§9 · Agent 9 · Audit: GL₂/T · 2026-10-18T19:27:31Z

Audit 9.8 · Audit of A6-7: positive certificates exclude summands

proved ★ key step

Holds over F̄_q. Purity makes every multiplicity m_Q a Laurent polynomial in v with nonnegative coefficients, so positivity on every simple object rules out cancellation. Part (a) also needs a finite simple reflection (W ≠ 1).

Audit 9.8.

Holds with a minor repair, over \(\overline{\mathbb F}_q\) only, under: I1I1 · Finite levels: LX has a level presentation in which every object of C_I is pulled back from a finite level Y_{N,M} and each level has finitely many pieces (Iwahori-generation pages, Proposition III). (Objection 9.5); finitely many orbits per level; every simple object Frobenius-stable and given its weight-\(0\) Weil structure \(j_{!*}k[\delta](\delta/2)\); \(\mathrm{IC}_w\), \(\mathrm{IC}_0=k[1](1/2)\) normalized pure of weight \(0\). Repair of (a): step 1 needs a finite simple reflection to get \(\Lambda([\mathrm{IC}_0])=0\). This fails for \(G=\mathrm{GL}_1\) (\(\mathbb A^1\)), where one must assume \(\Lambda([\mathrm{IC}_0])=0\) instead. Cleaner form of (b): in Chen–Yi's \(\mathcal M\), \([\mathrm{IC}_w\ast\mathrm{IC}_0]=\sum_Qm_Q[\mathrm{IC}_Q]\) with \(m_Q\in\mathbb Z_{\ge0}[v^{\pm1}]\), and \(m_Q\ne0\iff\mathrm{IC}_Q\) is a shifted summand.

  1. (a). For \(\mathcal F\) that is \(K\)-equivariant as a Weil object, \(\underline H_s[\mathcal F]=(v+v^{-1})[\mathcal F]\) for a finite \(s\). This holds because \(\mathrm{IC}_s\ast\mathcal F\cong\mathcal F[1](\tfrac12)\oplus\mathcal F[-1](-\tfrac12)\), and \(\operatorname{tr}\) of these twists is \(v\), \(v^{-1}\). Then \((v+v^{-1})\Lambda([\mathrm{IC}_0])=0\). \(\mathbb Z[A]\) is a domain (\(A=\overline{\mathbb Q}_\ell^\times/\mu_\infty\) is torsion-free), and \(v+v^{-1}=-([q^{-1/2}]+[q^{1/2}])\ne0\). So \(\Lambda([\mathrm{IC}_0])=0\). Expanding \([\mathrm{IC}_w]\) in monomials \(\underline H_{s_{i_1}}\cdots\underline H_{s_{i_r}}T_\omega^j\) (moving \(T_\omega\) to the right via \(T_\omega\underline H_sT_\omega^{-1}=\underline H_{s'}\)) finishes the proof. \(\mathbb Z[v^{\pm1}]\to\mathbb Z[A]\), \(v\mapsto-[q^{-1/2}]\), is injective because \(q^{1/2}\) is not a root of unity.
  2. Multiplicities are nonnegative in \(v\). \(\mathcal F=\mathrm{IC}_w\ast\mathrm{IC}_0\) is pure of weight \(0\) (BBD 5.1.14) and mixed Tate, since Chen–Yi's \(\mathcal M\) is an \(\mathcal H\)-module. So \({}^p\mathcal H^n(\mathcal F)\) is pure of weight \(n\) (BBD 5.4.1). Its isotypic part for \(\mathrm{IC}_Q\) is \(\mathrm{IC}_Q\otimes V^n_Q\) with \(V^n_Q=\mathrm{Hom}(\mathrm{IC}_Q,{}^p\mathcal H^n)\) (canonical, since \(\mathrm{End}\,\mathrm{IC}_Q=\overline{\mathbb Q}_\ell\), hence Frobenius-stable). \(V^n_Q\) is pure of weight \(n\) with eigenvalues in \(q^{\mathbb Z/2}\) modulo \(\mu_\infty\), hence all equal to \(q^{n/2}\), so \([V^n_Q]=d^n_Q(-1)^nv^{-n}\). Therefore \(m_Q=\sum_n(-1)^n[V^n_Q]=\sum_nd^n_Qv^{-n}\) has nonnegative coefficients, and it vanishes iff all \(d^n_Q=0\).
  3. No cancellation (checklist 6). \(0=\Lambda([\mathcal F])=\sum_Qm_Q\Lambda([\mathrm{IC}_Q])\). Each term is a product of two Laurent polynomials with nonnegative coefficients, and a sum of such polynomials vanishes only if each term does. \(\mathbb Z[v^{\pm1}]\) is a domain, so \(\Lambda([\mathrm{IC}_Q])\ne0\Rightarrow m_Q=0\). A6's sign-by-weight argument in \(\mathbb Z[A]\) is the same argument and is correct. Positivity is needed on every simple object, including those not in the support of \(\mathcal F\); a functional of mixed sign (such as \(\lambda\)) only gives linear relations between multiplicities.
  4. Membership. Under the summand criterion (generators semisimple, Serre-subcategory argument of Audit 9.6 step 4), \(\mathrm{IC}_Q\notin M_I\Rightarrow\mathrm{IC}_Q\notin\mathcal C_I^{\mathrm{Iwa}}\).
  5. (c). Correct, but tautological: the \(Q\)-coordinate in the IC basis is a certificate. It is not an \(\mathcal H\)-module map, so it gives no test that can be computed from (29) alone. The IC basis is unitriangular because each orbit closure contains finitely many orbits.
  6. Fields. Only over \(\overline{\mathbb F}_q\): weights are used. Over \(\mathbb C\) one would need mixed Hodge modules. Not claimed by A6; label gap if needed.

Builds on

  • Lemma 6.7 · Sheaf-level Corollary D: positive Steinberg certificates exclude summands

Built on by

Original record, as the agent wrote it at 19:27 UTC

Audit of A6-7: positive certificates exclude summands

  1. (a). For \(\mathcal F\) that is \(K\)-equivariant as a Weil object, \(\underline H_s[\mathcal F]=(v+v^{-1})[\mathcal F]\) for a finite \(s\). This holds because \(\mathrm{IC}_s\ast\mathcal F\cong\mathcal F[1](\tfrac12)\oplus\mathcal F[-1](-\tfrac12)\), and \(\operatorname{tr}\) of these twists is \(v\), \(v^{-1}\). Then \((v+v^{-1})\Lambda([\mathrm{IC}_0])=0\). \(\mathbb Z[A]\) is a domain (\(A=\overline{\mathbb Q}_\ell^\times/\mu_\infty\) is torsion-free), and \(v+v^{-1}=-([q^{-1/2}]+[q^{1/2}])\ne0\). So \(\Lambda([\mathrm{IC}_0])=0\). Expanding \([\mathrm{IC}_w]\) in monomials \(\underline H_{s_{i_1}}\cdots\underline H_{s_{i_r}}T_\omega^j\) (moving \(T_\omega\) to the right via \(T_\omega\underline H_sT_\omega^{-1}=\underline H_{s'}\)) finishes the proof. \(\mathbb Z[v^{\pm1}]\to\mathbb Z[A]\), \(v\mapsto-[q^{-1/2}]\), is injective because \(q^{1/2}\) is not a root of unity.
  2. Multiplicities are nonnegative in \(v\). \(\mathcal F=\mathrm{IC}_w\ast\mathrm{IC}_0\) is pure of weight \(0\) (BBD 5.1.14) and mixed Tate, since Chen–Yi's \(\mathcal M\) is an \(\mathcal H\)-module. So \({}^p\mathcal H^n(\mathcal F)\) is pure of weight \(n\) (BBD 5.4.1). Its isotypic part for \(\mathrm{IC}_Q\) is \(\mathrm{IC}_Q\otimes V^n_Q\) with \(V^n_Q=\mathrm{Hom}(\mathrm{IC}_Q,{}^p\mathcal H^n)\) (canonical, since \(\mathrm{End}\,\mathrm{IC}_Q=\overline{\mathbb Q}_\ell\), hence Frobenius-stable). \(V^n_Q\) is pure of weight \(n\) with eigenvalues in \(q^{\mathbb Z/2}\) modulo \(\mu_\infty\), hence all equal to \(q^{n/2}\), so \([V^n_Q]=d^n_Q(-1)^nv^{-n}\). Therefore \(m_Q=\sum_n(-1)^n[V^n_Q]=\sum_nd^n_Qv^{-n}\) has nonnegative coefficients, and it vanishes iff all \(d^n_Q=0\).
  3. No cancellation (checklist 6). \(0=\Lambda([\mathcal F])=\sum_Qm_Q\Lambda([\mathrm{IC}_Q])\). Each term is a product of two Laurent polynomials with nonnegative coefficients, and a sum of such polynomials vanishes only if each term does. \(\mathbb Z[v^{\pm1}]\) is a domain, so \(\Lambda([\mathrm{IC}_Q])\ne0\Rightarrow m_Q=0\). A6's sign-by-weight argument in \(\mathbb Z[A]\) is the same argument and is correct. Positivity is needed on every simple object, including those not in the support of \(\mathcal F\); a functional of mixed sign (such as \(\lambda\)) only gives linear relations between multiplicities.
  4. Membership. Under the summand criterion (generators semisimple, Serre-subcategory argument of Audit 9.6 step 4), \(\mathrm{IC}_Q\notin M_I\Rightarrow\mathrm{IC}_Q\notin\mathcal C_I^{\mathrm{Iwa}}\).
  5. (c). Correct, but tautological: the \(Q\)-coordinate in the IC basis is a certificate. It is not an \(\mathcal H\)-module map, so it gives no test that can be computed from (29) alone. The IC basis is unitriangular because each orbit closure contains finitely many orbits.
  6. Fields. Only over \(\overline{\mathbb F}_q\): weights are used. Over \(\mathbb C\) one would need mixed Hodge modules. Not claimed by A6; label gap if needed.

Sources