MathematicsPreprintTheory4 min read

WHEN SHOULD VAR OVERRULE THE REFEREE?

On 26 June 2026, in the last round of the World Cup group stage, Iran faced Egypt in Seattle. The authors were in the stadium. In the 93rd minute, Shoja Khalilzadeh scored what looked like the winner, and the on-field refereeing team ruled the goal onside. After minutes of deliberation, the video assistant referee (VAR) overturned the call: according to its 3D reconstruction, a small part of the scorer’s foot had been in an offside position. The match ended in a draw. According to The Athletic, Iran still had a 92% chance of reaching the knockout rounds; the next day’s results took it to zero.

The authors’ question is simple: should a very small offside margin reported by VAR be enough to overturn the referee’s call?

17 centimetres between two frames

Precision on screen is not precision on the pitch. In video recorded at 30 frames per second, an athlete running at 5 metres per second moves about 17 centimetres between two consecutive frames. Even a slight uncertainty about the exact moment the ball was played can matter more than the millimetres of a 3D model.

Amir Ali Ahmadi, of Princeton University, and Parsa and Radeen Alavi, of Algonquin Regional High School, turned the dispute into a probability model. The true margin, θ, is the distance of the attacker beyond the offside line: positive means offside. The referee perceives it with a random error of spread σ_H, but only says yes or no. VAR measures a number, with its own random error of spread σ_AI. Both errors are assumed Gaussian — spread along the familiar bell curve — and independent.

The √(2/π) rule

The authors then look for the margin that makes what was observed most likely — a call from the referee and a number from VAR — a method called maximum likelihood. They prove that this best estimate always exists and is unique. Better still, the final decision needs no computation at all, only a comparison with a threshold:

c = √(2/π) × σ_AI² / σ_H

When the two disagree, VAR overturns the referee only if the margin it measures is larger than c. Below that, the referee’s call stands. The number √(2/π), about 0.80, comes from the bell curve itself.

Take a referee whose typical error is 10 cm and a VAR whose typical error is 2 cm. The threshold is 0.32 cm. A conflicting VAR margin of 0.20 cm leaves the referee’s call in place; one of 0.50 cm overturns it. Placing the attacker beyond the line is not enough: VAR has to place them beyond it by more than the threshold.

Two errors that do not weigh the same

The formula holds an asymmetry. The threshold grows with the square of VAR’s error, but is only inversely proportional to the referee’s: double VAR’s error and the threshold quadruples, double the referee’s and it only halves. A noisier VAR needs stronger evidence; a noisier referee is easier to overrule. The authors trace the different powers to the two kinds of information: a number for VAR, a mere sign for the referee.

Combining both also pays off. With equal errors of 2 cm and a true margin of 2 cm, the referee alone or VAR alone is wrong 15.87% of the time; the combined rule, 9.70%.

The model also handles bias. A referee inclined to flag offside, by 5 cm on average, gets two different thresholds: 0.20 cm to overturn an offside call, 0.46 cm to overturn an onside one. And if VAR itself has a known 1 cm bias, ignoring it can flip the verdict.

To use the rule, one needs the four error parameters. The paper shows how to estimate them from calibration plays where the true margin is known: a simple average and spread for VAR, and a “probit” fit — a standard statistical method that recovers a bias and a spread from yes-or-no answers — for the referee’s calls.

Beyond the pitch

The same structure appears wherever a human gives a verdict and a machine gives a measurement of the same quantity: a doctor calling a lesion benign or malignant next to an AI malignancy score, an engineer clearing an aircraft part next to a defect-severity score, a fund manager next to a model’s expected return. The authors suggest testing the rule’s real effect where calibration is easy, such as finance, medical imaging and prediction markets.

The paper is written as a teaching piece, for undergraduates and advanced high school students who know basic probability and calculus. Its parameters are illustrative: no real referee or VAR system has been calibrated. Its last line belongs to the acknowledgements, where the authors say they used ChatGPT for figures, proofreading and sanity checks, “while following their √(2/π) rule to ensure that AI did not overturn their human judgment”.