# Church–Farb: what fails, what survives, and what can be repaired

Igor Rivin · Math Referee · 21 November 2026  
Public scoped report. Prepared with independent model assessments,
focused challenges and coordinator source checks. These are AI-assisted mathematical
adjudications, not an independent human review or a formal proof certificate.

**Paper:** Thomas Church and Benson Farb, *Representation theory and homological
stability*, Advances in Mathematics 245 (2013), 250–314,
DOI 10.1016/j.aim.2013.06.016; final author version arXiv:1008.1368v3.

The checked portions contain several false statements, not merely typographical
errors. They include nonuniform closure and cancellation assertions, implications
about mixed stability, an all-rank stability condition, and the prime-two
coefficient assertion in Theorem 8.4. These failures do **not** refute the entire
representation-stability framework. Finite-dimensional uniform statements have
substantially different hypotheses; important ones admit replacement arguments.
The pure-braid argument survived the scoped checks described below.

All eleven submitted requests have returned. Eight contain completed mathematical
assessments; three initial Fable attempts exhausted their output allowance.
There are no pending model jobs and no further paid calls scheduled. Conservative
usage accounting is **$11.963289 of the authorized $25**. The separate uncached
Batch-tier estimate is **$10.344260**. Both include failed attempts; neither is an
invoice. The source packets, original answers and failures remain preserved.


[Reviewed author edition](../files/paper.pdf) · [Download this report](../files/church-farb-scoped-review.pdf) · [Exact-check bundle](../files/church-farb-checks.zip)

The uniform cancellation repair was examined in both focused Astra packets. The eventual and polynomial homology reconstructions were checked by the coordinator against the source and Astra proposals; the full replacements have not received a separate fresh final audit. Their improvement records retain that distinction.

## 1. Scope and versions

The complete final-author TeX was supplied to the initial panel. Its hash is
`5f5f379bd8e23f2d3911be054310baedf23dc35a7b2693e6d3e96bef44654698`.
The supplied journal PDF, retained privately, has hash
`b791b01ac888a0e74341bbf2f2931ceb7335f0ef02fd52e46d1e442573ba08d2`.
Selected journal passages retain the disputed wording: p.269 (mixed stability and
Theorem 3.1), pp.275–276 (Theorem 3.5), pp.286 and 289–291 (Lie coefficients and
simple stability), and p.311 (Theorem 8.4). This is a scoped edition comparison,
not a complete collation of all 65 journal pages.

The journal records receipt on **19 January 2012**. The initial packets had
already frozen a conservative first-arXiv cutoff, 7 August 2010; their prompts
were not retroactively changed. The focused packets use the verified received
date. In particular Hemmer's 2010 online article and Putman's August 2011 preprint
are eligible inputs. Later results are retrospective context, not justification
of the original proof. Model training knowledge cannot literally be time-filtered.

| Part | Checked outcome | Boundary |
| --- | --- | --- |
| Theorem 3.1, nonuniform SL clauses | False | Explicit finite-dimensional examples |
| Theorem 3.1, finite uniform clauses | Supported on the specified classical operation formulas | External formula proofs not exhaustively audited |
| Theorem 3.5 and Remark 3.6 | Nonuniform conclusions false; projection proof invalid | Finite uniform replacement below |
| Mixed/dual implications after Definition 2.16 | False in the stated generality | The definition itself is not being rejected |
| Theorem 4.1, pure braids | No defect found in the checked argument | Earlier cohomology inputs not all reproved |
| Theorem 5.3 | Compatible-coefficient version has a replacement argument | Clarify what the transition maps mean |
| Theorem 5.6 and Corollaries 5.7–5.9 | Literal all-rank conclusions fail | Polynomial and eventual versions can survive |
| Theorem 8.4 | False at prime two as stated; exterior-cube table reversed | Corrected full periodicity and map claims remain partly open in this audit |
| Other claims, especially §7 and other §8 applications | Not finally adjudicated | Initial observations are not accepted findings |

## 2. Nonuniform closure and cancellation

Let $E_n=\mathbb Q^n$ and let $a_n$ alternate between 1 and 2. All representations
in the following finite-dimensional examples are rational algebraic
representations, so restricting to the intended category does not remove them.

For $SL_n$, $E_n^*$ has label $(1^{n-1})$, which escapes every fixed label as
$n$ grows. Consequently $E_n$ and $(E_n^*)^{\oplus a_n}$ are multiplicity stable,
but their tensor product has trivial multiplicity $a_n$, since
$E_n\otimes E_n^*=\mathbf1\oplus\mathfrak{sl}(E_n)$. Similarly,
$\operatorname{Sym}^2(E_n\oplus(E_n^*)^{\oplus a_n})$ has exactly $a_n$
invariants for $n\ge3$. Taking one-box inner or outer Schur functors carries
these counterexamples to the other nonuniform closure clauses.

A second example exposes unbounded multiplicity rather than oscillation. Set
$X_n=\bigwedge^*E_n$. Its two trivial summands and its distinct remaining exterior
powers give stable multiplicity at each fixed label, while

$$\dim (X_n\otimes X_n)^{SL_n}=4+(n-1)=n+3.$$

The source of the failure is that moving labels can produce a fixed output:
removing a full column changes an $SL_n$ label. Pointwise stabilization does not
bound all contributors to a fixed tensor-product constituent.

For Theorem 3.5 take $W_n=\operatorname{Sym}^n E_n$ and
$V_n=\mathbf1^{\oplus a_n}$. This works for $SL$, $GL$ and $Sp$. Both $W_n$ and
$V_n\otimes W_n$ are nonzero and every fixed multiplicity eventually vanishes;
$V_n$ is not multiplicity stable. Even a fixed nonzero limiting factor is not
enough: for $SL_n$, take $W_n=E_n$ and alternate $V_n$ between $\mathbf1$ and
$\mathfrak{sl}(E_n)$. The product always contains one standard constituent, and
its other labels escape. The unknown factor still fails to stabilize.

Remark 3.6 also fails pointwise. Alternate the trivial module with the traceless
adjoint for $SL_n$ or $GL_n$. The adjoint square has one invariant. Every other
fixed ordinary label is eventually excluded by scalar degree for $GL_n$, or
weight-lattice degree modulo $n$ for $SL_n$. Thus the squares stabilize
pointwise while the original trivial multiplicity oscillates. This is a
weight-lattice argument, not an assumption about roots of unity in $\mathbb Q$.

There is a separate infinite-multiplicity obstruction: a countable direct sum of
trivial modules absorbs either one or two copies under tensoring. Ordinary
representation-ring cancellation is an integer-coefficient argument and cannot
be applied to infinite cardinal multiplicities.

## 3. A replacement for finite uniform cancellation

The printed rank projection is not a ring homomorphism. For example, its
$SL_3\to SL_2$ projection kills $E_3^*$ and the adjoint but sends
$E_3\otimes E_3^*=\mathbf1\oplus\mathfrak{sl}_3$ to $\mathbf1$.
The compactness argument using high-size labels as an ideal fails for the same
reason: a representation and its dual can produce the unit.

Nevertheless, **finite-dimensional uniform cancellation is valid**, assuming
the classical stable tensor-product tables for each fixed pair of labels.

Suppose $W_n$ and $C_n=V_n\otimes W_n$ are uniformly multiplicity stable and
$W_n$ is eventually nonzero. Their stable supports $B,F$ are finite. Fix a
constituent $\mu\in B$. For every constituent $\lambda$ of $V_n$, the Cartan
constituent $\lambda+\mu$ occurs in $C_n$. Hence every $\lambda$ belongs to the
fixed finite candidate set obtained by subtracting $\mu$ from members of $F$.
For $GL_n$, keep both the polynomial partition and the determinant exponent;
the paper's ordinary labels repeat their last entry with rank.

Beyond the finitely many relevant product-table thresholds, transport the
decompositions at two ranks to a single larger rank. They satisfy $xw=yw$ there,
with $w\ne0$. The finite-dimensional representation ring embeds by characters
in the integral group ring of a free weight lattice, a Laurent polynomial
domain. Thus $x=y$. This proves uniform stability of $V_n$.

The same argument treats uniformly stable squares: $2\lambda$ bounds the
possible input support, and $(x-y)(x+y)=0$ implies $x=y$ for effective classes.
The proof uses bounded support before cancellation; it does not validate the
printed projection or any nonuniform extension. Fixed-label tensor, restriction
and Schur-functor formulas also support finite uniform closure. Their imported
classical proofs and sharp ranges were not all re-audited here.

## 4. Mixed stability and duals

The assertions following Definition 2.16 must distinguish ordinary labels from
mixed labels and distinguish multiplicities from transition maps.

Here is an example with actual injective, orbit-generating maps. Put
$B_n=\mathbf1\oplus E_n^{\oplus a_n}\oplus\operatorname{Sym}^2E_n$ and
$T_n=\det(E_n)\otimes B_n$. Write $E_{n+1}=E_n\oplus\mathbb Qz$.
At a one-to-two increase in $a_n$, use

$$ (q,v,w)\longmapsto(q,v,qz,w); $$

at a two-to-one decrease, use

$$ (q,v_1,v_2,w)\longmapsto(q,v_1,w+v_2z). $$

Old symmetric tensors and tensors involving $z$ are disjoint, proving injectivity.
At an increase, the orbit of $(1,z)$ generates both the trivial and new standard
summands by taking differences of translates; the pure old summands supply the
others. At a decrease the pure old summands already generate. Twist these maps
by the canonical determinant-line inclusion.

Every fixed mixed-label multiplicity of $T_n$ is eventually zero, because its
highest weights have a long middle of ones. But the fixed ordinary label
$\det(E_n)\otimes E_n$ has multiplicity $a_n$. Thus mixed stability does not
imply ordinary $GL$ stability under the printed naming convention. The analogous
implication for $SL$ is not refuted: the empty negative mixed label tests each
ordinary normalized $SL$ label.

For the dual assertion use
$Z_n=\mathbf1\oplus(E_n^*)^{\oplus a_n}\oplus\mathfrak{sl}(E_n)$.
The increase map inserts $qz^*$ as the second covector; the decrease map inserts
$z\otimes f_2$ into the lower-left block of the new adjoint matrix. Old blocks
and the new row are disjoint, and the same orbit argument proves injectivity
and generation. Only the trivial ordinary label persists in $Z_n$, whereas
$Z_n^*$ has standard mixed-label multiplicity $a_n$. Ordinary stability therefore
does not imply mixed stability of the duals. Duality does interchange the two
mixed labels of an individual simple module; natural transposes also reverse
the arrows and are not automatically forward stabilization maps.

## 5. Lie homology: separate the eventual theorem from the all-rank claim

### Coefficients and the eventual statement

Definition 5.2 must require that the coefficient transition $f_n$ intertwines
the Lie transition $i_n$:

$$f_n(xm)=i_n(x)f_n(m).$$

If “maps” already means morphisms of these modules, this is an implicit
convention to make explicit. If only graded equivariant linear maps are meant,
there is a counterexample: take $L_n=\mathbb Qz$ in degree one,
$M_n=\mathbb Qu\oplus\mathbb Qv$ in degrees zero and one, identity transitions,
trivial group actions, and alternate $zu=v$ with $zu=0$. All representation
pieces are strongly stable, but $H_0(L_n;M_n)^1$ alternates between zero and
$\mathbb Q$. Trivial and natural adjoint coefficients already have compatibility.

With compatible maps, finite-dimensional graded pieces and the finite uniform
operation inputs, the eventual theorem has a replacement proof. Fix total degree
$d$. Its Chevalley–Eilenberg complex has only finitely many chain terms. In the
stable range write each chain term as a sum $S_{\lambda,n}\otimes A^i_{\lambda,n}$.
Strong stability and matching branching multiplicity one make each transition
$u_{\lambda,n}\otimes T^i_{\lambda,n}$ with $T^i$ an isomorphism. The differential
acts on the multiplicity spaces; the chain square makes the $T^i$ chain
isomorphisms. Homology therefore has injective, generating, type-preserving maps
and constant finite multiplicities. Homology injectivity is proved here; it
does not follow from arbitrary injections of complexes.

For reverse implications use the finite-degree Euler identity
$\sum_i(-1)^i[C_i^d]=\sum_i(-1)^i[H_i^d]$ and induction on the Lie degree.
With trivial or adjoint coefficients the unknown Lie piece occurs with the unit
coefficient. For a general compatible coefficient take its least eventually
nonzero grade and use the finite uniform cancellation proved above. This supports
the intended finite eventual equivalence, relative to the specified operation
and branching inputs. The symplectic multiplicity-only reverse statement has the
same Euler argument with trivial coefficients; no symplectic map theorem follows.

### Failure at the first nonzero rank

Definition 5.5 demands orbit generation at every rank. But
$\bigwedge^2\mathbb Q=0$ cannot generate $\bigwedge^2\mathbb Q^2\ne0$.
This directly contradicts the stated conclusions of:

- Corollary 5.7 at free-Lie degree two;
- Corollary 5.8 at nilpotence class one and homological degree two;
- Corollary 5.9 at class one and homological degree one: the image in
  $E_2\otimes E_2$ generates its symmetric square, missing its exterior square.

These are canonical inclusions, not artificial transition maps. They represent
one definition-level defect, not three independent incorrect homology computations.
For Theorem 5.6 the abelian Lie algebra $L_n=E_n$, trivial coefficients and
$H_2=\bigwedge^2E_n$ give the same failure under the usual eventual type-preserving
hypothesis. Reading constant-tail type preservation as an extra all-rank
hypothesis changes this test at rank one, but does not rescue the three explicit
corollaries or the proof's asserted all-rank equivalences.

There is a further small-rank error in identifying SIV with the unipotent test
IV′. On $E_n\oplus\bigwedge^2E_n$, change only the first map to
$e_1\mapsto(e_1,e_1\wedge e_2)$. It is injective, orbit-generating, and its image
is fixed by the last-column unipotents. The old standard line generates two
new simple summands, so SIV fails.

### A useful changed conclusion

Replace full orbit generation at a new type's first rank by injectivity,
constant polynomial Schur multiplicities wherever defined, and preservation of
each existing polynomial type. For such a sequence the old space identifies with
the new space's weight-zero subspace for scaling the last coordinate. All weights
are nonnegative; tensor and exterior operations preserve this identification,
and compatible differentials preserve weights. It passes to homology.

Free Lie words, nilpotent truncations and their trivial/adjoint chain complexes
have precisely this property: count occurrences of the new generator, and retract
by killing it. This supplies a checked forward replacement for the applications.
At a fixed degree only finitely many shapes occur, so eventual full strong
stability follows once all those shapes exist.

This changed condition does not by itself prove the printed five-way equivalence.
A reverse all-rank claim also needs polynomial control of the Lie maps and
coefficient detection at the tested ranks. Astra supplied a further formal-series
Euler argument using the whole coefficient module; it is retained as a separate
conditional proposal, not needed for the forward replacement accepted here.

## 6. Prime two and the exterior-cube table

Write $A_g=H_1(\operatorname{Mod}_{g,1}(2);\mathbb Z)$. Sato's 2008 Proposition
4.9, for the same one-boundary-component group and $g\ge3$, states

$$A_g\cong(\mathbb Z/8)^{2g}\oplus
(\mathbb Z/4)^{\binom{2g}{2}}\oplus
(\mathbb Z/2)^{\binom{2g}{3}}.$$

Its exponent is eight, so it is not an $\mathbb F_2$ vector space. The theorem
statement and group conventions were checked; Sato's whole proof was not audited.

A separate obstruction needs less: on a level-two symplectic matrix
$M=I+2X$, write $X$ in blocks $(A,B;C,D)$. Then

$$f_i(M)=B_{ii}+2\sum_j A_{ij}B_{ij}\pmod4$$

is a surjective homomorphism. The symplectic relation makes $B,C$ symmetric and
$D=A^t$ modulo two. In a product, the defect in the linear term equals twice the
polarization of the quadratic term; their sum vanishes modulo four. The
transvection $I+2E_{i,g+i}$ maps to one. Pullback along the standard surjection
from the level mapping-class group gives a $\mathbb Z/4$ quotient. The surjectivity
of the mapping-class action is an explicit imported premise. This already rules
out an $\mathbb F_2$-vector-space structure.

Integral homology, $A_g/2A_g$, and the simple factors of a finite
$\mathbb Z[Sp_{2g}(\mathbb F_2)]$-module are different objects. Changing between
them changes the theorem. Putman's relevant extension is integral; tensoring it
with $\mathbb F_2$ is not left exact.

The displayed exterior-cube table has its congruence cases interchanged. For
$H=\mathbb F_p^{2g}$ let $\epsilon(x)=x\wedge\omega$ and let $\delta$ be symplectic
contraction. Then

$$\delta\epsilon=(g-1)I.$$

For $g\ge3$ with $g\equiv1\pmod p$, the injected $H$ lies in $\ker\delta$, and
the quotient by that kernel is another $H$. The printed table lists only one.
This filtration alone disproves that row without assuming a formula for the
remaining simple quotient's dimension.

The complete corrected class is

$$[\bigwedge^3H]=[L(\omega_3)]+(1+[g\equiv1\pmod p])[H].$$

The source route uses the Weyl-module formula in Baranov–Suprunenko Theorem 2.1,
its characteristic-two extension in Corollary 2.9, the characteristic-independent
character identity $[\bigwedge^3H]=[W(\omega_3)]+[H]$, and restriction of
$p$-restricted simple modules to the finite symplectic group. These premises and
conventions are explicit; their complete proof histories are not newly certified.
Fable also supplied an algebraic-group weight argument for the simple quotient.
It must be done over an algebraic closure before finite-group restriction, since
the split torus of the finite group at $p=2$ cannot justify that weight argument.

With Putman's integral sequence and the quoted remaining factor calculations,
the corrected integral $\mathbb Z[G]$ lists are $(2,5,2,1)$ for odd $g$ and
$(4,4,2,1)$ for even $g$, in the order
$\mathbf1,H,L(\omega_2),L(\omega_3)$. These are conditional factor lists, not
lists for $A_g/2A_g$, and they do not establish stabilization-map properties.
For odd primes, a central $-I$ argument repairs the exponent inference in the
equivariant extension: multiplication by $p$ would give a map from an even
central-character quotient to an odd central-character kernel, so is zero.
Naturality of the required extensions under stabilization still needs checking
before asserting the full corrected periodic representation-stability theorem.

**Later work matters.** Tudur Lewis's *Representation stability for the first
homology of congruence subgroups*, arXiv:2507.20902v2 (1 June 2026), states
corrected field-coefficient and uniform-periodicity results. Those statements
were inspected, not their full proofs. They do not validate CF's original
argument, and their existence prevents a claim that this issue has never been
noticed or that the corresponding modern result remains wholly unsettled.

## 7. Successful checks, coverage and remaining work

Sol and Fable independently completed the braid packet. The argument uses at most
$2i$ active labels in degree $i$, followed by induction from a fixed symmetric
group and Hemmer's stable range $n\ge2k$, giving $n\ge4i$. Hemmer's original
four-page article and its induction/Pieri proof were inspected. Earlier
Orlik–Solomon inputs and every downstream twisted-cohomology assertion were not
independently re-established.

| Round | Model and effort | Completed mathematical assessments | Limit |
| --- | --- | ---: | --- |
| Independent | Sol 6.2, high | 4 / 4 | Four assigned scopes; findings still needed adjudication |
| Independent | Fable 5.2, high | 1 / 4 | Foundations, Lie and modular attempts hit the 65,536 combined token cap |
| Focused | Astra 6.2, max | 2 / 2 | Non-blind foundations and Lie challenges |
| Focused | Fable 5.2, high | 1 / 1 | Non-blind modular challenge; partial supplied context and metadata recovery |

The focused Fable answer's mathematical fields were preserved verbatim while
source aliases and partial-access declarations were reconciled in a separate
derivative. The original validation failure and its cost remain. Its packet
mistakenly included a forward-reference page instead of Putman's Theorem 7.8;
the coordinator subsequently checked the correct PDF page 29. That later check
is not attributed to Fable. Missing theorem pages and full antecedent proofs
remain different coverage questions.

Exact local checks passed for rational tensor and symmetric-square invariants at
ranks 2–7, exterior maps at genera 2–8 over four prime fields, and 400 products of
level-two symplectic matrices modulo eight. Their script and result hashes are
preserved. Finite checks support conventions and examples; the general arguments
above do the mathematical work.

The selected findings and finite uniform/forward polynomial repairs have now
been consolidated. Remaining mathematical obligations are the full modular
stabilization-map repair, exhaustive auditing of the classical imported inputs,
and disposition of broader initial observations outside these focused targets.
None is silently marked as verified. This publication preserves those limits; it does not claim a complete audit of the paper.
