# Review of FI-modules and stability for representations of symmetric groups

**Igor Rivin · Math Referee · 20 November 2026**

This AI-assisted, source-adjudicated review identifies false application
statements in Church, Ellenberg and Farb's paper, together with proof defects
that admit repairs. It does **not** overturn the central theory of FI-modules.
The distinction between a polynomial formula for every size and one valid only
for sufficiently large sizes is essential to the findings.

The reference edition is the published paper, *Duke Mathematical Journal*
**164** (2015), 1833–1910,
[DOI 10.1215/00127094-3120274](https://doi.org/10.1215/00127094-3120274).
The independent readers received the complete
[final author version, arXiv:1204.4533v4](https://arxiv.org/abs/1204.4533v4).
We then compared the disputed application passages with the supplied 78-page
journal PDF. The journal retains them. This is a targeted comparison, not a
claim that the two editions are identical.

The [counterexamples and repairs](counterexamples-and-repairs.md) supply
the detailed mathematics. The [check bundle](cef-checks.zip) contains
the executable arithmetic checks and recorded results. The neutral reading
page and its source map are separate from these findings.

## Findings at a glance

| Published location | Finding | What survives |
| --- | --- | --- |
| Theorem 1.9(c), pp.1838–1839 | False: the integral minimum-generator count need not be polynomial for all sizes. | An exact maximum-of-polynomials formula, and eventual polynomiality. |
| Theorem 5.3.2, p.1888 | False in its stated generality, including a homogeneous, nonconstant example for the minor scheme. The general setup also need not be graded. | Ordinary determinantal conditions retain the all-size conclusion; positive-power conditions admit an eventual replacement. |
| Remark 5.3.3, p.1888 | False for the reduced variety defined by rank of the square at most one. | Eventual polynomiality in this example; more general repairs need compatible maps. |
| Proof of Theorem 6.4.3, pp.1897–1898 | A subquotient does not give the claimed inequality after reduction modulo a maximal ideal. | A universal-coefficient proof recovers bounded generation; finite Krull dimension is unnecessary in that proof. |
| Lemma 6.3.2(b), p.1893 | The converse from a filtered module to its associated graded pieces is false. | The forward implication used in the stability argument holds. |
| Main FI-module structure theorems | No counterexample to the principal finite-generation, Noetherian, classification or eventual-character results was established by this audit. | The scoped positive assessments retain their imported prerequisites and qualifications. |

The first three rows are false statements, rather than merely missing
explanations. A repaired statement does not make the original quantifiers true.

## 1. Integral generators: why a maximum appears

For a finitely generated abelian group, the minimum number of generators is
its free rank plus the largest generator count among its primary torsion
parts. Different primes do not add their requirements: a single generator
can serve coprime cyclic factors simultaneously.

An FI-sharp decomposition gives finite binomial sums

$$R(n)=\sum_j r_j\binom nj,
\qquad T_p(n)=\sum_j t_{j,p}\binom nj.$$

The exact integral minimum is

$$\mu(n)=R(n)+\max_p T_p(n).$$

Each fixed-prime count is polynomial. Their maximum need not be a single
polynomial on all nonnegative integers. It is eventually polynomial because
only finitely many primes occur and the signs of finitely many polynomial
differences eventually stabilize.

This is realized by an actual manifold satisfying the paper's hypotheses.
Let

$$M=\bigl(L^7(3)\mathbin{\#}UTS^4\mathbin{\#}UTS^4\bigr)
\setminus\{\mathrm{point}\}.$$

Here $UTS^4$ is the unit tangent sphere bundle of the four-sphere. The space
$M$ is a connected oriented seven-manifold, diffeomorphic to the interior
of a compact manifold with nonempty boundary. Standard bundle, Künneth and
universal-coefficient calculations give

$$H^4(\operatorname{Conf}_n(M);\mathbb Z)
\cong (\mathbb Z/2)^{2n}\oplus
(\mathbb Z/3)^{n+\binom n2+\binom n3}.$$

Removing the pair diagonals does not change these low-degree groups: the
diagonals have codimension seven. Therefore

$$\mu(n)=\max\left\{2n,\frac{n^3+5n}{6}\right\}.$$

This equals the cubic for every $n\ge3$, but at $n=1,2$ its values are
$2,4$ rather than $1,3$. No polynomial can agree with a cubic on an infinite
tail and disagree with it at those two inputs. This refutes Theorem 1.9(c).
The [detailed note](counterexamples-and-repairs.md) supplies the topology
and group calculation; an abstract FI-sharp example alone would not have
established a counterexample to the manifold theorem.

The rational, fixed-field and fixed-prime assertions are distinct. This
counterexample does not refute them. Degree assertions should also be read
as upper bounds: vanishing cohomology cannot have the prescribed exact degree.

## 2. Rank varieties: three separate failures

The paper studies matrices satisfying rank conditions on polynomials in
the matrix and claims all-size polynomiality for fixed homogeneous degrees.

First, arbitrary polynomials need not give homogeneous equations. For
$p(T)=T-1$, the rank conditions generally involve constants and different
degrees. There is then no inherited standard grading of the asserted kind.
One must specify homogeneous data, or formulate a different filtered problem.

Second, the claimed partial-injection maps need not preserve the conditions.
Put

$$N=\begin{pmatrix}1&1\\-1&-1\end{pmatrix},
\qquad A=\operatorname{diag}(N,N).$$

Then $A^2=0$. Its principal submatrix on indices $1,3$ is $I_2$, whose
square has rank two. Thus restriction takes a matrix with
$\operatorname{rank}(A^2)\le1$ outside that rank variety. The FI-sharp
construction used in the proof does not exist as asserted.

Third, this is accompanied by false numerical conclusions. The following
two tests keep the reduced variety and the minor scheme separate.

### The reduced variety

Over $\mathbb C$, let $X_n=\{A:\operatorname{rank}(A^2)\le1\}$.
For $n\ge2$ its quadratic vanishing ideal is precisely the span of the
second elementary characteristic coefficient $e_2(A)$. Consequently,

$$\dim\mathcal O(X_n)^{2}_{\mathrm{red}}=
\binom{n^2+1}{2}-1\qquad(n\ge2).$$

At $n=1$ the variety is the entire affine line, so the dimension is one,
whereas the displayed polynomial gives zero. This refutes Remark 5.3.3.
The proof checks both the vanishing of $e_2$ and the absence of other
quadratic equations; it is not an inference from sampled matrices.

### The minor scheme

Let $J_n$ be the ideal generated by the $2\times2$ minors of $A^2$.
Its generators have degree four, so the preceding degree-two calculation
does not apply to $\mathbb C[x_{ij}]/J_n$. In degree five, set
$g(n)=\dim(J_n)_5$. We independently constructed the integer coefficient
matrices of all variables times all these minors and computed their ranks
modulo the verified prime $1{,}000{,}003$.

An explicit commutator family supplies matching characteristic-zero upper
bounds. Thus the following are exact characteristic-zero ranks, not merely
probable ranks inferred from floating-point computation:

| $n$ | 0 | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| $g(n)$ | 0 | 0 | 4 | 73 | 541 | 2401 | 7876 | 21169 | 49393 |

There are at most $n^2\binom n2^2=O(n^6)$ products. If the quotient's
degree-five dimension were a polynomial for all $n$, its difference from
the ambient dimension $\binom{n^2+4}{5}$ would make $g(n)$ a polynomial.
The growth bound forces its degree to be at most six. But

$$\Delta^7g(0)=-6,\qquad\Delta^7g(1)=1.$$

This is impossible. The second calculation uses only positive sizes, so
discarding the empty matrix does not fix the assertion. This gives a
homogeneous, nonconstant counterexample to Theorem 5.3.2 itself.

Finite computations generally cannot prove a universal formula. Here they
disprove a polynomial claim because an independent degree bound makes a
finite contradictory difference decisive. The detailed note explains the
rank bounds and the exact checker. A separate proposed formula for every
$n$ is not needed for this disproof.

## 3. Repairs worth retaining

For integral generators, replace the all-size polynomial by the exact
maximum formula above and conclude eventual polynomiality. This preserves
the useful quantitative bound without asserting an invalid equality.

For rank varieties, ordinary determinantal conditions are preserved by
principal restriction and zero extension. Their homogeneous ideals and
radicals consequently support the original all-size argument. Positive
powers $p_\ell(T)=c_\ell T^{s_\ell}$, with $s_\ell\ge1$, retain zero-extension
compatibility but generally lose restriction compatibility. A direct
ordered-monomial argument gives eventual dimension polynomiality, with
degree at most $2d$ in grading $d$; the characteristic-zero character
conclusion additionally uses the general character-polynomial theorem.

For Theorem 6.4.3, reducing a submodule modulo a maximal ideal is not left
exact. The replacement proof first obtains polynomial field-cohomology
bounds from the forgetful configuration-space fibration. Integral homology
is degreewise finitely generated. The FI-sharp decomposition and field
detection bound the support of its coefficient groups. The natural
cohomological universal-coefficient sequence then gives the same support
bound over any commutative Noetherian ring. The proof uses finite generation
of annihilator ideals, but not finite Krull dimension. Its standard
topological premises and the arrow directions are stated explicitly in
the [repair note](counterexamples-and-repairs.md).

These are separately attributed reconstructions. They do not retroactively
validate the printed residue-field inference or the rank-variety maps.

## 4. Other qualifications from the broader reading

The filtered-module equivalence in Lemma 6.3.2(b) has a simple counterexample.
Take the constant FI-module $V_n=\mathbb Q$ and its submodule $W_n=0$ below
a positive threshold $N$, $W_n=\mathbb Q$ above it. The module $V$ has
stability type $(0,0)$; the transition $W_{N-1}\to W_N$ is not surjective.
The associated graded pieces need not inherit the bounds. The forward
implication from the pieces to the filtered object follows by exact
coinvariants and is the direction needed for the spectral-sequence proof.
Numerical injectivity-degree bounds also need truncation at zero when the
printed expression is negative. These changes do not reduce the stated
final rational stable range in the checked argument.

In Proposition 3.3.3 the displayed multiplicity argument cancels a common
summand without a finite-multiplicity hypothesis. Infinite cardinal
multiplicities cannot be cancelled in that way. The finite-dimensional
applications do not encounter this obstruction. The reviewers propose
repairs in the unrestricted setting, but this release does not certify
those additional proofs.

The Section 7 reviewers identify a non-proper Gysin-map issue in the
definition of the tautological classes, a normalization issue for their
pullbacks, and low-genus exceptions to an all-genus Torelli identification.
They propose compact-universal-curve normalization and restriction to the
genus-at-least-three tail. These observations concern proof interfaces;
they do not establish false intended eventual-polynomiality theorems.
The detailed external Johnson, Malcev and moduli inputs were not supplied
to the panel and have not been exhaustively verified here. These subsidiary
findings are reported for follow-up, not promoted to cleared repairs.

## 5. Evidence, coverage and attribution

The initial round comprised four Sol 6.2 high and four Fable 5.2 high
responses. Each received the full native author-version TeX and one of
four assigned scopes: foundations, algebraic applications, configuration
spaces, or moduli/Torelli/Lie applications. The two owner-supplied earlier
reports were withheld from that round. Fourteen grouped targets were
assigned to each model; this was not a separate audit of every numbered
claim or cited reference.

Two focused Astra 6.2 high responses then challenged the proposed
counterexamples and repairs using original passages, prior-report pages
and journal images. Three final checks completed: Sol and Fable independently
examined the degree-five and reduced rank arguments, and Sol checked the
configuration-space examples and replacement coefficient proof. Their full
records state their source limitations and the small expository additions
needed in the longer kernel calculation.
The coordinating assistant inspected the arguments and sources, performed
the exact checks, and selected the publication findings. Model agreement
is evidence of scrutiny, not a proof certificate or human peer review.

Seven initial responses used descriptive source labels rather than the
strict identifiers required by the validator. Their complete mathematical
text was preserved in explicit metadata-reconciled derivatives. The
original validation failures and unavailable-reference declarations remain
in the provenance record. No additional inference was used to reformat them.

Earlier Astra and Fable reports supplied for this investigation are the
discovery leads for the principal counterexamples. This review credits that provenance and
does not claim priority for the mathematics. A scoped literature search is
not evidence that no correction has ever appeared elsewhere.

The journal records receipt on 14 July 2013. External results used to
evaluate its original proof are subject to that historical cutoff;
present-day self-contained repairs are identified separately. Model
training cannot be restricted retrospectively to that date.

The public coverage and cost record includes every submitted attempt,
actual model identifiers, token usage, validation status and the conservative
usage-cost estimate of **$10.544866** under the owner's $25 total authorization.
All 13 submitted assessments returned complete mathematical responses. It is not
a provider invoice. The report, reading map, separate decision graph and
checks are available together so readers can assess the evidence themselves.


## Improvements and alternative arguments

Counterexamples and narrower replacements for the selected rank and configuration assertions; the rest of the paper is not cleared.

Improvement checks have their own scope. They do not change the recorded assessment of the original argument or certify the whole paper.

### Improvement cef-integral-minimum

**An exact maximum formula and eventual integral polynomiality**

**Independently checked within the stated scope** · repair

For every size the minimum is R(n)+max_p T_p(n), with a finite collection of binomial polynomials. It is polynomial for all sufficiently large sizes.

Affected claims: [thm:intro:openbetti](/papers/human-church-ellenberg-farb-fi-modules/files/decision-graph.roadmap.html?claim=thm%3Aintro%3Aopenbetti).

**Original statement:** The total integral minimum is polynomial for every size.

**Proposed statement:** For every size the minimum is R(n)+max_p T_p(n), with a finite collection of binomial polynomials. It is polynomial for all sufficiently large sizes.

**Statement comparison:** restricted scope.

Use finite abelian subset coefficients. Primary counts add, their total minimum is a maximum, and a finite maximum of polynomials has one controlling polynomial on a tail.

Proposed by Astra reconstruction, with owner-supplied discovery leads (gpt-6.2-astra; 2026-11-20). [Candidate argument](counterexamples-and-repairs.md).

**Inputs:**

- FI-sharp subset classification; additional classical topology inputs are listed in the repair note.

**Limits:**

- A validity and fidelity entry may refer to the same scoped response; these are not separate votes.
- The final exposition includes coordinator clarifications. No formal verification or new mathematical priority claim.
- The positive-power character extension additionally uses the general FI character theorem; it is not part of the standalone dimension repair.

**Validity check:** sol62-cef-verify-repairs (gpt-6.2-sol; fresh context same provider). Supports: The named repair, its hypotheses and its difference from the original claim; supplied source limitations remain explicit. [Check record](sol62-verify-repairs.md).

**Fidelity check:** sol62-cef-verify-repairs (gpt-6.2-sol; fresh context same provider). Supports: The named repair, its hypotheses and its difference from the original claim; supplied source limitations remain explicit. [Check record](sol62-verify-repairs.md).

### Improvement cef-general-ring

**A universal-coefficient proof without the Krull-dimension restriction**

**Independently checked within the stated scope** · weaker hypotheses

The same support and generation bounds hold over every commutative Noetherian ring, with the stated manifold hypotheses.

Affected claims: [thm:ConfMFIsharpfg](/papers/human-church-ellenberg-farb-fi-modules/files/decision-graph.roadmap.html?claim=thm%3AConfMFIsharpfg).

**Original statement:** Bounded FI-sharp generation over a Noetherian ring of finite Krull dimension.

**Proposed statement:** The same support and generation bounds hold over every commutative Noetherian ring, with the stated manifold hypotheses.

**Statement comparison:** weaker hypotheses.

Field growth and finite integral homology detect FI-sharp support. The natural integral UCT and exact subset extraction transfer support to arbitrary coefficients; Noetherianity controls annihilator modules.

Proposed by Astra reconstruction, with owner-supplied discovery leads (gpt-6.2-astra; 2026-11-20). [Candidate argument](counterexamples-and-repairs.md).

**Inputs:**

- FI-sharp subset classification; additional classical topology inputs are listed in the repair note.

**Limits:**

- A validity and fidelity entry may refer to the same scoped response; these are not separate votes.
- The final exposition includes coordinator clarifications. No formal verification or new mathematical priority claim.
- The positive-power character extension additionally uses the general FI character theorem; it is not part of the standalone dimension repair.

**Validity check:** sol62-cef-verify-repairs (gpt-6.2-sol; fresh context same provider). Supports: The named repair, its hypotheses and its difference from the original claim; supplied source limitations remain explicit. [Check record](sol62-verify-repairs.md).

**Fidelity check:** sol62-cef-verify-repairs (gpt-6.2-sol; fresh context same provider). Supports: The named repair, its hypotheses and its difference from the original claim; supplied source limitations remain explicit. [Check record](sol62-verify-repairs.md).

### Improvement cef-determinantal

**Retain all-size formulas for ordinary determinantal conditions**

**Independently checked within the stated scope** · repair

For p_l(T)=c_l T, the homogeneous minor ideals and their radicals support all-size dimension formulas; in characteristic zero the FI-sharp character formula applies.

Affected claims: [theorem:sections:polygrowth](/papers/human-church-ellenberg-farb-fi-modules/files/decision-graph.roadmap.html?claim=theorem%3Asections%3Apolygrowth).

**Original statement:** All-size dimension and character formulas for arbitrary polynomial rank conditions.

**Proposed statement:** For p_l(T)=c_l T, the homogeneous minor ideals and their radicals support all-size dimension formulas; in characteristic zero the FI-sharp character formula applies.

**Statement comparison:** restricted scope.

Principal restriction and zero extension preserve minors, and pullback preserves radicals. Partial inversion gives an FI-sharp quotient generated in each fixed degree.

Proposed by Astra reconstruction, with owner-supplied discovery leads (gpt-6.2-astra; 2026-11-20). [Candidate argument](counterexamples-and-repairs.md).

**Inputs:**

- Finite-dimensional linear algebra and the supplied FI-sharp classification; positive-power dimension proof uses the explicit ordered-monomial argument.

**Limits:**

- A validity and fidelity entry may refer to the same scoped response; these are not separate votes.
- The final exposition includes coordinator clarifications. No formal verification or new mathematical priority claim.
- The positive-power character extension additionally uses the general FI character theorem; it is not part of the standalone dimension repair.

**Validity check:** sol62-cef-verify-rank (gpt-6.2-sol; fresh context same provider). Supports: The named repair, its hypotheses and its difference from the original claim; supplied source limitations remain explicit. [Check record](sol62-verify-rank.md).

**Fidelity check:** sol62-cef-verify-rank (gpt-6.2-sol; fresh context same provider). Supports: The named repair, its hypotheses and its difference from the original claim; supplied source limitations remain explicit. [Check record](sol62-verify-rank.md).

**Validity check:** fable52-cef-verify-rank (claude-fable-5-2; fresh context other provider). Supports: The named repair, its hypotheses and its difference from the original claim; supplied source limitations remain explicit. [Check record](fable52-verify-rank.md).

**Fidelity check:** fable52-cef-verify-rank (claude-fable-5-2; fresh context other provider). Supports: The named repair, its hypotheses and its difference from the original claim; supplied source limitations remain explicit. [Check record](fable52-verify-rank.md).

### Improvement cef-positive-powers

**Eventual dimension formulas for positive monomial conditions**

**Independently checked within the stated scope** · repair

For positive monomials c_l T^(s_l), the minor scheme and reduced variety have eventually polynomial dimensions in each fixed grading d, of degree at most 2d, over any field.

Affected claims: [theorem:sections:polygrowth](/papers/human-church-ellenberg-farb-fi-modules/files/decision-graph.roadmap.html?claim=theorem%3Asections%3Apolygrowth).

**Original statement:** All-size polynomial dimensions for arbitrary polynomial rank conditions.

**Proposed statement:** For positive monomials c_l T^(s_l), the minor scheme and reduced variety have eventually polynomial dimensions in each fixed grading d, of degree at most 2d, over any field.

**Statement comparison:** restricted scope.

Zero-extension pullbacks give a dual FI-submodule. Leading monomials under increasing injections have finitely many patterns and upward gap sets; finite orthant counting gives eventual dimensions. No explicit uniform threshold is asserted.

Proposed by Astra reconstruction, with owner-supplied discovery leads (gpt-6.2-astra; 2026-11-20). [Candidate argument](counterexamples-and-repairs.md).

**Inputs:**

- Finite-dimensional linear algebra and the supplied FI-sharp classification; positive-power dimension proof uses the explicit ordered-monomial argument.

**Limits:**

- A validity and fidelity entry may refer to the same scoped response; these are not separate votes.
- The final exposition includes coordinator clarifications. No formal verification or new mathematical priority claim.
- The positive-power character extension additionally uses the general FI character theorem; it is not part of the standalone dimension repair.

**Validity check:** sol62-cef-verify-rank (gpt-6.2-sol; fresh context same provider). Supports: The named repair, its hypotheses and its difference from the original claim; supplied source limitations remain explicit. [Check record](sol62-verify-rank.md).

**Fidelity check:** sol62-cef-verify-rank (gpt-6.2-sol; fresh context same provider). Supports: The named repair, its hypotheses and its difference from the original claim; supplied source limitations remain explicit. [Check record](sol62-verify-rank.md).

**Validity check:** fable52-cef-verify-rank (claude-fable-5-2; fresh context other provider). Supports: The named repair, its hypotheses and its difference from the original claim; supplied source limitations remain explicit. [Check record](fable52-verify-rank.md).

**Fidelity check:** fable52-cef-verify-rank (claude-fable-5-2; fresh context other provider). Supports: The named repair, its hypotheses and its difference from the original claim; supplied source limitations remain explicit. [Check record](fable52-verify-rank.md).
