Algebra, representation theory and topology

FI-modules and stability for representations of symmetric groups

Thomas Church, Jordan S. Ellenberg and Benson Farb

54 pages114 claims indexed66 referencesManuscript 2014-10-30

claude-fable-5-2 high: cef-verify-rank

Reviewer record, not the publication verdict. This preserves a completed model response, including unaccepted proposals and its stated source limitations. The curated report and repair note specify which conclusions were adopted. Seven initial source-label failures were reconciled in explicit derivatives; native statuses and original access declarations remain in the provenance record.

Completed visible assessment, preserved as a readable check record. It is a model assessment, not a formal certificate. This document contains no private reasoning stream.

The candidate's mathematics survives a line-by-line challenge on both targets. V1: the degree-five kernel exhaustion is correct over C once three terse steps are written out (conjugation-invariance of the kernel, the compressed block trace, the first-order expansion of G^{-1}BG). The cases n=2 and n=3 check directly (kernel dimensions 0 and 8), the centralizer is the scalars, and g(n)=(n^2-1)binom(n,2)^2+1 holds for n>=2 with g(0)=g(1)=0. This refutes the all-n scheme statement of Theorem 5.3.2 at p=T^2, r=1, d=5. V2: the reduced quadratic computation is correct and refutes Remark 5.3.3 at n=1. The determinantal all-n repair is valid from the supplied FI#-classification alone. The positive-monomial repair yields eventual dimension polynomiality over any field by a self-contained leading-index argument, while its eventual character statement stays conditional on an unsupplied theorem. Nothing here was machine-checked.

V1

Checked statement: Candidate REX3: over C, for p=T^2, r=1, every linear relation among the products x_ij * (2-minor of A^2) is of the form L(A)=[D(A),C]. Hence dim ker(mu)=m^2-1 and g(n)=dim (J_n)_5=(n^2-1)*binom(n,2)^2+1 for n>=2, with g(0)=g(1)=0, m=binom(n,2).

Returned assessment: no_issue_found

I reworked every step by hand and tried to break it; I found no error. The formula is proved over C for all n>=2, and over any characteristic-zero field since the rank of mu is unchanged by field extension. Three steps are terse in the candidate and are expanded below; none changes the conclusion.

Setup. mu(L)(A)=tr(L(A)R(A)) with R(A)=wedge^2(A^2). The entries of R are exactly the 2-minors of A^2, so im(mu) is the span of all x_ij times a generator, which is (J_n)_5; the domain has dimension n^2 m^2. Commutators lie in the kernel: tr([D,C]R)=tr(C[R,D])=0 because D(A) commutes with wedge^2(A) and R=(wedge^2 A)^2. The kernel is stable under L -> L^g, L^g(A)=G^{-1}L(gAg^{-1})G with G=wedge^2 g, because R(gAg^{-1})=G R(A) G^{-1}. The candidate uses this silently ('works in every eigenbasis').

(i) Diagonal and shear reduction. For h diagonal, tr(L(h)R(h)) is the sum over k and i<j of (L(E_kk)){ij,ij} h_k h_i^2 h_j^2. These monomials are pairwise distinct (h_i^3 h_j^2 when k=i, three variables otherwise), so L(h) has zero exterior diagonal. For I != J impose h_i=h{j(i)} along a matching of I\J to J\I. The shear g=1+tsum E_{j(i),i} then commutes with h, its inverse is 1-tsum E_{j(i),i} (the sum squares to zero), and L^g(h)=G^{-1}L(h)G has zero diagonal. One checks e_I^* G^{-1}=e_I^, since g^{-1} only adds vectors with labels in J\I. G e_I is e_I plus lower terms, with top t-coefficient +-e_J (degree 1 for one replacement, degree 2 for disjoint pairs). Vanishing of that polynomial in t gives e_I^ L(h) e_J=0 on the imposed subspace. For one replacement this subspace is the hyperplane chi_I=chi_J. For disjoint pairs the two matchings give two codimension-2 subspaces meeting in codimension 3, so their sum is the hyperplane chi_I=chi_J. Hence L(h){I,J}=c{I,J}(chi_I-chi_J)(h), that is L(h)=[D(h),C] on diagonals. Correct.

(ii) Support of B=L(E_ab) after subtracting [D(.),C]. With g=1+tE_ab, g h g^{-1}=h+t(h_b-h_a)E_ab, so L^g(h)=t(h_b-h_a)G^{-1}BG. By (i) its (I,J) entry is a multiple of chi_I-chi_J, and this is proportional to h_b-h_a only when {I\J,J\I}={{a},{b}}. So G^{-1}BG has zero diagonal and is supported on entries linking u_i=e_b^e_i and w_i=e_a^e_i, for all t (polynomial in t). Here G=1+tD with D=D(E_ab)=sum_i w_i u_i^, and D^2=2wedge^2(E_ab)=0, so G^{-1}=1-tD and G^{-1}BG=B+t[B,D]-t^2 DBD. Writing B=sum_i(beta_i w_i u_i^+gamma_i u_i w_i^) gives [B,D]=sum_i gamma_i(u_i u_i^-w_i w_i^). Zero diagonal forces gamma_i=0. So L(E_ab) sends e_b^e_i to l_{ab,i} e_a^e_i and nothing else. Correct. For n=2 there is no i, so L=0.

(iii) Cycle identities and potentials. For A a 2- or 3-cycle on a coordinate set S (diagonalizable over C), applying (i) to L^g, with g the change to an eigenbasis on S and identity elsewhere, shows L(A) has zero diagonal in the exterior eigenbasis. The blocks U_i=span{e_x^e_i: x in S} (i outside S) and wedge^2(span S) are spanned by eigenbasis wedges, and the complementary eigen-wedges span the same coordinate subspace as the complementary standard wedges. So the coordinate projection P onto the block is the eigenbasis projection. D(A) is diagonal on the block in the eigenbasis, hence tr(P L(A) D(A)^p P)=0, and this compressed trace may be computed in the standard basis. I recomputed all three.

  • A=E_ab+E_ba, p=1 on U_i: l_{ab,i}+l_{ba,i}=0.
  • A=E_ab+E_bc+E_ca, p=2 on U_i: l_{ab,i}+l_{bc,i}+l_{ca,i}=0.
  • Same A on the internal block: D(A) sends e_a^e_b to -e_b^e_c and so on cyclically, D(A)^2 sends e_a^e_b to e_c^e_a, and L(A)D(A)^2 has diagonal -(l_{bc,a}, l_{ca,b}, l_{ab,c}); so l_{ca,b}+l_{ab,c}+l_{bc,a}=0. The first two give b_{xi} with l_{ab,i}=b_{bi}-b_{ai} (base vertex). Then d_{ab}=b_{ab}-b_{ba} satisfies d_{ab}+d_{bc}+d_{ca}=l_{ca,b}+l_{ab,c}+l_{bc,a}=0, so d_{ab}=t_a-t_b. Shifting b_{ai} by t_i symmetrizes without changing l. The diagonal C with entries b_{ai}+t_i then has [D(E_ab),C]u_i=(c_{bi}-c_{ai})w_i=L(E_ab)u_i and [D(h),C]=0. Exhaustion follows.

Small n. n=2: W is a line, R=det(A)^2, kernel 0=m^2-1, g=4. n=3: the spectator triangle identity is vacuous; the six unknowns l_{ab,c} are cut to three by antisymmetry and to two by the internal identity, which equals m-1. Total kernel is 6 (off-diagonal C) + 2 = 8, so g=81-8=73. Cross-check: for n=3, R is adj(A)^2 up to the identification of wedge^2 V with V* tensor det, and the 8 relations are the entries of [A,adj(A)^2]=0. n>=4: all three identities are available.

(iv) Centralizer. Commuting with all D(h) forces C diagonal (pair weights are distinct). Commuting with D(E_ab) gives c_{bi}=c_{ai}. The pair graph is connected, so C is scalar; equivalently wedge^2 V is irreducible for gl_n. So dim ker(mu)=m^2-1 and g(n)=n^2 m^2-(m^2-1)=(n^2-1)m^2+1.

Consequence. The formula gives 4, 73, 541, 2401, 7876, 21169, 49393, 103681 for n=2..9, which I recomputed. Let G(n) be the degree-6 polynomial (n^2-1)binom(n,2)^2+1; G(0)=G(1)=1 but g(0)=g(1)=0. A direct evaluation gives Delta^7 g(0)=-84+2555-18935+50421-55132+21169=-6. More simply, the degree-5 quotient dimension equals binom(n^2+4,5)-G(n) for n>=2, but at n=1 it is 1 while that polynomial gives 0. So the all-n statement of Theorem 5.3.2 (source lines 1811-1815) is false for p=T^2, r=1, d=5, also if n=0 is excluded, and the character version fails at the identity.

Premises used

  • Base field C: infinite field for coefficient extraction, and diagonalizable 2- and 3-cycles.
  • (J_n)_5 is spanned by variable-times-generator products; no independence of the generators is assumed.
  • For n=0,1 there are no 2-minors, so J_n=0.

Missing inputs

  • The prior report's rank table and any computation behind it were not supplied. The numbers above are certified here only through the formula, over characteristic zero.

Scope changes

  • Proved over characteristic zero only; no statement for r>1 or other powers of T.
  • Journal pages and prior reports cited by the candidate were not available to me; my confirmation rests on the arXiv TeX excerpt and the candidate's mathematics alone.

Located concern

candidate-rank REX3 (i), 'this test works in every choice of eigenbasis'

Conjugation-invariance of ker(mu) is used but not stated.

It holds: R(gAg^{-1})=G R(A) G^{-1}, so tr(L(gAg^{-1})R(gAg^{-1}))=tr(G^{-1}L(gAg^{-1})G R(A)).

State L^g(A)=G^{-1}L(gAg^{-1})G as a lemma before (i).

None; expository only.

Located concern

candidate-rank REX3 (iii), 'the trace of L(A)D(A)^p is zero' on a block

L(A) is zero-diagonal only in the eigen-wedge basis, so the block trace needs the remark that the block projection is the same in both bases.

The complementary eigen-wedges and complementary standard wedges span the same coordinate subspace, and D(A) is diagonal on the block in the eigenbasis. So the compressed trace is sum mu_v^p (v^* L(A) v)=0 and can be evaluated in the standard basis. The three identities were recomputed and agree with the candidate.

Insert the projection remark; no change to the identities.

None.

Located concern

candidate-rank REX3 (ii), 'Since G=1+tD(E_ab)'

The first-order expansion needs G^{-1}=1-tD(E_ab), that is D(E_ab)^2=0.

D(N)^2(v^w)=N^2v^w+2Nv^Nw+v^N^2w=2wedge^2(N)=0 for rank-one nilpotent N. Then diag(G^{-1}BG)=diag(B)+tdiag([B,D])-t^2*diag(DBD) vanishes identically in t.

Add this line.

None.

Located concern

candidate-rank REX3 premises and scope_changes

The proof does not transfer to positive characteristic.

The swap is not diagonalizable in characteristic 2, nor the 3-cycle in characteristic 3, and small fields break coefficient extraction. The inequality g(n)<=(n^2-1)m^2+1 does hold over every field, because D(A) commutes with R(A) integrally and the centralizer is at least the scalars; equality was not shown there.

Keep the claim restricted to characteristic zero, as the candidate does. One field suffices for the counterexample.

None for the refutation; a modular Hilbert-function claim would be unproved.

V2

Checked statement: Candidate REX2: over C with p=T^2, r=1, the degree-2 part of sqrt(J_n) is Ce_2 for n>=2, so the reduced degree-2 dimension is binom(n^2+1,2)-1 for n>=2 and 1 at n=1. Candidate REX3 repairs: (a) for p_l=c_lT, dimension and character polynomiality for all n>=0, for both J_n and its radical; (b) for p_l=c_l*T^{s_l} with s_l>=1, eventual dimension polynomiality of degree <=2d over any field, with eventual characters only conditionally.

Returned assessment: no_issue_found

I found no error in REX2 or in either repair. The original Remark 5.3.3 and Theorem 5.3.2 are false as stated; the repairs are new arguments with narrower hypotheses or weaker conclusions.

REX2. (1) Rank-one matrices lie in X_n. The kernel of x_ij -> u_i v_j in degree 2 is spanned by 2-minors, since each image monomial has at most two preimages and they differ by a minor. So a vanishing quadric is q=sum c_{I,J} Delta_{I,J}, and the minors are independent. (2) For disjoint I,J the test matrix has A^2=0 and unique nonzero minor Delta_{I,J}. For overlap, A=E_cb+E_ac has A^2=E_ab of rank 1 and unique nonzero minor on rows {a,c}, columns {b,c}. So off-principal coefficients vanish. (3) I recomputed (1+tE_31)(E_12+E_23)(1-tE_31)=E_12+E_23+tE_32-tE_21. Its principal minors are t on {1,2}, -t on {2,3}, 0 otherwise, so c_12=c_23. Relabeling and connectivity give q in C*e_2 (n>=3; n=2 is immediate). (4) rank(A^2)<=1 allows at most one nonzero eigenvalue counted with multiplicity, so e_2(A)=0. The generic-point argument modulo primes containing J_n, or simply the Nullstellensatz over C, puts e_2 in sqrt(J_n). e_2 is nonzero for n>=2. Hence the reduced degree-2 dimension is binom(n^2+1,2)-1 for n>=2 and 1 at n=1, where J_1=0 and the polynomial gives 0. No single polynomial fits, also if n=0 is dropped. Remark 5.3.3 (source line 1829) is a false statement for this datum. The side remark is also right: (J_n)_2=0, so this degree does not test the scheme statement; V1 supplies that in degree 5.

Repair (a), determinantal, all n.

  • Directions. A partial injection (A,B,phi): S->T acts on matrices by principal A-block, relabel, zero-pad (source line 1820). On coordinates, x_{phi(s),phi(s')} -> x_{s,s'} and other variables -> 0.
  • Ideal compatibility. A target (r+1)-minor of cA pulls back to c^{r+1} times a signed source minor if its rows and columns lie in B, and to 0 otherwise. This also covers |S|<r+1.
  • So J_T -> J_S for every FI#-morphism. Radicals follow, since ring maps send sqrt into sqrt, and both ideals are homogeneous.
  • Via FI# = FI#^op (source lines 1269, 1823) the degree-d quotients are FI#-modules whose FI-action relabels monomials. A degree-d monomial uses at most 2d labels, so the module is generated in degree <=2d with finite-dimensional values.
  • By Theorem 4.1.5 (source lines 1288-1360, proof supplied and spot-checked) and Remark 2.3.8, V=M(W) with W_i=0 for i>2d. So dim V_n=sum_{i<=2d} dim W_i binom(n,i) for all n>=0. In characteristic 0 the fixed-subset computation at lines 1392-1397 gives a character polynomial of weighted degree <=2d for all n.
  • This is unconditional on the supplied text and does not need Theorems thm:fisubalgebra or thm:persinomial, which are not in the excerpt.
  • Consistency check: for r=1, d=2 the count binom(n^2+1,2)-binom(n,2)^2 is right at n=0,1.

Repair (b), positive monomials, eventual.

  • Only zero-extension is compatible: p(A+0)=p(A)+0, so target minors pull back to source minors or 0. Principal restriction is not; the candidate's example [[I,I],[-I,-I]] has square zero and principal block I_2.
  • So O(X)^d is a co-FI quotient of the ambient. Its dual U is an FI-submodule of the ambient dual, where an injection f sends the dual monomial of m to the dual monomial of f(m). I checked this pairing.
  • Leading-index argument. An increasing injection only inserts zero exponents at non-image pairs and keeps the relative lex order of image pairs, so it preserves the monomial order and lead(phi_* u)=phi(lead u). The lead set is closed under increasing injections, and dim U_n equals the number of leads at size n.
  • Monomials correspond to a pattern (finitely many, q<=2d used labels) plus a gap tuple in N^{q+1}. Increasing injections realize exactly the componentwise enlargements of gaps. Dickson's lemma gives finitely many minimal tuples.
  • An orthant above b contributes binom(n-sum b,q) at size n. This is in fact valid for all n>=sum b, slightly better than the candidate states. Inclusion-exclusion with coordinatewise maxima gives an eventual polynomial of degree <=q<=2d.
  • The argument is field-independent and applies equally to the radical quotient.
  • Finite generation versus dimension. The count gives eventual polynomiality directly, without any Noetherian or Hilbert-function theorem. Finite generation of U follows separately by leading-term reduction at fixed size and is needed only for the character statement.

Characters. In case (a) they are unconditional for all n. In case (b), the eventual character polynomial and its degree bound <=2d depend on the general characteristic-zero theorem for finitely generated FI-modules and its weight bound, whose proofs are outside the supplied ranges. The dual has the same character because sigma and sigma^{-1} are conjugate. No all-n character or dimension statement survives in case (b): V1 and REX2 give counterexamples at n=1.

Premises used

  • FI#-classification Theorem 4.1.5 and the computation in the proof of Theorem 4.1.7 as supplied (lines 1288-1397); Remark 2.3.8 (lines 793-797).
  • k a field; characteristic 0 for character statements; C for the REX2 computation.
  • Standard grading; homogeneity of minors of c*T^s.

Missing inputs

  • Proofs of the general eventual character-polynomial theorem and weight bound for finitely generated FI-modules in characteristic 0 (referenced as thm:persinomial; not in the excerpt). Needed only for eventual characters in repair (b).
  • Theorem thm:fisubalgebra, cited by the original proof at source line 1825, is not supplied. The repair (a) argument avoids it.

Scope changes

  • Repair (a) covers tuples in which every p_l is a scalar multiple of T; mixed tuples fall under (b) or fail.
  • Repair (b) concerns standard-graded quotients compatible with zero extension; inhomogeneous data or data with p_l(0) nonzero are excluded.
  • REX2 is verified over C; one field suffices against the universal remark.

Located concern

CEF-rank-source lines 1820-1829 versus candidate-rank REX2 and REX3

The status of each statement needs to be kept separate.

False as stated: Theorem 5.3.2 (p=T^2, r=1, d=5 from V1; also p=1) and Remark 5.3.3 (p=T^2, r=1, d=2). Invalid proof step: 'the rank varieties are preserved by these maps' (line 1821) fails for principal restriction once some p_l is not a scalar multiple of T. True, with the original preservation step valid: all p_l=c_l*T. True only eventually and by a different proof: positive monomials.

Restate Theorem 5.3.2 and Remark 5.3.3 for determinantal data (all n), or for homogeneous zero-extension-compatible data with 'for n sufficiently large'.

The candidate's classification stands; no change.

Located concern

candidate-rank REX3 third issue, final sentences on characters

The eventual character statement for positive monomials is conditional on an unsupplied theorem.

Finite generation of the dual FI-module is proved by the leading-term argument, but passing from finite generation to an eventual character polynomial of degree <=2d uses results outside the excerpt.

Either cite that theorem as an external input, or state only the eventual dimension result, which is fully proved here over any field.

Dimension repair unconditional; character repair conditional. The candidate already flags this correctly.

Located concern

candidate-rank REX3 third issue, orthant count

Minor sharpening and no explicit threshold.

binom(n-sum b,q) is the exact slice count for every n>=sum b, since it is 0 when 0<=n-sum b<q. The eventual range is governed by the largest coordinatewise maximum among minimal gap tuples, which the argument does not bound.

State the result as 'for all n>=N(P,r,d)' with N not effective.

None; the examples at n=1 show the threshold cannot be 0 or 1 in general.

Outside the audit

  • The journal pages and prior reports cited by the candidate (journal-rank-text, journal-rank-p55, journal-rank-p56, prior-report-A-rank-, prior-report-B-rank-) were not supplied; statements about their content, including the reported rank table, are unverified.
  • The candidate's REX1 items outside the two targets, apart from the block-matrix preservation example used for map directions in V2.
  • Unsupplied results of the source (thm:fisubalgebra, thm:persinomial, weight and stability-degree propositions) and all external literature.
  • Positive-characteristic analogues of the degree-five relation count, bounds r>1, and other powers of T.
  • Any computer verification.
Source record

Additional supporting archive items are identified in the original report. View the manuscript hashes and repository snapshot.

Cite this review

If this review or its supporting reports inform your work, please cite it.

Igor Rivin. Review of FI-modules and stability for representations of symmetric groups. Math Referee, .

Download .bib
View BibTeX
@misc{rivin2026review-human-church-ellenberg-farb-fi-modules,
  author = {Rivin, Igor},
  title = {{Review of FI-modules and stability for representations of symmetric groups}},
  year = {2026},
  month = {November},
  date = {2026-11-20},
  howpublished = {Math Referee},
  url = {https://mathreferee.dimensionreducers.ai/papers/human-church-ellenberg-farb-fi-modules/panel/},
  note = {Review report}
}