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

gpt-6.2-sol 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.

Both proposed mathematical arguments pass their stated checks. The degree-five commutator calculation exhausts the kernel over C, including at n=2 and n=3. The quadratic radical computation also proves the surviving equation's radical membership, not just vanishing on test points. These give distinct numerical counterexamples to the original all-n claim. The narrower all-n determinantal repair and the eventual dimension repair for positive monomials are proved from the supplied material and the checked arguments. The general eventual character extension for positive monomials still uses a separate FI theorem.

V1

Checked statement: For p=T² and r=1 over C, the proposed exhaustion of the linear relations among variables times the quartic minors is valid, and g(n)=dim(J_n)_5=(n²−1)binom(n,2)²+1 for n≥2.

Returned assessment: no_issue_found

The kernel exhaustion works. In particular, the shear at a repeated eigenvalue gives the required divisibility on the diagonal algebra, and the internal three-cycle supplies a consistency relation that the spectator cycles alone do not give.

  1. Set H=End(Cⁿ), W=∧²Cⁿ, m=dim W, R(A)=∧²(A²), and let D(A) act by D(A)(v∧w)=Av∧w+v∧Aw. For L:H→End(W) linear, put μ(L)(A)=tr(L(A)R(A)). The entries of R(A) are precisely the indexed minors. Taking L to be a coordinate linear form times a matrix unit produces each variable times each minor, so im μ=(J_n)_5 without any prior assumption of independence. Since D(A) commutes with R(A), every map A↦[D(A),C] is in ker μ by cyclicity of trace. Also, if G=∧²g, then L_g(A)=G⁻¹L(gAg⁻¹)G is in the kernel whenever L is: μ(L_g)(A)=μ(L)(gAg⁻¹).

  2. For h diagonal, write e_I for the exterior basis vector on a pair I and w_I(h)=Σ_{i∈I}h_i. The kernel identity is Σ_{I={i,j}} L(h)_{I,I} h_i²h_j²=0. The monomials h_k h_i²h_j² determine both I and k, including when k∈I. Thus every diagonal entry of L(h) is zero. Applying the same computation to L_g makes this a statement in every eigenbasis, even when eigenvalues repeat.

For I≠J, match I\J to J\I by i↦j(i) and impose h_i=h_{j(i)}. The shear g=1+tΣ_i E_{j(i),i} commutes with h. Its inverse changes no row indexed by I, so e_IG⁻¹=e_I. The coefficient of the highest power t^{|I\J|} in G e_I is ±e_J. Hence the zero diagonal identity e_I*G⁻¹L(h)G e_I=0 forces L(h){I,J}=0 on the imposed equalities. For one replacement the equalities define the hyperplane w_I=w_J. For two replacements, the two matchings give codimension-two subspaces whose intersection, where all four coordinates agree, has codimension three. Their sum is therefore the hyperplane w_I=w_J. Linearity gives vanishing on that sum. Consequently L(h){I,J} is a scalar multiple of w_I−w_J, and these scalars are the off-diagonal entries of one constant matrix C₀ with L(h)=[D(h),C₀] on the fixed diagonal algebra.

  1. Subtract that commutator so that L(h)=0 there. For a≠b put B=L(E_ab). With g=1+tE_ab one has ghg⁻¹=h+t(h_b−h_a)E_ab and G=1+tD(E_ab); there is no quadratic term in G because E_ab has rank one. On diagonals, L_g(h)=t(h_b−h_a)G⁻¹BG. For t≠0, the preceding divisibility forces any nonzero off-diagonal entry of G⁻¹BG to have weight difference proportional to h_b−h_a. The only such entries join {b,i} and {a,i} for i distinct from a,b, in either direction. Its diagonal is zero as well. These zero-entry conditions also hold at t=0 by polynomiality. The first-order term BD(E_ab)−D(E_ab)B has zero diagonal. On the oriented two-vector block u_b=e_b∧e_i, u_a=e_a∧e_i, the operator D(E_ab) sends u_b to u_a and kills u_a. Thus the u_b diagonal entry of the first-order term is exactly the coefficient of u_b in B u_a. This eliminates the reverse entry separately for each i. The only possibly nonzero action is B(e_b∧e_i)=l_{ab,i} e_a∧e_i.

  2. Use A=E_ab+E_ba or A=E_ab+E_bc+E_ca, acting as a cycle on its supporting coordinates and zero elsewhere. These matrices are diagonalizable over C. For a fixed outside spectator i, the wedge of the cycle's coordinate span with e_i is preserved by L(A) and D(A), and the internal exterior square of the supporting span is preserved as well. Each such block has a basis from an eigenbasis of A, in which D(A) is diagonal and L(A) has zero diagonal by step 2. Thus the trace of L(A)D(A)^p on each block vanishes. With p=1 on a two-cycle spectator block this gives l_{ab,i}+l_{ba,i}=0. With p=2 on a three-cycle spectator block it gives l_{ab,i}+l_{bc,i}+l_{ca,i}=0. On the internal three-cycle block, D(A) sends e_a∧e_b to −e_b∧e_c, then to −e_c∧e_a with the same transition sign, and cycles back with sign −1. The corresponding three coefficients of L(A) are −l_{ca,b}, −l_{ab,c}, −l_{bc,a}. The trace with p=2 therefore gives l_{ca,b}+l_{ab,c}+l_{bc,a}=0.

For each fixed i, the spectator equations yield numbers b_{ai} with l_{ab,i}=b_{bi}−b_{ai}: choose a base vertex among labels other than i and use the triangle equation, or just antisymmetry when there are only two vertices. The internal equation then says d_{ab}+d_{bc}+d_{ca}=0 for d_{ab}=b_{ab}−b_{ba}. Antisymmetry and this equation give d_{ab}=t_a−t_b by choosing a base vertex. Replacing b_{ai} by b_{ai}+t_i makes them symmetric and leaves the differences unchanged. The diagonal C₁ with entry b_{ai}+t_i at {a,i} now has D(E_ab),C₁=l_{ab,i}e_a∧e_i. It also commutes with D(h). Thus the normalized L equals [D(−),C₁] on all matrix units, and the original L equals [D(−),C₀+C₁].

  1. At n=3 each spectator graph has two vertices, so it needs only the two-cycle equation; the internal three-cycle still gives the required symmetry of the potentials. At n=2 the exterior basis has one vector, the off-diagonal support from step 3 is empty, and no potential equations are needed. For every n≥2, commuting with all D(h) makes a centralizing C diagonal because the pair weights are distinct. Commuting with D(E_ab) then equates its entries on pairs differing by one label. This graph is connected, including the singleton at n=2, so the centralizer consists exactly of scalars. The kernel dimension is m²−1, and rank-nullity gives g(n)=n²m²−(m²−1). In particular, g(2)=4 and g(3)=73. At n=0,1 there are no indexed minors and g(n)=0; the formula for n≥2 must not be used there.

The degree-five quotient dimension is binom(n²+4,5)−g(n). The polynomial obtained from the formula for n≥2 is forced by its infinite tail but gives 0 at n=1, whereas the quotient at n=1 is a polynomial ring in one variable, with degree-five dimension 1. This refutes the original all-n theorem for the minor scheme itself, not just its proof. Evaluating a putative all-n character polynomial at the identity gives the same contradiction.

Premises used

  • The base field for the exact kernel calculation is C, so coefficient comparisons are polynomial identities over an infinite field, and the short cycle matrices are diagonalizable.
  • J_n is the ideal generated by the 2-minors of A², not its radical. Since the generators are homogeneous quartics, multiplying them by linear forms spans its degree-five piece.

Missing inputs

Scope changes

  • The all-n formula in the original source is false; the candidate's rank formula for n≥2 is proved.
  • The argument independently proves the rank formula over C. It does not verify any reported computation or assert the same relation count in positive characteristic.

V2

Checked statement: The reduced quadratic counterexample is correct, as are the all-n determinantal repair and the eventual dimension repair for positive monomials. The general eventual character statement for the latter repair requires a separate input.

Returned assessment: no_issue_found

Reduced quadratic calculation. Work over C with p=T², r=1 and J_n the minor ideal. A quadratic in √J_n vanishes on every rank-one matrix. Under x_ij=u_i v_j, monomials in degree two are grouped by the row and column multisets of their two factors. A group has one monomial unless both multisets have distinct entries; in that case it has two, and their difference is a 2-minor. Since the substitution vanishes identically, the quadratic is a linear combination of the minors Δ_{I,J}. For disjoint I,J, map the basis vectors in J bijectively to those in I and kill all other vectors. Then A²=0 and the only nonzero 2-minor is Δ_{I,J}=±1. For I={c,a}, J={b,c}, use the chain b→c→a→0 and kill the remaining vectors. Its square has rank one, and again only Δ_{I,J} is nonzero. These examples eliminate every off-principal coefficient.

On a coordinate triple let N=E₁₂+E₂₃ and conjugate by 1+tE₃₁. The result is N+tE₃₂−tE₂₁. Its square still has rank one; its principal 2-minors on 12 and 23 are t and −t, and all others are zero. Relabeling and taking t=1 equates the coefficients of any two pairs sharing a label. This connects all pairs for n≥3; for n=2 there is just one pair. The only possible quadratic is therefore a multiple of e₂=Σ_I Δ_{I,I}.

Membership in the radical is a separate step. For each prime containing J_n, pass from the quotient domain to the algebraic closure of its fraction field. There A² has rank at most one. In a triangular form for A, two nonzero diagonal entries would give a nonzero principal 2-minor of the triangular form for A². So at most one eigenvalue, counted with multiplicity, is nonzero. This kills the coefficient of t² in det(1+tA), which is e₂. Thus e₂ is zero modulo every prime containing J_n and belongs to √J_n. For n≥2 it is nonzero as a polynomial, so the reduced quadratic dimension is binom(n²+1,2)−1. At n=1 it is 1, not the 0 predicted by that polynomial. The scheme quotient does not have this quadratic relation: (J_n)_2=0. Its separate all-n failure is the degree-five result in V1.

Directions of maps and the determinantal repair. A partial injection A⊆S→T acts on matrices by restriction to the principal A-block, then relabeling and zero padding. The pullback goes from the target polynomial ring to the source polynomial ring. For p_l=c_l T, each target generator pulls back to a signed source minor times the scalar power, or to zero. This gives pullback compatibility of the homogeneous ideals for every partial injection, including when some size has no such minors. For any compatible ideals, radical compatibility follows by pulling back a power. Moreover, the radical of a homogeneous ideal is homogeneous: in the graded quotient the highest homogeneous component of a nilpotent is nilpotent, and subtracting it and descending proves the assertion. These are statements about the scheme ideals and their radicals, not about functions on just the k-rational points of a finite field.

Use partial inverse to turn contravariant pullbacks into a covariant FI#-algebra. On a total injection, the resulting covariant map on ambient polynomials relabels the variables. Each degree-d monomial uses at most 2d labels, so these monomials generate the degree-d ambient FI-module in sizes at most 2d, and its quotient has the same generation bound. Each value is finite dimensional. The supplied FI# decomposition then gives H₀ supported in sizes at most 2d and dim V_n=Σ_{a≤2d}dim H₀(V)_a·binom(n,a) for every n, over any field. In characteristic zero the supplied fixed-subset trace calculation gives an all-n character polynomial of weighted degree at most 2d, with weight i assigned to X_i. No general FI character theorem or characteristic-zero weight argument is needed for the dimension statement.

Positive monomials and leading indices. For p_l=c_l T^{s_l} with s_l≥1, total zero extensions on matrices preserve the minor equations because p_l(M⊕0)=p_l(M)⊕0. Pullbacks of target minors are source minors or zero. The minors are homogeneous, and the radical argument still applies. But for this repair only the zero-extension pullbacks are available in general, so the coordinate quotients are co-FI. Their degree-d duals U embed covariantly in the ambient degree-d dual. On dual monomials an injection relabels the exponent array and puts zeros in the new positions, because pullback on polynomials kills any monomial using a new label and otherwise returns its unique relabeling. This is not an FI# action on the quotient or on U.

Restrict to increasing injections. At each size order the dual monomials lexicographically by exponent arrays in lexicographic pair order. An increasing injection preserves comparison because relative pair order is preserved and every new position is zero in both arrays. Accordingly it takes the leading index of a nonzero vector of U to the leading index of its image. Encode an index by its exponent pattern on the q used labels, in increasing order, and its q+1 gaps of unused labels. There are finitely many patterns and q≤2d. Within a pattern, componentwise gap enlargement is exactly what can be realized by an increasing injection: send the used labels to the corresponding used labels and embed each intervening gap separately. Hence the leading indices in that pattern form an upward set of gap tuples.

Each such upward set has finitely many minimal tuples. Indeed, any infinite sequence in a finite product of copies of the nonnegative integers has an infinite componentwise nondecreasing subsequence, by extracting a constant or increasing subsequence successively in each coordinate. This precludes infinitely many mutually incomparable minimal tuples; descent shows that every tuple of the set dominates one. At a fixed size, elimination gives one independent leading index per basis vector of U_n. For a minimum b, the number of tuples above it with total unused labels n−q is binom(n−Σ_i b_i,q) once n≥q+Σ_i b_i, by counting the residual gaps. Intersections of these sets use coordinatewise maxima of the minima. Finite inclusion-exclusion and the finite number of patterns give an eventual dimension polynomial of degree at most 2d, over any field. The counting threshold matters: below it, an orthant can be empty even when the polynomial binomial expression is nonzero.

Finite generation is a separate consequence of the same leading-index construction: choose one vector for each minimal tuple of each pattern. For any leading index, an increasing injection carries a chosen vector's leading index to it. Scale and subtract to cancel that term; the process terminates in the finite basis at the given size. These vectors generate under increasing injections and hence under FI. The eventual dimension result above comes from counting, not merely from finite generation or a Noetherianity assertion. For the general eventual character conclusion in characteristic zero, one can apply the general finitely generated FI-module character-polynomial theorem to U, if taken as an additional input. The original quotient has the same character as its dual because permutations are conjugate to their inverses. Neither the gaps nor the supplied FI# character formula establishes this general FI character conclusion on its own.

Premises used

  • Reduced means the quotient by the radical of the supplied minor ideal. The quadratic counterexample is over C.
  • The determinantal repair uses the supplied FI# classification and its fixed-subset trace formula only after proving actual pullback compatibility.
  • Degree d is fixed; the ambient spaces have finite monomial bases, and ordinary finite-dimensional linear duality applies over the base field.

Missing inputs

  • For the general eventual character-polynomial assertion in the positive-monomial repair, the general characteristic-zero FI character theorem, in a version available by 2013-07-14, must be supplied or cited as an external input. Its proof is not in the supplied ranges. The eventual dimension assertion needs no such input.

Scope changes

  • The verdict accepts the candidate's counterexample and qualified repairs, not the unrestricted theorem or remark in the original source.
  • The all-n determinantal character conclusion is unconditional from the supplied FI# material. The general eventual character inference in the positive-monomial repair is conditional on a separate general FI theorem.
  • The eventual dimension proof applies more generally to homogeneous quotients of this matrix polynomial ring compatible with zero-extension pullbacks. It does not give an all-n bound or a threshold for arbitrary inhomogeneous polynomial data.

Located concern

CEF-rank-source lines 1794–1806 and 1820–1823

The original proof assumes both standard homogeneity and preservation by partial injections for arbitrary polynomial data.

For p=T−1, r=1, n=2, the proper ideal generated by det(A−I) is not homogeneous: its generator has constant component 1. Even with p=T² and r=1, the block matrix [[I₂,I₂],[−I₂,−I₂]] has square zero, but its first principal block is I₂ and violates the condition. Also p(M⊕0_s)=p(M)⊕p(0)I_s, so arbitrary zero padding alone need not preserve the bound when p(0)≠0.

Verify homogeneity and the relevant pullback compatibilities separately. Use FI# only when all partial pullbacks work; use the dual FI argument when only the zero extensions work.

Some original degree pieces are undefined, and the general FI# construction fails even for homogeneous data. Failure of that construction alone does not refute a numerical polynomiality statement.

Located concern

CEF-rank-source lines 1811–1815 and 1829; candidate-rank, REX2 and REX3

The original all-n conclusions are false even where the standard graded pieces are defined.

For the same homogeneous data p=T², r=1 over C, the radical quotient in degree two and the minor quotient in degree five each disagree at n=1 with their uniquely determined tail polynomial.

Retain the all-n statement under homogeneous FI# compatibility, in particular for ordinary determinantal conditions. For homogeneous positive monomials, replace all-n dimension polynomiality by eventual dimension polynomiality.

The counterexamples refute the original dimension statements and, by evaluating at the identity, the corresponding all-n characteristic-zero character statements. They do not refute the abstract FI# polynomiality result.

Outside the audit

  • Journal pages and prior-report pages named inside candidate-rank were not supplied as source texts; no comparison with them was made.
  • The omitted general FI character theorem and omitted source passages outside the supplied line ranges were not audited.
  • No reported code execution, positive-characteristic version of the exact kernel count, higher-rank analogue, or historical priority claim was checked.
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}
}