooks underplay and exams love.
Why elimination order changes the arithmetic, not the answer
Order of elimination is free, but order of arithmetic is not. Every row operation you perform multiplies through coefficients you'll later carry, so a bad pivot choice inflates the numbers and the chances of a slip. Skilled solvers scan the column first and pick the coefficient that divides cleanly into its column-mates.
A practical rule: pivot on the coefficient with the smallest absolute value that still divides the others evenly, or on a 1 if one exists. In a system like 2x + 3y β z = 5, 4x β y + 2z = 3, x + 2y + 3z = 8, the third equation's x-coefficient of 1 is the natural pivot β eliminating x from rows one and two costs two multiplications and produces no fractions.
This choice determines how much scratch paper you burn, which is the real bottleneck in a timed exam. The mathematics is indifferent; your error rate is not. Once you commit to an order, stick with it, because switching mid-solve is how sign errors creep in.
Pivot selection has a second payoff: it tells you early whether the system is degenerate. If eliminating a variable kills two entire rows to zeros, you're looking at infinite solutions, and you learned that in one step instead of four.
Setting up the augmented matrix and reading its columns
Write the augmented matrix by listing coefficients of x, y, and z in order, then the constants after a divider. For 2x β y + 3z = 7, x + 4y β 2z = β3, 3x + 2y + z = 10, the matrix reads [2 β1 3 | 7], [1 4 β2 | β3], [3 2 1 | 10]. Every row is one equation; every column but the last is one variable.
The columns must be in consistent variable order across all three rows, or the row operations mean nothing. A common early mistake is writing the second equation as 4y β 2z = β3 + x and then dropping the x-term when forming the row. That error is invisible until the answer fails a check.
The divider before the constants column is not decoration. Row operations on the coefficient block change the equations; treating the constant column as a fourth variable is exactly what makes the augmented form work. Once you're in reduced row echelon form, that last column is the answer.
Before writing anything, clear fractions by multiplying entire equations by a common denominator. A single 1/3 in row one will propagate into every subsequent operation, and the cleanest way to avoid fractional arithmetic is to never create it.
Triangularizing: the numbered elimination sequence
This is the workhorse sequence. It works on every 3Γ3 system with a unique solution, and it exposes the exceptions as a side effect.
1: Choose a pivot in the x-column β ideally a 1 or a small coefficient that divides the others. If the top row's x-coefficient is 0, swap rows. 2: Eliminate x from the other two equations: for each, multiply the pivot row by the ratio (target x-coefficient Γ· pivot coefficient) and subtract. 3: You now have two equations in y and z. Choose a pivot in the y-column of that 2Γ2 block and eliminate y from the remaining equation. 4: The last row is now a single-variable equation in z. Solve it. 5: Back-substitute z into the second row to get y, then y and z into the first row to get x. 6: Check the triple in all three original equations, not just one.
If step 2 produces a row of the form [0 0 0 | nonzero], stop: the system is inconsistent. If it produces [0 0 0 | 0], that equation carried no new information and the system has infinitely many solutions.
Failure handling matters more than the happy path. When a pivot is zero, swap down. When a swap isn't available, you're in a degenerate system β proceed to the inconsistency test rather than forcing arithmetic. Sign errors are caught by the check in step 6; if one equation fails, re-trace the last elimination you performed, since that's where the slip almost always lives.
Worked example one: a clean integer solution
Solve 2x + y β z = 3, x β y + 2z = 4, 3x + 2y + z = 10. Matrix: [2 1 β1 | 3], [1 β1 2 | 4], [3 2 1 | 10]. Swap rows one and two so the pivot is 1: [1 β1 2 | 4], [2 1 β1 | 3], [3 2 1 | 10].
R2 β 2R1 gives [0 3 β5 | β5]. R3 β 3R1 gives [0 5 β5 | β2]. The 2Γ2 block is 3y β 5z = β5 and 5y β 5z = β2.
Subtracting the second from the first after scaling β multiply the first by 5 and the second by 3 to clear y: 15y β 25z = β25 and 15y β 15z = β6. Subtract: β10z = β19, so z = 1.9. That's not the integer I promised, and it's the point: textbook problems often produce ugly decimals, and you keep going. Back-substitute into 3y β 5z = β5: 3y β 9.5 = β5, so y = 1.5. Then x β 1.5 + 3.8 = 4, so x = 1.7.
Check: 2(1.7) + 1.5 β 1.9 = 3.4 + 1.5 β 1.9 = 3. β. Second: 1.7 β 1.5 + 3.8 = 4. β. Third: 5.1 + 3 + 1.9 = 10. β. The decimal arithmetic is the honest cost of a system whose coefficients don't divide nicely β and a reminder that "clean" systems are constructed, not typical.
Worked example two: substitution on a system with a 1
Take x + y + z = 6, 2x β y + z = 3, x + 2y β z = 2. The first equation solves for x with no division: x = 6 β y β z. Substitute into the other two.
Second: 2(6 β y β z) β y + z = 3, giving 12 β 2y β 2z β y + z = 3, so β3y β z = β9, or 3y + z = 9. Third: (6 β y β z) + 2y β z = 2, giving 6 + y β 2z = 2, so y β 2z = β4.
Now solve 3y + z = 9 and y β 2z = β4. From the second, y = 2z β 4. Substituting: 3(2z β 4) + z = 9, so 7z β 12 = 9, z = 3. Then y = 2(3) β 4 = 2, and x = 6 β 2 β 3 = 1.
Check: 1 + 2 + 3 = 6 β; 2 β 2 + 3 = 3 β; 1 + 4 β 3 = 2 β. Substitution wins here because the pivot coefficient was exactly 1. It loses badly when no coefficient is 1 β you'll be juggling fractions from the first line.
Elimination versus substitution versus Cramer's rule on the same system
| Criterion | Elimination | Substitution | Cramer's rule |
|---|---|---|---|
| Best when | Coefficients are small integers, no 1s | A coefficient is 1 or β1 | You need only one variable |
| Determinant work | None | None | Four 3Γ3 determinants |
| Arithmetic risk | Sign errors in row ops | Fraction buildup | Determinant expansion slips |
| Detects no-solution | Yes, zero row with nonzero constant | Yes, contradiction | Only if det = 0, then re-check |
| Detects infinite solutions | Yes, zero row with zero constant | Yes, identity | Needs separate rank test |
| Hand-computation cost (typical 3Γ3) | 8β12 multiplications | 6β10 if a 1 exists, 20+ otherwise | 4 determinants β 40 multiplications |
The verdict splits by use. For hand arithmetic, elimination with a smart pivot is the default, and substitution is a shortcut only when a coefficient is 1 or β1. Cramer's rule is a legitimate choice when a question asks for z alone and nothing else β you compute three determinants instead of a full triangularization. It's the worst choice for a full solution, and it's silent on inconsistency unless you check the determinant of the coefficient matrix first.
Cramer's rule has a diagnosable failure: if the coefficient determinant is 0, the rule gives no answer, and you cannot say from that alone whether the planes are parallel or coincident. You must fall back to elimination to distinguish "no solution" from "infinitely many."
Reading the three failure cases: no solution, infinite solutions, and the row that lies
Three planes in space intersect in exactly one of four ways: at a point (unique solution), along a line or in a plane (infinitely many), nowhere (no solution), or β the case students forget β two coincident planes intersected by a third in a line.
A row like [0 0 0 | 5] is a lie: it claims 0 = 5. When elimination produces it, declare no solution immediately. Geometrically, at least two planes are parallel and distinct, or all three form a triangular prism with no common point.
A row like [0 0 0 | 0] is a truth that says nothing. Keep it and continue; the remaining nonzero rows tell you the solution set's dimension. One dependent row among three means a line of solutions β parametrize with one free variable. Two dependent rows means a plane of solutions.
Worked case: x + y + z = 2, 2x + 2y + 2z = 4, x β y = 0. The second equation is twice the first, so it eliminates to [0 0 0 | 0]. With x = y from the third, substitute into the first: 2x + z = 2, so z = 2 β 2x. Solution set: (t, t, 2 β 2t) for any real t. The check that matters: plug t = 0 to get (0, 0, 2) and t = 1 to get (1, 1, 0); both satisfy all three original equations, confirming the line, not a point.
The near-parallel warning: why 0.001 coefficients matter
Exact arithmetic hides a real-world failure. Consider 0.001x + y = 1 and x + y = 2. Solving gives x β 1.001 and y β 0.999 β fine on paper. But perturb the first constant by 0.001 and x swings roughly 1 unit. The system is ill-conditioned: tiny input changes produce huge output changes.
This matters whenever coefficients come from measurements β physics labs, economics models, survey regressions β where the numbers carry 2β3 significant figures. Elimination is exact on the symbols but amplifies the rounding error buried in the inputs. A triangularization that divides by 0.001 multiplies any error in that row by 1,000.
For hand computation the fix is partial pivoting: always eliminate using the row with the largest absolute coefficient in the current column, not the topmost one. That keeps every multiplier at or below 1 in magnitude and prevents error amplification. Textbooks rarely emphasize this because their coefficients are integers, but any course that models real data will.
The sanity check is cheap: after solving, perturb one input by its rounding unit and re-solve. If the answer moves more than a few percent, report the solution as approximate and say so. An honest "x = 1.0 Β± 0.2" beats a false "x = 1.001."
Word problems: turning three constraints into three equations
Most exam 3Γ3 problems arrive as a story. The translation discipline: name the variables explicitly, write one equation per constraint sentence, and never combine two constraints into one equation. Three sentences, three unknowns, three equations β if your count is off, you've merged or split a constraint.
Classic pattern: a total, a relationship, and a rate. "A theater sold 500 tickets for $4,800; adult tickets cost $12, child tickets $6, and there were 40 more child than adult tickets." Variables a, c, s for adult, child, student. Equations: a + c + s = 500; 12a + 6c + 4s = 4,800; c = a + 40. Substitute c everywhere, and you have a 2Γ2 in a and s.
A second pattern is the mixture problem: three ingredients with concentrations and a target. A third is the age or work-rate problem, which is linear only if you set it up in rates, not times. The setup step carries most of the marks; the algebra is routine once the equations are right.
Watch for the sentence that says "twice as many X as Y plus 10" β that's X = 2Y + 10, not 2X = Y + 10. Word-problem errors are translation errors far more often than arithmetic ones.
Checking your answer and catching the classic sign slip
Check all three original equations with the final triple. Solving two of three and checking one is how wrong answers survive. The third equation is your only independent audit.
The most common slip is a dropped minus in the subtraction step of elimination: R2 β 2R1 means subtract twice row one from row two, and students routinely subtract row two from row one instead. If your final check fails on exactly one equation, suspect that step first.
A second common slip: forgetting that multiplying a row by a ratio before subtracting changes the constant column too. If you scaled only the coefficient block, the constant is now wrong by the same factor, and the error propagates silently.
When the answer is fractional, convert to a common denominator once at the end rather than carrying decimals. And when a system has integer coefficients and produces a decimal answer, double-check before accepting it β constructed problems usually have cleaner solutions, and a decimal often signals the sign slip, not the problem.
FAQ
Can you solve three equations with three variables by graphing?
Only conceptually. Three planes intersect at a point, a line, a plane, or nowhere, and graphing them by hand gives you a rough picture, not coordinates. Software that renders 3D surfaces will show the intersection, but reading a triple like (1.7, 1.5, 1.9) off a screen image is unreliable. Graph to understand the geometry, compute to get the answer.
What does it mean when elimination gives 0 = 0?
That equation carried no independent information β it was a combination of the others. Keep eliminating. If you end with one nonzero equation and two zeros, the solution is a line with one free variable; two nonzero equations and one zero gives one free variable's worth of solutions. You have infinitely many solutions, not one.
Is there a formula for 3Γ3 systems like the quadratic formula?
Yes β Cramer's rule, which expresses each variable as a ratio of two 3Γ3 determinants. It's exact and gives a closed form, but it needs four determinant evaluations for a full solution, roughly 40 multiplications by hand. Elimination is faster in practice, and Cramer's rule tells you nothing useful when the coefficient determinant is zero.
Why did I get no solution when the problem says there is one?
Check whether you multiplied a row by a ratio using the wrong row as the pivot. A zero row with a nonzero constant almost always traces to a sign error or a mis-scaled row. Re-run the elimination with the pivot choice written down explicitly, and verify the original equations still match your starting matrix.
How many equations do you need to solve for three variables?
Three independent equations for a unique triple. Two equations give a line of solutions; one gives a plane. "Independent" matters more than the count: three equations can still describe only a line if one is a combination of the others, which is exactly the degenerate case elimination exposes as a zero row.
Can I use matrices and a calculator instead of hand elimination?
Yes, and it's standard practice above 3Γ3. Enter the augmented matrix, apply row-reduced echelon form, and read the last column. The caveat is entry error: a single mistyped coefficient produces a confidently wrong answer with no visible inconsistency, so check the triple in the originals. On an exam requiring shown work, the matrix is documentation, not a substitute for the steps.
What's the difference between a 3Γ3 system with no solution and one with infinite solutions?
No solution means the planes never share a point β elimination yields a row [0 0 0 | c] with c β 0. Infinite solutions means they share a line or plane β elimination yields [0 0 0 | 0] and leaves free variables. The two look similar in the matrix and are opposite in meaning, which is why the constant column must be tracked.
Do these methods work if the equations aren't linear?
No. Elimination, substitution, and Cramer's rule all depend on each variable appearing to the first power with constant coefficients. Systems with products like xy, squares like xΒ², or variables inside functions need numerical methods or algebraic tricks specific to the structure, and the three-plane picture no longer applies.