FI-modules: when eventually is not always
Specific all-size claims fail in the CEF paper. Counterexamples, exact computations and separately checked repairs show what must change and what survives.
A formula that works for every sufficiently large size can still fail as a formula for every size. That distinction is central to a new Math Referee audit of Church, Ellenberg and Farb's FI-modules and stability for representations of symmetric groups.
The paper develops a powerful way to organize families of mathematical objects as their size grows. Our audit finds false statements in specific applications, alongside proof gaps that admit repairs. It does not overturn the central theory of FI-modules. The useful outcome is a precise account of which conclusions must change, which can be recovered, and why.
We checked the disputed application passages against the published 2015 Duke Mathematical Journal version. They remain in that edition; the comparison is not limited to the arXiv preprint.
One generator can do two jobs
Consider two independent pieces of finite arithmetic, one modulo two and one modulo three. Together they need only one generator: repeatedly adding $(1,1)$ visits all six elements of $\mathbb Z/2\oplus\mathbb Z/3$.
More generally, for a finite abelian group with $a$ independent factors modulo two and $b$ independent factors modulo three, the minimum number of generators is $\max(a,b)$, not $a+b$.
Now suppose those counts depend polynomially on the size $n$. Their maximum need not be one polynomial at every size. Our review realizes this elementary obstruction inside the configuration spaces covered by Theorem 1.9(c). For a particular seven-dimensional manifold, the two counts are
| Number of points | Modulo-two count | Modulo-three count | Generators needed |
|---|---|---|---|
| 1 | 2 | 1 | 2 |
| 2 | 4 | 3 | 4 |
| 3 | 6 | 7 | 7 |
| 4 | 8 | 14 | 14 |
The cubic controls every size from three onward. Any polynomial agreeing with that entire tail must be the cubic itself, which gives the wrong answers at sizes one and two. This is a counterexample to the stated all-size conclusion, not just a missing step in its proof.
The repair is useful and exact: keep the maximum formula for every size, and conclude polynomiality for sufficiently large sizes. The rational and fixed-prime counts have separate polynomial formulas.
The maps matter as much as the formulas
A second application concerns matrices subject to bounds on the ranks of their powers. The proof assumes that passing to a principal submatrix preserves these conditions.
It does not. There is a four-by-four matrix whose square is zero but whose principal two-by-two submatrix is the identity. Squaring the smaller matrix gives rank two. The proposed map has taken a point out of the very space it was supposed to preserve.
A failed map does not, by itself, disprove a numerical formula. The report therefore goes further. It gives separate counterexamples for the reduced variety and for the scheme defined by the minors of the squared matrix. For the latter, a small exact rank computation combines with a proved degree bound to give a finite disproof of polynomiality. The executable checker and its recorded results are public.
Here too there are meaningful replacements. Ordinary determinantal conditions retain the all-size argument. Conditions on positive powers admit an eventual dimension formula by a different proof.
A repair can strengthen the surviving theorem
The general-coefficient configuration-space proof has another problem: reducing a submodule modulo a maximal ideal need not preserve its inclusion. The dimension estimate used in that step does not follow.
Our repair uses field-cohomology bounds, integral homology, and the natural universal-coefficient sequence. It recovers the stated generation bound over Noetherian coefficient rings, and removes the finite-Krull-dimension restriction from that proof. It does not restore the false all-size minimum-generator formula; those are different claims.
Read the evidence
The audit includes eight independent source-only assessments, two focused challenges, and three final checks. Earlier Astra and Fable reports supplied the discovery leads. The published package separates model responses from the findings selected after source checking, and distinguishes the neutral reading map from the decision graph.
The central FI-module structure theorems have no counterexample established by this audit. Several subsidiary questions, particularly in the later moduli applications, remain follow-up work. The report makes no claim of new mathematical priority or of an exhaustive search for earlier corrections.