Why the trace of Frobenius does not turn sheaf generation into function generation
A self-contained account of the categorical trace of Gaitsgory–Kazhdan–Rozenblyum–Varshavsky [GKRV] and Arinkin–Gaitsgory–Kazhdan–Raskin–Rozenblyum–Varshavsky [AGKRRV3], applied to the action map \(F\mapsto F\ast\mathrm{IC}_0\). It locates exactly where "sheaves generate, take traces, so functions generate" breaks, and says when it does not. The counterexamples are those of Proposition IX.
Sources
- [GKRV] D. Gaitsgory, D. Kazhdan, N. Rozenblyum, Y. Varshavsky, A toy model for the Drinfeld–Lafforgue shtuka construction, Indag. Math. (N.S.) 33 (2022), 39–189; arXiv:1908.05420. Cited sections:
- §3.1: trace and cyclicity;
- §3.2.1: the trace of a 1-morphism with a right adjoint;
- §3.2.3–3.2.4: composition and cyclicity;
- §3.3.1: \(\operatorname{Tr}\) as a symmetric monoidal functor;
- §3.3.2–3.3.3 and §3.4.4–3.4.6: algebra and module structures, including Proposition 3.4.6;
- §3.4.1–3.4.3: DG categories and the classes \(\operatorname{cl}(c,\alpha)\);
- §3.5.1: the trace as Hochschild homology.
- [AGKRRV1] D. Arinkin, D. Gaitsgory, D. Kazhdan, S. Raskin, N. Rozenblyum, Y. Varshavsky, The stack of local systems with restricted variation and geometric Langlands theory with nilpotent singular support, arXiv:2010.01906. Cited: §22.2 (the local-term map), Conjecture 22.3.7 (the Trace Conjecture), Conjecture 14.1.8.
- [AGKRRV2] the same authors, Duality for automorphic sheaves with nilpotent singular support, arXiv:2012.07665. This gives the self-duality of \(\mathrm{Shv}_{\mathrm{Nilp}}(\mathrm{Bun}_G)\) used to compute the trace.
- [AGKRRV3] the same authors, Automorphic functions as the trace of Frobenius, arXiv:2102.07906. Cited:
- Main Theorem 0.2.6 (Corollary 4.1.4);
- §0.4.2: the failure for \(\mathrm{Shv}(Y)\);
- (0.9) and Remark 0.6.5: local terms on classes;
- Theorem 0.6.8 (Theorem 5.2.3), Corollary 0.6.9, Remarks 0.6.10–0.6.11.
- [GR1] D. Gaitsgory, N. Rozenblyum, A study in derived algebraic geometry, Vol. I, Math. Surveys Monogr. 221, AMS, 2017. Cited: Ch. 1 (\(\mathrm{DGCat}\), duality) and Ch. 10 (2-categories).
- The run: Theorems 1.2 and 1.3, Examples 1.4, Proposition 1.8, Lemma 2.8, Theorems 2.6 and 3.5, and Corollary 3.9.
Proposition IX as a paper · the verdict, the trace formalism and the proofs (PDF) →
Definitions
Group data over a finite field
\(p\) is an odd prime and \(q\) a power of \(p\). \(G\) is a connected reductive group over \(\mathbb F_q\) and \(\theta\) an involution of \(G\) over \(\mathbb F_q\):
$$X=G/G^\theta,\qquad x_0=eG^\theta .$$\(S\) is a maximal \(\theta\)-split torus and \(T\supseteq S\) a \(\theta\)-stable maximal torus. Labels and the leading-term map are those of the pages:
$$\Lambda_T^+\subseteq X_*(T),\qquad\Lambda_S^+=X_*(S)\cap\Lambda_T^+,\qquad\varphi(\lambda)=\lambda-\theta(\lambda),$$with the dominance order on \(\Lambda_S^+\), in which lower sets \(\{\nu\le\mu\}\) are finite.
Sheaves over \(k=\overline{\mathbb F}_q\)
\(LX\) and \(L^+G\) are the loop space and the arc group over \(k\); the orbits are \(\mathcal O_\mu=L^+G\cdot x_\mu\) with \(\tau(x_\mu)=t^\mu\). The category is Chen–Yi's \(\ell\)-adic one:
$$\mathcal C_{K,k}=\{\text{bounded }L^+G\text{-equivariant constructible }\overline{\mathbb Q}_\ell\text{-complexes on }LX\text{, supported on finitely many orbit closures}\}.$$\(\mathrm{IC}_\nu\) is the intersection complex of \(\overline{\mathcal O_\nu}\) with constant coefficients, \(V_\lambda\) the Satake sheaf of \(\operatorname{Gr}_{\le\lambda}\), and \(\mathrm{IC}_0\) the constant sheaf on \(L^+X\):
$$\mathrm{IC}_\nu|_{\mathcal O_\nu}\simeq\overline{\mathbb Q}_\ell[\delta(\nu)],\qquad\mathcal C^{\mathrm{Sat}}_{K,k}=\operatorname{Thick}_{\mathcal C_{K,k}}\{V_\lambda\ast\mathrm{IC}_0:\lambda\in\Lambda_T^+\}.$$The Betti category \(\mathcal C_K\) of the pages is the same construction over \(\mathbb C\) with complex coefficients.
Functions over \(F=\mathbb F_q(\!(t)\!)\)
\(F=\mathbb F_q(\!(t)\!)\) has ring of integers \(\mathcal O=\mathbb F_q[\![t]\!]\). Its points are the \(\mathbb F_q\)-points of the loop spaces:
$$LX(\mathbb F_q)=X(F),\qquad LG(\mathbb F_q)=G(F),\qquad L^+G(\mathbb F_q)=K=G(\mathcal O).$$The spherical Hecke algebra \(\mathcal H_K\) consists of the compactly supported \(K\)-bi-invariant functions on \(G(F)\); it has the basis \(T_\lambda\), \(\lambda\in\Lambda_T^+\) (Cartan decomposition), and acts by convolution on the \(K\)-invariant functions on \(X(F)\), with Haar measure normalized by \(\operatorname{vol}K=1\):
$$\mathcal H_K=C_c(K\backslash G(F)/K),\qquad T_\lambda=\mathbf 1_{Kt^\lambda K},\qquad(f\ast\varphi)(x)=\sum_{gK\in G(F)/K}f(g)\,\varphi(g^{-1}x),$$ $$C_K=C_c^\infty(X(F))^K,\qquad\varphi_0=\mathbf 1_{X(\mathcal O)}.$$Spherical generation.
$$\text{(SG)}\qquad\mathcal H_K\ast\varphi_0=C_K.$$For a field \(E\) of characteristic \(0\), \(\mathcal H_K\ast\varphi_0\otimes E\) is the \(E\)-span of the \(\mathbb Z\)-valued functions \(T_\lambda\ast\varphi_0\) and \(C_K\otimes E\) the \(K\)-invariant compactly supported \(E\)-valued functions; (SG) does not depend on \(E\) (IIb, step 7).
Other residue fields. Proposition II needs only \(F\). Since \(\overline{\mathbb F}_{q^n}=\overline{\mathbb F}_q\), replacing \(q\) by \(q^n\) leaves \(\mathcal C_{K,k}\) and its objects unchanged and gives the same statements over \(F_n=\mathbb F_{q^n}(\!(t)\!)\), with \(K_n=G(\mathbb F_{q^n}[\![t]\!])\), \(\mathcal H_{K_n}\), \(C_{K_n}\), \((\mathrm{SG})_n\), the trace \(\operatorname{tr}_n\) of the \(n\)-th iterate of \(\Phi\) and the matrix \(m^{(n)}\) (trace dictionary). Several \(n\) at once are used only for the converse (II, limits; Proposition IX).
Weil objects and trace functions
A Weil object is a pair \((M,\Phi)\) with \(M\in\mathcal C_{K,k}\) and \(\Phi:\mathrm{Fr}^*M\xrightarrow{\sim}M\), where \(\mathrm{Fr}\) is the \(\mathbb F_q\)-linear Frobenius endomorphism of \(LX\) (coordinates raised to the \(q\)-th power), so that \(LX(\mathbb F_q)=X(F)\) is its set of fixed points. For \(x\in X(F)\) and a geometric point \(\bar x\) over it, \(\mathrm{Fr}(\bar x)=\bar x\), so \(\Phi_{\bar x}\) acts on each stalk cohomology \(\mathcal H^i(M)_{\bar x}\), and the trace function of \(M\) is
$$\operatorname{tr}(M)(x)=\sum_i(-1)^i\operatorname{tr}\big(\Phi_{\bar x},\ \mathcal H^i(M)_{\bar x}\big)\in\overline{\mathbb Q}_\ell .$$The function \(\operatorname{tr}(M)\) is \(K\)-invariant and lies in \(C_K\otimes\overline{\mathbb Q}_\ell\) (IIb, steps 1–2); it is additive on Frobenius-compatible distinguished triangles, and \(\operatorname{tr}(M[1])=-\operatorname{tr}(M)\). \(V_\lambda\), \(\mathrm{IC}_0\) and \(\mathrm{IC}_\nu\) carry canonical Weil structures (eigenvalue \(1\) on the lowest stalk cohomology of the open stratum, no Tate twist), and \(V_\lambda\ast\mathrm{IC}_0\) the induced one. The trace dictionary: properties, conventions and sources →
Multiplicity spaces and the matrix \(m\)
By D2 each Satake object is a finite sum of shifted intersection complexes. With \(P_i={}^p\mathcal H^i(V_\lambda\ast\mathrm{IC}_0)\), the \(i\)-th perverse cohomology sheaf, the multiplicity space \(E_{\nu,i}(\lambda)\) counts the copies of \(\mathrm{IC}_\nu[-i]\):
$$V_\lambda\ast\mathrm{IC}_0\simeq\bigoplus_{\nu,i}\mathrm{IC}_\nu[-i]\otimes E_{\nu,i}(\lambda),\qquad E_{\nu,i}(\lambda)=\operatorname{Hom}\big(\mathrm{IC}_\nu,P_i\big).$$The Weil structure induces the Frobenius \(A_{\nu,i}(\lambda)\in\mathrm{GL}(E_{\nu,i}(\lambda))\) (IIb, step 4). The matrix has rows indexed by orbit labels and columns by Hecke labels:
$$m_{\nu\lambda}=\sum_i(-1)^i\operatorname{tr}A_{\nu,i}(\lambda)\qquad(\nu\in\Lambda_S^+,\ \lambda\in\Lambda_T^+).$$Each column \((m_{\nu\lambda})_\nu\) has finitely many nonzero entries. "Every row is nonzero" and "the columns span \(\bigoplus_{\Lambda_S^+}\overline{\mathbb Q}_\ell\)" are the two conditions compared on these pages.
Proposition I (summand criterion)
Assume H1–H3 and that every \(V_\lambda\ast\mathrm{IC}_0\) is a finite sum of shifted \(\mathrm{IC}_\nu\) with constant coefficients. Put
$$\mathcal M=\{\mu\in\Lambda_S^+:\mathrm{IC}_\mu[s]\text{ is a direct summand of some }V_\lambda\ast\mathrm{IC}_0,\ s\in\mathbb Z\}.$$Then
$$\mathcal C_K^{\mathrm{Sat}}=\operatorname{Thick}\{\mathrm{IC}_\nu:\nu\in\mathcal M\},\qquad\mathcal C_K^{\mathrm{Sat}}=\mathcal C_K\iff\mathcal M=\Lambda_S^+ .$$So \(\mathcal M\) is the set of labels whose intersection complex occurs, up to shift, as a direct summand of a Satake object, and the sheaf statement \(\mathcal C_K^{\mathrm{Sat}}=\mathcal C_K\) is the statement \(\mathcal M=\Lambda_S^+\). For \(\mathcal C_{K,k}\): H1 follows from D0, H2 from D1, H3 from the connectedness hypothesis. Proposition I · proof →
Hypotheses (R), (C) and inputs D0–D3, D5
(R) · Rationality.
\(G\), \(\theta\) and \(X\) are defined over \(\mathbb F_q\), \(q\) odd; \(T\supseteq S\) is split over \(\mathbb F_q\); and \(G\) has a \(\theta\)-stable Borel pair \((B_0,T_0)\) over \(\mathbb F_q\) with \(T_0\) split (Chen–Yi, §8.2). Then every \(t^\mu\), \(\mu\in\Lambda_S\), lies in \(S(F)\).(C) · Connectedness.
$$G^\theta\ \text{ and }\ Z_{G^\theta}(S)\ \text{ are connected}.$$D0 · Orbit geometry over \(k\).
For \(p\ne2\) (Chen–Yi, Prop. 17, Thm 18 and the pages' Proposition III): $$LX(k)=\bigsqcup_{\mu\in\Lambda_S^+}\mathcal O_\mu(k),\quad\tau(x_\mu)=t^\mu,\qquad\overline{\mathcal O_\mu}=\bigsqcup_{\nu\le\mu}\mathcal O_\nu,\qquad\operatorname{supp}(V_\lambda\ast\mathrm{IC}_0)\subseteq\overline{\mathcal O_{\varphi(\lambda)}},$$ $$\operatorname{Stab}_{L^+G}(x_\mu)=Z_{G^\theta}(\mu)\ltimes U_\mu,\ \ U_\mu\ \text{pro-unipotent},\qquad\pi_0Z_{G^\theta}(\mu)=1\ \text{ for all }\mu\ \text{ under (C)}.$$D1 · Traces and t-structure.
Proper base change and the Grothendieck–Lefschetz trace formula on the fibres of the convolution map; Chen–Yi's sheaf theory on \(L^+G\backslash LX\) (App. A.5) with a Frobenius-compatible perverse t-structure of finite length whose simple objects are the \(\mathrm{IC}_\nu\), \(\operatorname{End}(\mathrm{IC}_\nu)=\overline{\mathbb Q}_\ell\).D2 · Decomposition.
\(V_\lambda\ast\mathrm{IC}_0\) is pure, and over \(k\) $$V_\lambda\ast\mathrm{IC}_0\simeq\bigoplus_{\nu,i}\mathrm{IC}_\nu[-i]\otimes E_{\nu,i}(\lambda),\qquad\dim E_{\nu,i}(\lambda)\lt\infty$$ (Sun, Thm 1.2, on finite-type equivariant models; Chen–Yi, Prop. 70 and the proof of Thm 93).D3 · Lang's theorem.
For \(L^+G=\varprojlim_mG(\mathbb F_q[t]/t^m)\) and its stabilizers: $$H^1(\mathbb F_q,L^+G)=1,\qquad H^1\big(\mathbb F_q,\operatorname{Stab}_{L^+G}(x_\mu)\big)=H^1\big(\mathbb F_q,\pi_0\operatorname{Stab}_{L^+G}(x_\mu)\big).$$D5 · Comparison with \(\mathbb C\).
The statement "\(\mathrm{IC}_\mu\) is a summand of \(V_\lambda\ast\mathrm{IC}_0\)", which lives on the finite-type model \(\operatorname{Gr}_{\le\lambda}\mathbin{\widetilde\times}\overline{\mathcal O_0}\to\overline{\mathcal O_{\varphi(\lambda)}}\), has the same answer over \(\overline{\mathbb F}_q\) and over \(\mathbb C\) (BBD §6.1; Bezrukavnikov–Finkelberg, Prop. 5, for \(p\) outside a finite set). Used only for the Betti form of the conclusion.
§1 the argument · §2 traces of matrices · §3 traces of endofunctors · §4 classes and functoriality · §5 the Hecke square · §6 generation versus surjectivity · §7 three toy models · §8 the two geometric failures · §9 when the argument works · §10 what AGKRRV prove · Sources references.
Coefficients are \(\overline{\mathbb Q}_\ell\) throughout, and \(\mathrm{Vect}\) is the category of chain complexes of \(\overline{\mathbb Q}_\ell\)-vector spaces. \(\mathrm{Sph}_k\) is the spherical Hecke category: bounded \(L^+G\)-equivariant constructible complexes on \(\operatorname{Gr}\) over \(k=\overline{\mathbb F}_q\) with finite support, with the Satake sheaves \(V_\lambda\). It is written \(\mathcal H\) in the run. For a small category such as \(\mathcal C_{K,k}\), \(\operatorname{Tr}(-,\mathcal C_{K,k})\) means the trace on \(\operatorname{Ind}\mathcal C_{K,k}\), whose compact objects are \(\mathcal C_{K,k}\).
1. The argument to be tested
The argument
- The action map \(\mathrm{act}:\mathrm{Sph}_k\to\mathcal C_{K,k}\), \(F\mapsto F\ast\mathrm{IC}_0\), commutes with Frobenius.
- \(\mathcal C^{\mathrm{Sat}}_{K,k}=\mathcal C_{K,k}\): the objects \(\mathrm{act}(F)\) generate \(\mathcal C_{K,k}\) under shifts, cones and retracts.
- The categorical trace of Frobenius is functorial, so it produces a linear map \(\operatorname{Tr}(\mathrm{act}):\operatorname{Tr}(\mathrm{Fr}_*,\mathrm{Sph}_k)\to\operatorname{Tr}(\mathrm{Fr}_*,\mathcal C_{K,k})\).
- The two traces are \(\mathcal H_K\) and \(C_K\), and \(\operatorname{Tr}(\mathrm{act})\) is \(f\mapsto f\ast\varphi_0\).
- Hence \(\mathcal H_K\ast\varphi_0=C_K\).
Steps 1 and 3 are theorems (§4). Step 4 holds when the run's Theorem 1.3 applies (§5): this needs the rationality hypothesis (R) (split tori, rational orbit representatives), the connectedness hypothesis (C) (connected stabilizers), and the inputs D0–D3 (orbit geometry, base change with the trace formula, decomposition, Lang's theorem). Step 5 needs \(\operatorname{Tr}(\mathrm{act})\) to be onto, and step 2 does not give that. Generation is a statement about objects. Surjectivity of a trace map is a statement about morphisms \(c\to\mathrm{Fr}_*c\) (§6). Proposition IX shows that the gap is real (§8).
2. Traces in a symmetric monoidal category
Definition 2.1 (dualizable object, trace) [GKRV, §3.1.1]
Let \((\mathcal O,\otimes,\mathbf 1)\) be a symmetric monoidal category. An object \(o\) is dualizable if there are an object \(o^\vee\) and maps
$$\mathrm{unit}:\mathbf 1\to o\otimes o^\vee,\qquad\mathrm{counit}:o\otimes o^\vee\to\mathbf 1$$satisfying the two zigzag identities. The trace of an endomorphism \(F\) of \(o\) is the composite
$$\operatorname{Tr}(F,o)=\Big(\mathbf 1\xrightarrow{\ \mathrm{unit}\ }o\otimes o^\vee\xrightarrow{\ F\otimes\mathrm{id}\ }o\otimes o^\vee\xrightarrow{\ \mathrm{counit}\ }\mathbf 1\Big)\in\operatorname{End}_{\mathcal O}(\mathbf 1).$$Example 2.2. In finite-dimensional vector spaces, take \(o=V\) and \(o^\vee=V^*\), with \(\mathrm{unit}(1)=\sum_ie_i\otimes e^i\) and \(\mathrm{counit}(v\otimes\xi)=\xi(v)\). Then \(\operatorname{Tr}(F,V)=\sum_ie^i(Fe_i)\) is the usual trace. The dualizable objects are exactly the finite-dimensional spaces. For chain complexes they are the perfect complexes, with \(\operatorname{Tr}(F)=\sum_i(-1)^i\operatorname{tr}H^i(F)\).
Properties 2.3 [GKRV, §3.1.2, §3.1.4]. For \(f:o_1\to o_2\) and \(g:o_2\to o_1\):
$$\operatorname{Tr}(F_1\otimes F_2,o_1\otimes o_2)=\operatorname{Tr}(F_1,o_1)\cdot\operatorname{Tr}(F_2,o_2),\qquad\operatorname{Tr}(f\circ g,o_2)=\operatorname{Tr}(g\circ f,o_1).$$At this level a trace is a number, and nothing compares the traces of two different objects. One categorical level up, traces become vector spaces, and functors between categories give maps between them.
3. One level up: the trace of an endofunctor
Definition 3.1 (the categorical trace) [GKRV, §3.4.1; AGKRRV3, §0.2.2; GR1, Ch. 1]
\(\mathrm{DGCat}\) is the symmetric monoidal 2-category whose objects are the presentable stable \(\overline{\mathbb Q}_\ell\)-linear categories, whose 1-morphisms are the colimit-preserving (continuous) functors, and whose 2-morphisms are the natural transformations. Its tensor product is Lurie's, and its unit is \(\mathrm{Vect}\). Since \(\operatorname{End}(\mathrm{Vect})=\mathrm{Vect}\), the trace of a continuous endofunctor \(F\) of a dualizable \(\mathcal C\) is a chain complex:
$$\operatorname{Tr}(F,\mathcal C)=\operatorname{ev}_{\mathcal C}\circ(F\otimes\mathrm{Id})\circ u_{\mathcal C}\ \in\ \mathrm{Vect}.$$Proposition 3.2 (the trace is Hochschild homology) [GKRV, §3.5.1]
A compactly generated \(\mathcal C\) is dualizable, with \(\mathcal C^\vee=\operatorname{Ind}((\mathcal C^c)^{\mathrm{op}})\). Let \(g\) be a compact generator and \(E=\operatorname{End}(g)\), so that \(\operatorname{Hom}(g,-)\) identifies \(\mathcal C\) with right \(E\)-modules. A continuous endofunctor \(F\) corresponds to the \(E\)-bimodule \(Q=\operatorname{Hom}(g,Fg)\), with \(a\cdot q\cdot b=F(a)\circ q\circ b\). Then
$$\operatorname{Tr}(F,\mathcal C)=E\otimes^{\mathbf L}_{E\otimes E^{\mathrm{op}}}Q=\mathrm{HH}_\bullet(E;Q).$$If \(E\) and \(Q\) are concentrated in degree \(0\), then \(\mathrm{HH}_\bullet\) lives in degrees \(\ge0\), and
$$H_0\operatorname{Tr}(F,\mathcal C)=Q\big/\big\langle F(a)\circ q-q\circ a\ :\ a\in E,\ q\in Q\big\rangle .$$The left side is defined by \((\mathcal C,F)\) alone, so the right side does not depend on the choice of generator (Morita invariance). With several generators \(g_i\), replace \(E\) by the category on the \(g_i\) and \(Q\) by \(Q(i,j)=\operatorname{Hom}(g_i,Fg_j)\).
Example 3.3 (this page's case). For \(\mathcal C=\operatorname{Ind}\mathcal C_{K,k}\) and \(F=\mathrm{Fr}_*\), we have \(\mathrm{Fr}_*\simeq(\mathrm{Fr}^*)^{-1}\), since Frobenius is a universal homeomorphism. Proposition 3.2 gives the Frobenius-twisted Hochschild homology of the run's Theorem 1.3:
$$\operatorname{Tr}(\mathrm{Fr}_*,\mathcal C_{K,k})=\mathrm{HH}_\bullet\big(\mathcal C_{K,k};\ (M,N)\mapsto\operatorname{Hom}(M,\mathrm{Fr}_*N)=\operatorname{Hom}(\mathrm{Fr}^*M,N)\big).$$4. Classes and functoriality
Definition 4.1 (the class of an object) [GKRV, §3.4.3; AGKRRV3, Remark 0.6.5]
Let \(c\in\mathcal C\) and \(\alpha:c\to F(c)\). The functor \(T_c:\mathrm{Vect}\to\mathcal C\), \(V\mapsto V\otimes c\), has a continuous right adjoint exactly when \(c\) is compact, and \(\alpha\) is a 2-morphism \(T_c\circ\mathrm{Id}\Rightarrow F\circ T_c\). Its trace (Definition 4.2) is a map \(\overline{\mathbb Q}_\ell=\operatorname{Tr}(\mathrm{Id},\mathrm{Vect})\to\operatorname{Tr}(F,\mathcal C)\), that is, an element
$$\operatorname{cl}(c,\alpha)\in H_0\operatorname{Tr}(F,\mathcal C).$$In the model of Proposition 3.2, \(\operatorname{cl}(g,q)\) is the class of \(q\in Q\). For \(F=\mathrm{Fr}_*\), adjunction turns \(\alpha:M\to\mathrm{Fr}_*M\) into \(\alpha^\flat:\mathrm{Fr}^*M\to M\). A Weil object \((M,\Phi)\) is the case where \(\alpha^\flat=\Phi\) is invertible, and we write \(\operatorname{cl}(M,\Phi)\). AGKRRV allow any \(\alpha\) ("possibly lax Weil sheaves").
Definition 4.2 (the trace of a functor) [GKRV, §3.2.1, §3.4.1]
A 1-morphism \((\Psi,\gamma):(\mathcal C,F_{\mathcal C})\to(\mathcal D,F_{\mathcal D})\) consists of two things.
- A continuous functor \(\Psi:\mathcal C\to\mathcal D\) with a continuous right adjoint. For compactly generated categories this means that \(\Psi\) preserves compact objects.
- A natural transformation \(\gamma:\Psi\circ F_{\mathcal C}\Rightarrow F_{\mathcal D}\circ\Psi\), not necessarily invertible.
Pasting \(\gamma\) and the adjunction \((\Psi,\Psi^R)\) into the unit–trace–counit composite gives a linear map
$$\operatorname{Tr}(\Psi,\gamma):\operatorname{Tr}(F_{\mathcal C},\mathcal C)\longrightarrow\operatorname{Tr}(F_{\mathcal D},\mathcal D).$$GKRV follow Ben-Zvi–Nadler and Kondyrev–Prikhodko for this construction.
Theorem 4.3 (functoriality) [GKRV, §3.2.3, §3.3.1, §3.4.2–3.4.3]
- Composition. \(\operatorname{Tr}(\Psi'\circ\Psi,\gamma'\circ\gamma)=\operatorname{Tr}(\Psi',\gamma')\circ\operatorname{Tr}(\Psi,\gamma)\). More precisely, \(\operatorname{Tr}\) is a symmetric monoidal functor from pairs (dualizable category, endofunctor), with these 1-morphisms, to \(\mathrm{Vect}\).
- Classes go to classes. \(\operatorname{Tr}(\Psi,\gamma)\big(\operatorname{cl}(c,\alpha)\big)=\operatorname{cl}\big(\Psi c,\ \gamma_c\circ\Psi(\alpha)\big)\).
- Additivity and cyclicity. \(\operatorname{cl}(c\oplus c',\alpha\oplus\alpha')=\operatorname{cl}(c,\alpha)+\operatorname{cl}(c',\alpha')\). For \(u:c\to c'\) and \(\beta:c'\to F(c)\), \(\operatorname{cl}(c,\beta\circ u)=\operatorname{cl}(c',F(u)\circ\beta)\).
- Vector spaces. For \(\mathcal C=\mathrm{Vect}\), \(F=\mathrm{Id}\) and \(V\) perfect, \(\operatorname{cl}(V,\alpha)=\sum_i(-1)^i\operatorname{tr}H^i(\alpha)\).
Proof
- (b). \(\operatorname{cl}(c,\alpha)=\operatorname{Tr}(T_c,\alpha)\), and \((\Psi,\gamma)\circ(T_c,\alpha)=(T_{\Psi c},\gamma_c\circ\Psi(\alpha))\). Apply (a).
- (c). In the model of Proposition 3.2, take a generator containing \(c\oplus c'\) as a summand. Then \(q\mapsto[q]\) is linear, and the cyclicity identity is the defining relation \(F(a)\circ q\sim q\circ a\) of \(H_0\). The 2-categorical form is [GKRV, §3.2.4].
- (d). \(\operatorname{Tr}(\mathrm{Id},\mathrm{Vect})=\overline{\mathbb Q}_\ell\) [GKRV, §3.4.2], and \(\operatorname{Tr}(T_V,\alpha)\) is the trace of Definition 2.1 for the dualizable object \(V\).
Proposition 4.4 (traces of module categories) [GKRV, §3.3.2–3.3.3, §3.4.4, Proposition 3.4.6]
Let a monoidal \(\mathcal R\) act on \(\mathcal M\), with multiplication and action preserving compact objects. Let \(F_{\mathcal R}\) be right-lax monoidal and \(F_{\mathcal M}\) right-lax compatible with the action. Then \(\operatorname{Tr}(F_{\mathcal R},\mathcal R)\) is an associative algebra and \(\operatorname{Tr}(F_{\mathcal M},\mathcal M)\) a module over it. For compact \(r\) with \(\alpha:r\to F_{\mathcal R}(r)\), the class \(\operatorname{cl}(r,\alpha)\) acts by \(\operatorname{Tr}(H_r,\alpha_{r,\mathcal M})\), with \(H_r=r\ast-\).
Corollary 4.5 (the trace of the action map)
Take \(\mathcal R=\operatorname{Ind}\mathrm{Sph}_k\), \(\mathcal M=\operatorname{Ind}\mathcal C_{K,k}\) and \(F=\mathrm{Fr}_*\) on both, and let \(\Phi_0\) be the canonical Weil structure of \(\mathrm{IC}_0\). Then
$$\operatorname{Tr}(\mathrm{act})(x)=x\cdot\operatorname{cl}(\mathrm{IC}_0,\Phi_0)\qquad(x\in\operatorname{Tr}(\mathrm{Fr}_*,\mathrm{Sph}_k)),$$ $$\operatorname{Tr}(\mathrm{act})\ \text{is onto}\iff\operatorname{Tr}(\mathrm{Fr}_*,\mathcal C_{K,k})\ \text{is generated by}\ \operatorname{cl}(\mathrm{IC}_0,\Phi_0)\ \text{as a}\ \operatorname{Tr}(\mathrm{Fr}_*,\mathrm{Sph}_k)\text{-module}.$$Proof
- Factorization. \(\mathrm{act}\) is the composite \(\mathrm{Sph}_k\simeq\mathrm{Sph}_k\otimes\mathrm{Vect}\xrightarrow{\mathrm{Id}\otimes T_{\mathrm{IC}_0}}\mathrm{Sph}_k\otimes\mathcal C_{K,k}\xrightarrow{\ \ast\ }\mathcal C_{K,k}\), compatibly with Frobenius.
- Traces. By Theorem 4.3(a), \(\operatorname{Tr}\) is symmetric monoidal, so the trace of the middle functor is \(x\mapsto x\otimes\operatorname{cl}(\mathrm{IC}_0,\Phi_0)\). The trace of the last functor is the module structure of Proposition 4.4.
| categorical | after the trace | |
|---|---|---|
| statement | \(\mathrm{IC}_0\) generates \(\mathcal C_{K,k}\) as a \(\mathrm{Sph}_k\)-module category: \(\mathcal C^{\mathrm{Sat}}_{K,k}=\mathcal C_{K,k}\) | \(\operatorname{cl}(\mathrm{IC}_0)\) generates \(\operatorname{Tr}(\mathrm{Fr}_*,\mathcal C_{K,k})\) as a module: (SG), by §5 |
| built from | objects \(V_\lambda\ast\mathrm{IC}_0\), every \(\lambda\) | classes \(\operatorname{cl}(V_\lambda\ast\mathrm{IC}_0,\gamma\circ\mathrm{act}(\alpha))\), \(\alpha:V_\lambda\to\mathrm{Fr}_*V_\lambda\) |
| operations | shifts, cones, retracts | linear combinations |
| sees | objects up to isomorphism | morphisms \(c\to\mathrm{Fr}_*c\), modulo cyclicity |
5. The Hecke square
Definition 5.1 (local terms) [AGKRRV1, §22.2; AGKRRV3, (0.9) and Remark 0.6.5]
For a quasi-compact algebraic stack \(Y\) over \(\mathbb F_q\), there is a canonical local-term map \(\mathrm{LT}_Y:\operatorname{Tr}((\mathrm{Frob}_Y)_*,\mathrm{Shv}(Y))\to\mathrm{Funct}(Y(\mathbb F_q))\). On classes it is the Grothendieck–Deligne function:
$$\mathrm{LT}_Y\big(\operatorname{cl}(M,\alpha)\big)(x)=\sum_i(-1)^i\operatorname{tr}\big(\alpha^\flat_{\bar x},\ \mathcal H^i(M)_{\bar x}\big).$$On the loop-space category, the finite stabilizer presentations and rational stalk functors used in Proposition 5.2(c) define the local-term map on the full trace complex. Its value on a Weil class is the stalk trace; checking degree-zero cyclicity alone would not construct this map for an arbitrary differential graded category.
Proposition 5.2 (the square)
$$\begin{CD}\operatorname{Tr}(\mathrm{Fr}_*,\mathrm{Sph}_k) @>{\operatorname{Tr}(\mathrm{act})}>> \operatorname{Tr}(\mathrm{Fr}_*,\mathcal C_{K,k})\\ @V{\mathrm{LT}_{\mathrm{Sph}}}VV @VV{\mathrm{LT}_{\mathcal C_K}}V\\ \mathcal H_K @>{f\,\mapsto\, f\ast\varphi_0}>> C_K\end{CD}$$- The square commutes on classes of Weil objects: \(\operatorname{tr}(F\ast\mathrm{IC}_0)=\operatorname{tr}(F)\ast\varphi_0\).
- Suppose every Frobenius-stable orbit has connected stabilizer. This holds under (C) and in the examples of Proposition IX. Then \(\mathrm{LT}_{\mathrm{Sph}}\) and \(\mathrm{LT}_{\mathcal C_K}\) are onto, and hence, with \(W=\operatorname{span}\{\operatorname{cl}(V_\lambda\ast\mathrm{IC}_0,\Phi):\sigma\lambda=\lambda\}\subseteq\operatorname{Tr}(\mathrm{Fr}_*,\mathcal C_{K,k})\), the span of the classes that Weil Satake objects produce, $$\mathrm{LT}_{\mathcal C_K}(W)=\mathcal H_K\ast\varphi_0,\qquad(\mathrm{SG})\iff\mathrm{LT}_{\mathcal C_K}(W)=C_K .$$
- Assume (R), (C) and D0–D3. Then both traces are concentrated in degree \(0\) and spanned by classes, both local-term maps are isomorphisms, and $$(\mathrm{SG})\iff\operatorname{Tr}(\mathrm{act})\ \text{is onto}.$$
Proof
- (a). The Grothendieck–Lefschetz trace formula and proper base change on the fibres of the convolution map: IIa, and the convolution row of the trace dictionary.
- (b). The basis elements \(T_\lambda\) (\(\sigma\lambda=\lambda\)) and the orbit indicators \(\mathbf 1_{\mathcal O_\nu(\mathbb F_q)}\) are the trace functions of the constant sheaves on Frobenius-stable cells and orbits, extended by zero with their canonical Weil structure. Lang's theorem and connectedness of the stabilizers make each set of rational points one \(K\)-orbit. Every \(K\)-orbit is of this form, since an orbit with a rational point is Frobenius-stable. For \(W\): by Lemma 2.8, \(V_\lambda\) is Weil iff \(\sigma\lambda=\lambda\), and \(\mathcal H_K\ast\varphi_0\) is spanned by the \(\operatorname{tr}(V_\lambda\ast\mathrm{IC}_0)\) with \(\sigma\lambda=\lambda\).
- (c). On a single orbit with connected stabilizer, finite-level equivariant descent gives \(\operatorname{Perf}C^*(BL)\), where \(L\) is its reductive quotient; the pro-unipotent kernel contributes no cohomology. The free even cohomology algebra has generators \(v_i\) in degree \(2d_i\), with Frobenius eigenvalues \(q^{d_i}\). Choosing cocycle representatives gives a polynomial differential graded model, compatible with Frobenius up to homotopy. Its twisted Hochschild complex is the Koszul complex on
$$ (1-q^{d_i})v_i. $$
These form a regular sequence, so the trace is one copy of the coefficient field in degree zero. Frobenius-stable finite orbit filtrations give triangular differential graded gluings: a mixed cyclic-bar term would need a morphism in both directions between strata, and is zero. Induction therefore gives one trace line for each rational orbit. The compatible finite presentations in D1 identify the loop-space category with these finite computations. Rational stalk functors define local terms on the entire trace complex; convolution and proper base change give the action square. See the finite-stabilizer calculation.
So steps 1, 3 and 4 of §1 are sound. Everything hinges on whether \(\operatorname{Tr}(\mathrm{act})\) is onto.
6. Generation versus surjectivity
Theorem 6.1 (the image of a trace map, in degree 0)
Let \((\Psi,\gamma):(\mathcal C,F_{\mathcal C})\to(\mathcal D,F_{\mathcal D})\) be as in Definition 4.2. Let \(\{c_i\}\) be compact objects with
$$\mathcal C^c=\operatorname{Thick}\{c_i\},\qquad\mathcal D^c=\operatorname{Thick}\{\Psi c_i\},$$and suppose all the spaces \(\operatorname{Hom}(c_i,c_j)\), \(\operatorname{Hom}(c_i,F_{\mathcal C}c_j)\), \(\operatorname{Hom}(\Psi c_i,\Psi c_j)\) and \(\operatorname{Hom}(\Psi c_i,F_{\mathcal D}\Psi c_j)\) are concentrated in degree \(0\). Then:
- \(H_0\operatorname{Tr}(F_{\mathcal D},\mathcal D)=\bigoplus_i\operatorname{Hom}(\Psi c_i,F_{\mathcal D}\Psi c_i)\big/\big\langle q\circ f-F_{\mathcal D}(f)\circ q\big\rangle\), where \(f:\Psi c_i\to\Psi c_j\) and \(q:\Psi c_j\to F_{\mathcal D}\Psi c_i\). Every element is a sum of classes \(\operatorname{cl}(\Psi c_i,q)\).
- \(\operatorname{im}H_0\operatorname{Tr}(\Psi,\gamma)=\operatorname{span}\big\{\operatorname{cl}\big(\Psi c_i,\gamma\circ\Psi(\alpha)\big):\alpha\in\operatorname{Hom}(c_i,F_{\mathcal C}c_i)\big\}\).
- Hence \(\operatorname{Tr}(\Psi,\gamma)\) is onto in degree \(0\) iff $$\bigoplus_i\operatorname{Hom}(\Psi c_i,F_{\mathcal D}\Psi c_i)=\bigoplus_i\gamma\circ\Psi\big(\operatorname{Hom}(c_i,F_{\mathcal C}c_i)\big)+\big\langle q\circ f-F_{\mathcal D}(f)\circ q\big\rangle .$$
Proof
- Step 1 (Morita). \(\mathcal D^c=\operatorname{Thick}\{\Psi c_i\}\) makes \(\mathcal D\) the category of modules over the full subcategory on the \(\Psi c_i\), an ordinary linear category by the degree hypothesis. Proposition 3.2 with several generators computes the trace as its Hochschild homology with coefficients in \(Q(i,j)=\operatorname{Hom}(\Psi c_i,F_{\mathcal D}\Psi c_j)\). In degree \(0\) the cyclic bar complex gives the displayed quotient, and \(q\in Q(i,i)\) maps to \(\operatorname{cl}(\Psi c_i,q)\) by Definition 4.1. This proves (i).
- Step 2. The same argument for \(\mathcal C\) shows that \(H_0\operatorname{Tr}(F_{\mathcal C},\mathcal C)\) is spanned by the \(\operatorname{cl}(c_i,\alpha)\).
- Step 3. Theorem 4.3(b) sends \(\operatorname{cl}(c_i,\alpha)\) to \(\operatorname{cl}(\Psi c_i,\gamma_{c_i}\circ\Psi(\alpha))\). This proves (ii), and (iii) follows from (i) and (ii).
Reading Theorem 6.1. The generation hypothesis \(\mathcal D^c=\operatorname{Thick}\{\Psi c_i\}\) is used only to compute \(\operatorname{Tr}(F_{\mathcal D},\mathcal D)\) from the \(\Psi c_i\). Whether the map is onto is then decided by (iii), which concerns morphisms. Twisted endomorphisms of \(\Psi c_i\) must come from \(\mathcal C\), up to commutators. This is a fullness condition, and generation (essential surjectivity up to cones and retracts) does not imply it. It fails in two ways:
- (M1) Missing Weil structures. A generator \(c_i\) with \(\operatorname{Hom}(c_i,F_{\mathcal C}c_i)=0\) contributes no class, but its image may be fixed by \(F_{\mathcal D}\) and carry twisted endomorphisms. Generation uses \(c_i\); the trace sees only pairs \((c_i,\alpha)\). Generation without retracts does not help either: in Toy B below, \(\Psi\) is essentially surjective on the nose.
- (M2) Retracts. Suppose \(\Psi c_i\simeq N\oplus N'\) in \(\mathcal D\), and the projector \(e\) onto \(N\) is not in \(\Psi(\operatorname{End}c_i)\) modulo commutators. Then \(\operatorname{cl}(N)\) is not reached. The thick closure contains \(N\), while the trace contains only the combinations that \(\Psi\) produces, such as \(\operatorname{cl}(N)+\operatorname{cl}(N')\).
Without the degree-zero Hom hypothesis, the ordinary commutator quotient in Theorem 6.1 need not compute \(H_0\): higher cyclic-bar terms can contribute [AGKRRV3, Remark 0.6.11]. The equivariant loop categories have higher Ext groups, so Proposition 5.2(c) uses its separate stabilizer-and-gluing calculation; concentration of the resulting trace in degree zero does not force the Hom complexes to lie in degree zero.
7. Three toy models
Each model is computed in full from Theorem 6.1. The categories are semisimple and the generators lie in the heart, so the degree-0 hypothesis holds. These are the run's Examples 1.4.
Toy A (a retract): \(\Psi:\mathrm{Vect}\to\mathrm{Vect}\times\mathrm{Vect}\), \(V\mapsto(V,V)\), \(F=\mathrm{Id}\)
$$\operatorname{Tr}(\mathrm{Id},\mathrm{Vect})=\overline{\mathbb Q}_\ell\ \xrightarrow{\ \operatorname{Tr}(\Psi)\ }\ \operatorname{Tr}(\mathrm{Id},\mathrm{Vect}^{2})=\overline{\mathbb Q}_\ell\operatorname{cl}(\overline{\mathbb Q}_\ell,0)\oplus\overline{\mathbb Q}_\ell\operatorname{cl}(0,\overline{\mathbb Q}_\ell),\qquad1\mapsto(1,1).$$- Generation. \((\overline{\mathbb Q}_\ell,0)\) is a retract of \(\Psi(\overline{\mathbb Q}_\ell)\).
- Image. The diagonal. In 6.1(iii), \(\operatorname{End}(\Psi\overline{\mathbb Q}_\ell)=\overline{\mathbb Q}_\ell^2\supsetneq\Psi(\operatorname{End}\overline{\mathbb Q}_\ell)=\overline{\mathbb Q}_\ell(1,1)\), with no commutators.
- Missing class. The idempotent \((1,0)\), which is mechanism (M2).
- The same toy as modules. \(\mathrm{Vect}^2=(\overline{\mathbb Q}_\ell\times\overline{\mathbb Q}_\ell)\text{-mod}\) and \(\Psi\) is the free-module functor. In the same way, \(\mathcal C^{\mathrm{Sat}}_{K,k}=\mathcal C_{K,k}\) makes \(\operatorname{Ind}\mathcal C_{K,k}\) the category of \(A\)-modules in \(\operatorname{Ind}\mathrm{Sph}_k\) and \(\mathrm{act}\) the free-module functor (Theorem 1.3(d)).
- Variant A′. \(\Psi:\mathrm{Vect}\to\operatorname{Perf}(\overline{\mathbb Q}_\ell[x]/x^2)\) misses the nilpotent class \(\operatorname{cl}(\overline{\mathbb Q}_\ell[x]/x^2,x)\). If Frobenius acts on \(x\) by \(q\ne1\), the relation \(x\sim qx\) kills that class and the map is onto in degree \(0\). Frobenius weights remove nilpotent classes, not idempotent ones.
Toy B (no Weil structure): \(\Psi:(\mathrm{Vect}^2,\mathrm{swap})\to(\mathrm{Vect},\mathrm{Id})\), \((V_1,V_2)\mapsto V_1\oplus V_2\)
Here \(\gamma:V_2\oplus V_1\xrightarrow{\sim}V_1\oplus V_2\) is the canonical isomorphism.
$$\operatorname{Tr}(\mathrm{swap},\mathrm{Vect}^2)=0\ \xrightarrow{\ \operatorname{Tr}(\Psi,\gamma)\ }\ \operatorname{Tr}(\mathrm{Id},\mathrm{Vect})=\overline{\mathbb Q}_\ell .$$- Generation. \(\Psi(\overline{\mathbb Q}_\ell,0)=\overline{\mathbb Q}_\ell\), even without retracts.
- Source trace. For \(c=(\overline{\mathbb Q}_\ell,\overline{\mathbb Q}_\ell)\), \(\operatorname{Hom}(c,\mathrm{swap}\,c)\ni(q_1,q_2)\). The relation with \(f=(1,0)\) is \(q\circ f-\mathrm{swap}(f)\circ q=(q_1,-q_2)\), which spans everything, so \(H_0=0\).
- Image of a Weil object. \(\alpha=(\phi_1,\phi_2)\) goes to \(\gamma\circ\Psi(\alpha)=\begin{pmatrix}0&\phi_2\\\phi_1&0\end{pmatrix}\), of trace \(0\). In 6.1(iii) the image consists of the off-diagonal matrices, which lie in \([M_2,M_2]=\{\operatorname{tr}=0\}\ne M_2\).
- Missing class. \((\overline{\mathbb Q}_\ell^2,\mathrm{swap})=(\overline{\mathbb Q}_\ell,+1)\oplus(\overline{\mathbb Q}_\ell,-1)\). The Weil object \((\overline{\mathbb Q}_\ell,+1)\) is a summand of an image, but its class \(1\) is not reached; the image class is \(1+(-1)=0\). This is mechanism (M1).
- Squares. For \(F_{\mathcal C}=\mathrm{swap}^2=\mathrm{Id}\) the map is \((a,b)\mapsto a+b\), which is onto.
Toy C (fusion): \(\operatorname{Res}:\operatorname{Rep}\mathrm{SL}_2\to\operatorname{Rep}T\), \(T\) the diagonal torus, \(F=\mathrm{Id}\)
$$\operatorname{Tr}(\mathrm{Id},\operatorname{Rep}\mathrm{SL}_2)=\overline{\mathbb Q}_\ell[z+z^{-1}]\ \xrightarrow{\ \text{restriction of characters}\ }\ \operatorname{Tr}(\mathrm{Id},\operatorname{Rep}T)=\overline{\mathbb Q}_\ell[z^{\pm1}].$$- Generation. The weight \(z^m\) is a summand of \(\operatorname{Res}V_{|m|}\).
- Traces. For a semisimple \(\operatorname{Rep}H\), Theorem 6.1(i) gives one class per irreducible, \(\operatorname{Tr}=R(H)\otimes\overline{\mathbb Q}_\ell\), the class functions. \(\operatorname{Tr}(\operatorname{Res})\) is restriction.
- Image. The \(W\)-invariant Laurent polynomials.
- Missing class. \(\operatorname{cl}(z)\). It is the weight projector of \(\operatorname{Res}V_1=z\oplus z^{-1}\), which is not in \(\operatorname{End}_{\mathrm{SL}_2}V_1=\overline{\mathbb Q}_\ell\): mechanism (M2).
- Why. \(t\) and \(t^{-1}\) in \(T\) are conjugate in \(\mathrm{SL}_2\) but not in \(T\): two classes fuse. For finite groups \(H\subseteq G\), restriction of class functions is onto iff no two \(H\)-classes fuse in \(G\). The smallest instance is \(\mathbb Z/3\subset S_3\).
| toy | generation uses | \(\operatorname{Tr}(\Psi)\) | missing class | instance in Proposition IX |
|---|---|---|---|---|
| A | a retract | \(\overline{\mathbb Q}_\ell\to\overline{\mathbb Q}_\ell^2\), diagonal | an idempotent | the general shape: \(\mathrm{act}\) is a free-module functor |
| B | an object with no Weil structure | \(0\to\overline{\mathbb Q}_\ell\) | \((\overline{\mathbb Q}_\ell,+1)\) | (a) unitary twist, odd \(n\) |
| C | retracts (weights) | restriction of class functions | a weight projector | (b) Sakellaridis's \(X\), every \(n\) |
8. The two geometric failures
8.1 The unitary twist: Toy B with \(\mathrm{Vect}\) replaced by \(\mathrm{Sph}_H\)
Let \(G=\operatorname{Res}_{\mathbb F_{q^2}/\mathbb F_q}H\) with \(\theta=\vartheta\circ\sigma\) (Proposition IX(a)). Over \(k\), \(G_k=H\times H\) and Frobenius exchanges the factors, so \(\mathrm{Sph}_{G}=\mathrm{Sph}_H\otimes\mathrm{Sph}_H\) with \(\mathrm{Fr}_*=\mathrm{swap}\circ(\mathrm{Fr}_*\otimes\mathrm{Fr}_*)\).
- Hecke side. The trace of a swap is the trace of the square, by the same string diagram as \(\operatorname{tr}(\mathrm{swap}\circ(f_1\otimes f_2))=\operatorname{tr}(f_2f_1)\). So \(\operatorname{Tr}(\mathrm{Fr}_*,\mathrm{Sph}_G)\) is the Hecke algebra of \(H(\mathbb F_{q^2}(\!(t)\!))=G(F)\), as it must be.
- Generation. \(X_k\) is the group case, and \(V_{(c,0)}\ast\mathrm{IC}_0\simeq\mathrm{IC}_c\) for every \(c\in\Lambda_H^+\). The label \((c,0)\) is not Frobenius-fixed: \(\mathrm{Fr}^*V_\lambda\simeq V_{\sigma^{-1}\lambda}\), so \(V_\lambda\) is Weil iff \(\sigma\lambda=\lambda\) (Lemma 2.8).
- Traces. Weil Satake objects are built from fixed labels and from orbits \(\{\lambda,\sigma\lambda\}\) with the swap structure. Like \((\overline{\mathbb Q}_\ell^2,\mathrm{swap})\) in Toy B, such an orbit contributes \(0\) to \(\operatorname{tr}_n\) for odd \(n\). The objects \(V_{(c,0)}\ast\mathrm{IC}_0\) do carry Weil structures in \(\mathcal C_{K,k}\), but for \(c\notin2\Lambda_H\) none of them comes from \(\mathrm{Sph}_G\).
- Outcome. \(\mathbf 1_{O_c}\notin\mathcal H_K\ast\varphi_0\) for \(c\notin2\Lambda_H\) (Theorem 2.6, audited twice). For even \(n\), \(\mathrm{Fr}^n\) fixes every label and \((\mathrm{SG})_n\) holds, as in Toy B with \(\mathrm{swap}^2\).
8.2 Sakellaridis's \(X=\mathrm{SL}_2^3\backslash(\mathrm{Sp}_4\times\mathrm{Sp}_4)\): Toy C with \(\mathrm{SL}_2\supset T\) replaced by \(\mathrm{SO}_5^2\supset\iota\check G_X\)
Here \(G\) is split and every label is Frobenius-fixed, so every \(V_\lambda\) is Weil and (M1) cannot occur.
- Multiplicities. Given the decomposition theorem for \(X\) (the hypothesis of Corollary 3.9), \(V_\lambda\ast\mathrm{IC}_0\simeq\bigoplus_\nu\mathrm{IC}_\nu^{\,c_{\nu\lambda}}\), where \(c_{\nu\lambda}=[\operatorname{Res}^{\mathrm{SO}_5\times\mathrm{SO}_5}_{\iota\check G_X}V_\lambda:V_\nu]\) are branching numbers (Proposition 3.8), and \(m^{(n)}_{\nu\lambda}=c_{\nu\lambda}\,q^{n(\langle2\rho,\lambda\rangle-\delta(\nu))/2}\). Up to these diagonal powers of \(q\), \(\operatorname{Tr}(\mathrm{act})\) is restriction \(R(\check G)\otimes\overline{\mathbb Q}_\ell\to R(\iota\check G_X)\otimes\overline{\mathbb Q}_\ell\), exactly as in Toy C.
- Fusion. For \(a\ne\pm b^{\pm1}\), the points \((a,b,a)\) and \((b,a,b)\) of \(\check G_X\) are \(\check G\)-conjugate but not \(\check G_X\)-conjugate (Lemma 3.3(d)). Restriction is therefore not onto, and, given the same decomposition, \(\operatorname{tr}\mathrm{IC}_{(0,0,2)}\notin\mathcal H_{K_n}\ast\varphi_0\) for every \(n\) (Corollary 3.9(d)): mechanism (M2).
- Status. The function failure is proved unconditionally, by two independent periods (Theorem 3.5, audited). The sheaf half \(\mathcal C^{\mathrm{Sat}}_{K,k}=\mathcal C_{K,k}\) needs Chen–Yi's framework for this \(X\): Sakellaridis–Wang's degeneration in odd characteristic, the dimension theory \(\delta=3d_1+3d_2+e\), and Chen–Yi's Appendix A with Prop. 70 (obligations 5.10).
9. When the argument does work
Proposition 9.1 (split surjections)
Suppose \((\Psi',\gamma'):(\mathcal D,F_{\mathcal D})\to(\mathcal C,F_{\mathcal C})\) is a 1-morphism with \((\Psi,\gamma)\circ(\Psi',\gamma')\simeq(\mathrm{Id},\mathrm{id})\). Then \(\operatorname{Tr}(\Psi,\gamma)\) is split onto: \(\operatorname{Tr}(\Psi,\gamma)\circ\operatorname{Tr}(\Psi',\gamma')=\operatorname{Tr}(\mathrm{Id})=\mathrm{id}\) by Theorem 4.3(a). An instance is a Frobenius-compatible localization whose fully faithful right adjoint preserves compact objects. \(\mathrm{act}\) is not of this kind: it is a free-module functor (Toy A).
Proposition 9.2 (dominant lifts) [run's Proposition 1.8(e)–(f), audited]
Assume (R), (C) and D0–D3. If every orbit label \(\mu\) has a dominant lift, i.e. \(\mu=\varphi(\lambda_\mu)\) for some \(\lambda_\mu\in\Lambda_T^+\), then the columns \(\lambda_\mu\) of \(m^{(n)}\) are triangular with nonzero diagonal entries \(m^{(n)}_{\mu\lambda_\mu}\), and \((\mathrm{SG})_n\) holds for every \(n\).
In terms of §6: \(\mathrm{IC}_\mu\) is the top summand of \(V_{\lambda_\mu}\ast\mathrm{IC}_0\), with nonzero top coefficient, so \(\operatorname{cl}(\mathrm{IC}_\mu)\) is reached modulo lower labels without any projector. A triangular system of summands whose multiplicity spaces are split Tate of one parity suffices. Without dominant lifts the trace argument decides nothing: with \(U\) the labels without a lift, \((\mathrm{SG})\iff\{p_U(T_\lambda\ast\varphi_0)\}\) spans \(\mathbb C^{(U)}\), and for \(E_6/F_4\) it does not (Proposition IX(a)).
10. What AGKRRV prove, and why it does not apply
The theorem [AGKRRV3, Main Theorem 0.2.6 = Corollary 4.1.4; Theorem 0.6.8 = Theorem 5.2.3; Corollary 0.6.9]
Let \(\Sigma\) be a smooth projective curve over \(\mathbb F_q\) (AGKRRV write \(X\)) and \(G\) a reductive group as in [AGKRRV3, §0.2.1], with \(\mathrm{Bun}_G\) the stack of \(G\)-bundles on \(\Sigma\) and \(\mathrm{Shv}_{\mathrm{Nilp}}\) the ind-constructible sheaves with nilpotent singular support. Then
$$\operatorname{Tr}\big((\mathrm{Frob}_{\mathrm{Bun}_G})_*,\ \mathrm{Shv}_{\mathrm{Nilp}}(\mathrm{Bun}_G)\big)\simeq\mathrm{Funct}_c(\mathrm{Bun}_G(\mathbb F_q)).$$Assume [AGKRRV1, Conjecture 14.1.8], that compact objects of \(\mathrm{Shv}_{\mathrm{Nilp}}\) are constructible. Then this isomorphism is the composite of the trace of the inclusion into \(\mathrm{Shv}(\mathrm{Bun}_G)\) with the local-term map. The statement is the Trace Conjecture [AGKRRV1, Conjecture 22.3.7].
- The trace of one category. The theorem identifies the trace of a single category with a space of functions. Its local analogue here is Proposition 5.2(c), \(\operatorname{Tr}(\mathrm{Fr}_*,\mathcal C_{K,k})\simeq C_K\), which identifies the target. Given \(\mathcal C^{\mathrm{Sat}}_{K,k}=\mathcal C_{K,k}\), it yields \(\operatorname{Tr}(\mathrm{Fr}_*,\mathcal C^{\mathrm{Sat}}_{K,k})\simeq C_K\). This is true, but it is a tautology, because the two categories are equal.
- The map. (SG) concerns the map \(\operatorname{Tr}(\mathrm{act})\) into that target. GKRV's Proposition 3.4.6 identifies it with the action on \(\operatorname{cl}(\mathrm{IC}_0)\) (Corollary 4.5). Nothing in either theorem makes this module cyclic.
- The singular-support condition. AGKRRV stress the same distinction. For a general stack \(Y\), it is "emphatically not true" that \(\operatorname{Tr}((\mathrm{Frob}_Y)_*,\mathrm{Shv}(Y))\simeq\mathrm{Funct}_c(Y(\mathbb F_q))\) [AGKRRV3, §0.4.2]. Without the singular-support condition there would "obviously" be enough sheaves to recover all functions, but "the relations imposed by sheaves would not match the relations on functions" [AGKRRV3, Remark 0.6.10]. Generation gives enough sheaves. (SG) asks that the relations among classes in the image of one functor be the relations among functions.
- Classes. In general \(H_0\) of a trace is not spanned by classes [AGKRRV3, Remark 0.6.11]. This is why Theorem 6.1 assumes degree \(0\), and why §5 needs the run's Theorem 1.3.