Mathwords logoMathwords

Graeco-Latin Square — Definition & Example

A Graeco-Latin square is a grid built by overlaying two Latin squares, one of Latin letters and one of Greek letters, so that every possible Latin-Greek pair appears in exactly one cell. In experiments, it lets you test one treatment while controlling three nuisance factors in only n² runs.

A Graeco-Latin square of order nn is an n×nn \times n array whose cells contain ordered pairs (aij,bij)(a_{ij}, b_{ij}) from S×TS \times T, where ∣S∣=∣T∣=n|S| = |T| = n. The arrays (aij)(a_{ij}) and (bij)(b_{ij}) must each be Latin squares, and each of the n2n^2 pairs in S×TS \times T must occur exactly once. Two Latin squares with this property are called orthogonal. Graeco-Latin squares exist for every order n≥3n \ge 3 except n=6n = 6.

Worked Example

Problem: You want to compare the stopping distance (in feet) of 3 tire brands A, B, C. Three nuisance factors could bias the results: the car (3 cars), the test day (3 days), and the driver (3 drivers, labeled α, β, γ). Design a Graeco-Latin square that balances all of them.
Step 1: Let rows be cars and columns be days. Assign tire brands with a cyclic Latin square. Each row shifts one place to the left.
ABCBCACAB\begin{array}{|c|c|c|} \hline A & B & C \\ \hline B & C & A \\ \hline C & A & B \\ \hline \end{array}
Step 2: Assign drivers with a second Latin square. Each row shifts two places to the left instead of one. This different shift is what makes the two squares orthogonal.
αβγγαββγα\begin{array}{|c|c|c|} \hline \alpha & \beta & \gamma \\ \hline \gamma & \alpha & \beta \\ \hline \beta & \gamma & \alpha \\ \hline \end{array}
Step 3: Superimpose the squares. All 9 brand-driver pairs are different, so every driver tests every brand exactly once. You need 3² = 9 runs, compared with 3⁴ = 81 for every combination of car, day, driver, and brand.
AαBβCγBγCαAβCβAγBα\begin{array}{|c|c|c|} \hline A\alpha & B\beta & C\gamma \\ \hline B\gamma & C\alpha & A\beta \\ \hline C\beta & A\gamma & B\alpha \\ \hline \end{array}
Answer: The superimposed 3 × 3 grid above is a valid Graeco-Latin design. Each brand appears once per car, once per day, and once per driver, using 9 runs instead of 81.

Why It Matters

Euler studied these squares in his 36 officers problem. That problem asks for a Graeco-Latin square of order 6, which does not exist; Gaston Tarry confirmed this in 1900. Euler conjectured that no square exists for any order of the form 4k + 2. Bose, Shrikhande, and Parker disproved that conjecture in 1959, and by 1960 they had shown such squares exist for every order of the form 4k + 2 except 2 and 6. In industrial and agricultural experiments, the design lets you study four factors at n levels each in only n² runs.

Common Mistakes

Mistake: Running a single unreplicated 3 × 3 Graeco-Latin square and expecting an ANOVA significance test.
Correction: The 8 total degrees of freedom split as 2 for each of the four factors, leaving (n − 1)(n − 3) = 0 for error. You cannot estimate error variance this way. Replicate the square or use a larger order.

Related Terms