Group-stage tiebreaks

Why can a team win two of its three group matches and go home while a team that won only one advances?

It reads like an error in the arithmetic. One team wins two of its three matches and flies home; another wins one, draws two, and plays on. Nobody cheated and everyone agreed the rules in advance.

So before blaming the rulebook, ask what a group stage is actually being asked to do — and whether what it produces is what we read off it.

Notice first that a group stage never ranks the field. It ranks each group separately and applies the cut inside each one. That is a partition, and a partition throws away every comparison crossing it. Nobody ever asks whether the third-placed team in one group is better than the second-placed team in another; the format has no machinery for the question.

That alone gets us most of the way. Two wins in a hard group and one win in a soft group are never weighed against each other, so of course the softer group can send forward the lower total. The tournament is not making a mistake; it is answering a question we did not notice it was asking.

But there is a second mechanism, and it is the more interesting one. Inside a group of four, each team plays three matches, and points arrive in a coarse currency — three, one or nothing — so possible totals are few and teams are many. Ties are not a rare accident of this design; they are close to its normal output.

Can we build the worst case exactly? Take teams A, B, C, D. Let A beat B, B beat C and C beat A — a perfect cycle — and let all three beat D. Now A, B and C each have two wins and one defeat: six points apiece, D on none. Two advance, so a team with two wins is going home; and if a team elsewhere finishes second on one win and two draws, we have precisely the situation in the question.

Now watch the rules try to separate A, B and C. The obvious appeal is head-to-head: how did the tied teams do against each other? But by construction that sub-table is the cycle. There is no ordering to find. The criterion does not give a wrong answer, it gives none.

So the rules fall back on something drawn from the whole group — goal difference, goals scored. And here is the point worth sitting with: these criteria do not measure the same thing. Head-to-head asks who beat whom, goal difference by how much, goals scored how much attacking was done. A team can top one and bottom another on identical results. So the order the criteria are listed in is itself a decision that changes who qualifies — and the common orderings differ, some competitions applying head-to-head before goal difference and others the reverse.

Does that make one order right? Tempting, but no. Ask what any tiebreak must satisfy. It must be announced in advance, or teams cannot know what they are playing for. It must be computable from the results, since nothing else is observed. And it must always return an answer, even when the results are cyclic. Those three demands do not sit comfortably together — the last guarantees that at some depth the rule stops measuring merit and starts simply deciding, which is why the cascades bottom out in a drawing of lots.

The first demand has its own consequence. If the criteria are public and final matches are played one after another, both teams in a late match can compute the result that suits them — which is why final-round group matches are now simultaneous, a fix aimed at the incentive rather than the ranking.

Three refinements.

First, keep the two mechanisms apart. The partition explains cross-group anomalies — a stronger record eliminated while a weaker one advances elsewhere. The coarse currency plus the cascade explains within-group ones, where a team goes out on a criterion unrelated to winning. The case in the question is usually the first; the one that generates most anger is the second.

Second, this is a close cousin of the problem in voting systems, and the resemblance is not decorative: in both, reasonable ways of aggregating pairwise results can disagree, and cycles can leave no consistent ordering at all. The difference is the input. An election aggregates preferences held simultaneously by voters; this aggregates results produced sequentially by the very parties the rule will judge — which is why gameability, not just consistency, is a design constraint here.

Third, it is distinct from the reason knockout stages surprise us. There the mechanism is sampling noise: single matches are unreliable, so a stronger side loses often enough to be eliminated. Here the results are taken exactly as they came, and the trouble is in how they are aggregated and where the cut falls. Noise is not doing the work.

What would falsify the account? Its distinctive claim is that the ordering of criteria changes outcomes, not merely their existence. So re-rank historical group tables under the alternative ordering and count how often the qualifying pair changes. A count of zero across a large sample would make the claim that these criteria measure different things close to empty, and the phenomenon would collapse back onto the partition effect alone. The partition claim has its own test: stronger seeding, which evens out group strength, should reduce cross-group anomalies with no rule rewritten.

• Why three points for a win changed which ties occur, compared with the older two-point system.
• How ranking third-placed teams across groups reintroduces the comparison the partition removed.

Every term the collection defines is gathered in the glossary.