Introduction
Basic notations
Let F F F be a non-Archimedean local field of residue characteristic p p p and residue-field cardinality q q q , and let G / F G/F G / F be a connected reductive group. Write O F \mathcal{O}_F O F for its ring of integers, fix a uniformizer ϖ \varpi ϖ , and normalize v F v_F v F and ∣ ⋅ ∣ |\mathord\cdot| ∣ ⋅ ∣ by v F ( ϖ ) = 1 v_F(\varpi)=1 v F ( ϖ ) = 1 and ∣ ϖ ∣ = q − 1 |\varpi|=q^{-1} ∣ ϖ ∣ = q − 1 . Throughout, G G G denotes the algebraic group and G ( F ) G(F) G ( F ) its locally compact group of F F F -points; we use the same convention for L L L , M M M , P P P , and other algebraic groups. We denote by M ( G ) \mathcal{M}(G) M ( G ) the category of smooth representations of G ( F ) G(F) G ( F ) (in complex vector spaces).
Invariant distributions and the Bernstein center
Fix a Haar measure on G ( F ) G(F) G ( F ) . Write H G = C ∞ ( G ( F ) , C ) \mathcal{H}_G=C^\infty(G(F),\mathbb{C}) H G = C ∞ ( G ( F ) , C ) for its Hecke algebra and D G \mathcal{D}_G D G for the space of conjugation-invariant distributions on G ( F ) G(F) G ( F ) . The subscripts in H G \mathcal{H}_G H G , D G \mathcal{D}_G D G , and Z G \mathcal{Z}_G Z G refer to the algebraic group G G G , while these spaces concern G ( F ) G(F) G ( F ) . We realize the Bernstein center Z G ⊂ D G \mathcal{Z}_G\subset\mathcal{D}_G Z G ⊂ D G as the algebra of essentially compact invariant distributions. By definition these are invariant distributions z z z such that z ∗ h ∈ H G z*h\in\mathcal{H}_G z ∗ h ∈ H G for every h ∈ H G h\in\mathcal{H}_G h ∈ H G .
Let D G c \mathcal{D}^c_G D G c consist of those invariant distributions α \alpha α for which there is a compact subset C ⊂ G ( F ) C\subset G(F) C ⊂ G ( F ) such that
Supp ( α ) ⊂ C G ( F ) , C G ( F ) = { g c g − 1 : g ∈ G ( F ) , c ∈ C } . \operatorname{Supp}(\alpha)\subset C^{G(F)},\qquad C^{G(F)}=\{gcg^{-1}:g\in G(F),\ c\in C\}. Supp ( α ) ⊂ C G ( F ) , C G ( F ) = { g c g − 1 : g ∈ G ( F ) , c ∈ C } .
Thus D G c \mathcal{D}^c_G D G c is the space of invariant distributions with bounded support modulo conjugation. The main examples of such distributions are orbital measures (a.k.a orbital integrals): for every adjoint orbit Ω ⊂ G ( F ) \Omega\subset G(F) Ω ⊂ G ( F ) it is known that
a) Ω \Omega Ω possesses a G ( F ) G(F) G ( F ) -invariant measure μ \mu μ (unique up to constant)
b) The integral ∫ Ω h d μ \int_{\Omega}h\,d\mu ∫ Ω h d μ is absolutely convergent.
Thus the above integral defines an invariant distribution which by definition lies in D G c \mathcal{D}^c_G D G c .
The rational Bernstein center
The main goal of this paper
Let R G \mathcal{R}_G R G be the rational Bernstein-center subspace:
R G = { α ∈ D G : z ∗ α ∈ Z G for some non-zero-divisor z ∈ Z G } . \mathcal{R}_G=\{\alpha\in\mathcal{D}_G:z*\alpha\in\mathcal{Z}_G\text{ for some non-zero-divisor }z\in\mathcal{Z}_G\}. R G = { α ∈ D G : z ∗ α ∈ Z G for some non-zero-divisor z ∈ Z G } .
The central action on arbitrary distributions is the transpose of its action on test functions. Explicitly, if h ∨ ( g ) = h ( g − 1 ) h^\vee(g)=h(g^{-1}) h ∨ ( g ) = h ( g − 1 ) and z ∨ ( h ) = z ( h ∨ ) z^\vee(h)=z(h^\vee) z ∨ ( h ) = z ( h ∨ ) , then ( z ∗ α ) ( h ) = α ( z ∨ ∗ h ) (z*\alpha)(h)=\alpha(z^\vee*h) ( z ∗ α ) ( h ) = α ( z ∨ ∗ h ) .
Remark. Here rationality means saturation of Z G \mathcal{Z}_G Z G inside D G \mathcal{D}_G D G under the action of non-zero-divisors. It does not assert that R G \mathcal{R}_G R G embeds in the total fraction ring of Z G \mathcal{Z}_G Z G : an invariant distribution annihilated by a non-zero-divisor already belongs to R G \mathcal{R}_G R G . Such nonzero central-torsion distributions occur for S L ( 2 ) \mathrm{SL}(2) SL ( 2 ) in Section 7.
Conjecture 1.1. For every connected reductive group G / F G/F G / F there exists a non-zero divisor z ∈ Z G z \in\mathcal{Z}_G z ∈ Z G such that z ∗ D G c ⊂ Z G z * \mathcal{D}^c_G \subset\mathcal{Z}_G z ∗ D G c ⊂ Z G .
Corollary 1.2.
D G c ⊂ R G . \mathcal{D}^c_G \subset\mathcal{R}_G. D G c ⊂ R G .
Remark 1.3. In other words, every invariant distribution supported in the conjugates of a compact set belongs to the rational Bernstein center.
The denominator for general linear groups
We prove this conjecture for G = G L ( N ) G = \mathrm{GL}(N) G = GL ( N ) , and more generally for products of general linear groups. More precisely, in this case we can describe the entire ideal of common denominators
J G = { z ∈ Z G : z ∗ D G c ⊂ Z G } . \mathcal{J}_G = \{z \in\mathcal{Z}_G : z * \mathcal{D}^c_G \subset\mathcal{Z}_G\}. J G = { z ∈ Z G : z ∗ D G c ⊂ Z G } .
(The fact that this is an ideal is obvious).
For G = G L ( N ) G = \mathrm{GL}(N) G = GL ( N ) , a Bernstein component s \mathfrak{s} s is specified by a cuspidal support
( ∏ a G L ( d a , F ) m a , ⨂ a σ a ⊗ m a ) , ∑ a d a m a = N , \left(\prod_a \mathrm{GL}(d_a,F)^{m_a},\ \bigotimes_a \sigma_a^{\otimes m_a}\right), \qquad\sum_a d_a m_a = N, ( a ∏ GL ( d a , F ) m a , a ⨂ σ a ⊗ m a ) , a ∑ d a m a = N ,
where the σ a \sigma_a σ a are pairwise inequivalent up to unramified twist. Let z a , i z_{a,i} z a , i be coordinates on their unramified-twist tori, after quotienting by the finite groups of self-twists. Then
Z s = ⨂ a C [ z a , 1 ± 1 , … , z a , m a ± 1 ] S m a , Z G = ∏ s Z s . \mathcal{Z}_{\mathfrak{s}} = \bigotimes_a \mathbb{C}[z_{a,1}^{\pm1},\ldots,z_{a,m_a}^{\pm1}]^{S_{m_a}}, \qquad\mathcal{Z}_G = \prod_{\mathfrak{s}} \mathcal{Z}_{\mathfrak{s}}. Z s = a ⨂ C [ z a , 1 ± 1 , … , z a , m a ± 1 ] S m a , Z G = s ∏ Z s .
Define the discriminant by
Δ s = ∏ a ∏ 1 ≤ i , j ≤ m a i ≠ j ( 1 − z a , i z a , j ) , Δ G = ( Δ s ) s ∈ Z G . (1) \Delta_{\mathfrak{s}} = \prod_a \prod_{\substack{1\leq i,j\leq m_a\\i\ne j}}\left(1-\frac{z_{a,i}}{z_{a,j}}\right), \qquad\Delta_G = (\Delta_{\mathfrak{s}})_{\mathfrak{s}} \in\mathcal{Z}_G. \tag*{(1)} Δ s = a ∏ 1 ≤ i , j ≤ m a i = j ∏ ( 1 − z a , j z a , i ) , Δ G = ( Δ s ) s ∈ Z G . (1)
The empty product is 1. The definition is independent of the chosen cuspidal representatives. For a product L L L of general linear groups, define Δ L \Delta_L Δ L by taking the product of the discriminants for its factors.
Theorem 1.4. For every product L L L of general linear groups,
J L = Δ L Z L . \mathcal{J}_L = \Delta_L \mathcal{Z}_L. J L = Δ L Z L . The element Δ L \Delta_L Δ L is a non-zero-divisor. In particular,
Δ L ∗ D L c ⊂ Z L , D L c ⊂ R L . \Delta_L * \mathcal{D}^c_L \subset\mathcal{Z}_L, \qquad\mathcal{D}^c_L \subset\mathcal{R}_L. Δ L ∗ D L c ⊂ Z L , D L c ⊂ R L . Thus Conjecture 1.1 holds for L L L , and every common denominator is divisible by Δ L \Delta_L Δ L component by component.
The endoscopic denominator ideal
For general G G G , our proposed description of J G \mathcal{J}_G J G uses Langlands parameters. We briefly explain the objects involved; precise algebraic constructions are given in Sections 2–4.
Let G ^ \widehat{G} G be the complex Langlands dual group of G G G , and let W F W_F W F be the Weil group of F F F . The L L L -group is L G = G ^ ⋊ W F {}^LG = \widehat{G} \rtimes W_F L G = G ⋊ W F , where W F W_F W F acts on G ^ \widehat{G} G through the action determined by G G G . A (Weil) Langlands parameter for G G G is a continuous homomorphism
φ : W F ⟶ L G \varphi: W_F \longrightarrow{}^L G φ : W F ⟶ L G
whose composition with the projection L G → W F {}^LG \to W_F L G → W F is the identity. Two parameters are identified if they are conjugate by G ^ \widehat{G} G . When G G G is split, the Weil action is trivial, so a parameter is simply a continuous homomorphism W F → G ^ W_F \to\widehat{G} W F → G , up to conjugation. We use Weil parameters here, rather than Weil–Deligne parameters; no additional monodromy operator is included.
The ring A G = Z G s p e c A_G = \mathcal{Z}_G^{\mathrm{spec}} A G = Z G spec , called the spectral Bernstein center , consists of regular functions on the stack of these parameters. Fargues–Scholze construct a homomorphism
Ψ G : A G ⟶ Z G . \Psi_G : A_G \longrightarrow\mathcal{Z}_G. Ψ G : A G ⟶ Z G .
The parameter space and this map are defined in § 2 .
An ordinary untwisted endoscopic datum e \mathfrak{e} e for G G G includes another quasi-split reductive group H H H and a group H ′ H' H ′ fitting into an exact sequence
1 ⟶ H ^ ⟶ H ′ → p e W F ⟶ 1 , 1 \longrightarrow\widehat{H} \longrightarrow H' \xrightarrow{p_{\mathfrak{e}}} W_F \longrightarrow1, 1 ⟶ H ⟶ H ′ p e W F ⟶ 1 ,
together with an embedding
η : H ′ ⟶ L G \eta: H' \longrightarrow{}^L G η : H ′ ⟶ L G
compatible with the projections to W F W_F W F . In particular, its restriction to H ^ \widehat{H} H identifies H ^ \widehat{H} H with a subgroup of G ^ \widehat{G} G . When an appropriate L L L -embedding L H → L G {}^L H \to{}^L G L H → L G has been chosen, we can take H ′ = L H H' = {}^L H H ′ = L H . A parameter with values in H ′ H' H ′ is a continuous homomorphism φ H : W F → H ′ \varphi_H : W_F \to H' φ H : W F → H ′ satisfying p e ∘ φ H = id W F p_{\mathfrak{e}} \circ\varphi_H = \operatorname{id}_{W_F} p e ∘ φ H = id W F . Composing φ H \varphi_H φ H with η \eta η gives a parameter for G G G . This is the transfer of parameters attached to e \mathfrak{e} e .
For such a parameter, let Z H ^ ( φ H ) Z_{\widehat{H}}(\varphi_H) Z H ( φ H ) denote the subgroup of H ^ \widehat{H} H commuting with the image of the parameter (in H ′ H' H ′ ), and similarly for Z G ^ ( η ∘ φ H ) Z_{\widehat{G}}(\eta\circ\varphi_H) Z G ( η ∘ φ H ) . The first centralizer is contained in the second. We call φ H \varphi_H φ H exceptional if
dim Z G ^ ( η ∘ φ H ) > dim Z H ^ ( φ H ) . \dim Z_{\widehat{G}}(\eta\circ\varphi_H) > \dim Z_{\widehat{H}}(\varphi_H). dim Z G ( η ∘ φ H ) > dim Z H ( φ H ) .
Thus a parameter is exceptional precisely when its transferred parameter has additional infinitesimal symmetries not present for H H H .
To turn this set into an ideal, we retain more information than just whether the inequality holds. Put V = Lie ( G ^ ) / Lie ( H ^ ) V = \operatorname{Lie}(\widehat{G})/\operatorname{Lie}(\widehat{H}) V = Lie ( G ) / Lie ( H ) . The parameter acts linearly on V V V through η \eta η and the adjoint action. On a finite-type parameter space, this action factors through a finitely generated group Γ \Gamma Γ ; choose generators γ 1 , … , γ m \gamma_1,\ldots,\gamma_m γ 1 , … , γ m . For each parameter form the linear map
T φ H : V ⟶ V ⊕ m , v ⟼ ( ( γ 1 − 1 ) v , … , ( γ m − 1 ) v ) . T_{\varphi_H} : V \longrightarrow V^{\oplus m}, \qquad v \longmapsto\bigl((\gamma_1-1)v,\ldots,(\gamma_m-1)v\bigr). T φ H : V ⟶ V ⊕ m , v ⟼ ( ( γ 1 − 1 ) v , … , ( γ m − 1 ) v ) .
The exceptional condition is equivalent to ker T φ H ≠ 0 \ker T_{\varphi_H} \ne0 ker T φ H = 0 . Writing n = dim V n = \dim V n = dim V , take the ideal generated by all n × n n \times n n × n minors of the matrix of T T T (and take the unit ideal if n = 0 n = 0 n = 0 ). Its vanishing set is exactly the exceptional locus. The minors also record the order of vanishing along that locus; consequently we do not replace this ideal by the ideal of functions merely vanishing on the underlying set. The construction does not depend on the chosen generators or bases; see § 3 .
We pass from the parameter space to its quotient by H ^ \widehat{H} H : intersect the ideal of minors with the ring of invariant regular functions. On each bounded parameter piece, denote the resulting ideal by J e , U \mathcal{J}_{\mathfrak{e},U} J e , U , and the induced transfer map by
f e , U : Y e , U ⟶ Y G , U . f_{\mathfrak{e},U} : Y_{\mathfrak{e},U} \longrightarrow Y_{G,U}. f e , U : Y e , U ⟶ Y G , U .
The condition imposed on a regular function a a a on the target is simply that its pullback belongs to J e , U \mathcal{J}_{\mathfrak{e},U} J e , U . In symbols, the corresponding ideal on the target is
I G , U e = { a ∈ A G , U : f e , U ∗ a ∈ J e , U } . \mathcal{I}^{\mathfrak{e}}_{G,U} = \{a \in A_{G,U} : f_{\mathfrak{e},U}^{*}a \in\mathcal{J}_{\mathfrak{e},U}\}. I G , U e = { a ∈ A G , U : f e , U ∗ a ∈ J e , U } .
Intersect these ideals over all ordinary untwisted endoscopic data, and impose the same condition on every bounded piece. This defines I G n ⊂ A G \mathcal{I}^{\mathrm{n}}_G \subset A_G I G n ⊂ A G ; see (10) . The definition is entirely in terms of regular functions and explicit matrix minors, although the more invariant formulation used later is convenient for proofs.
Conjecture 1.5 (Endoscopic description of the denominator ideal). The endoscopic ideal I G n \mathcal{I}^{\mathrm{n}}_G I G n is intrinsic and satisfies
J G = Z G ⋅ Ψ G ( I G n ) . (2) \mathcal{J}_G = \mathcal{Z}_G \cdot\Psi_G(\mathcal{I}^{\mathrm{n}}_G). \tag*{(2)} J G = Z G ⋅ Ψ G ( I G n ) . (2) This ideal contains a non-zero-divisor of Z G \mathcal{Z}_G Z G .
This refines Conjecture 1.1 by describing the ideal of common denominators. We also prove Conjecture 1.5 for GL ( N ) \operatorname{GL}(N) GL ( N ) . Under the spectral-to-Bernstein identification, the component of I G Γ I_G^\Gamma I G Γ on s \mathfrak{s} s is exactly ( Δ s ) (\Delta_{\mathfrak{s}}) ( Δ s ) ; see Theorem 6.3. Thus the endoscopic construction recovers the denominator ideal of Theorem 1.4.
Sketch of the proof of Theorem 1.4
Roughly speaking, the idea of the proof is this. Fix a congruence subgroup K K K of G ( F ) G(F) G ( F ) . Then one can show that the restriction of any invariant distribution with support bounded modulo conjugation to the corresponding Hecke algebra is a sum of distributions induced from Bernstein center elements of Levi subgroups. Comparing type A A A Plancherel measures shows that the discriminant kills their denominators. For the reverse inclusion, an induced cuspidal Bernstein projector has a numerator relatively prime to the discriminant.
Outline of the paper
We first give the precise general construction of the endoscopic ideal in Sections 2–4. Section 5 proves Theorem 1.4. Section 6 computes the endoscopic intersection. Section 7 proves both conjectures for SL ( 2 ) \operatorname{SL}(2) SL ( 2 ) . In that case the denominator is t 2 − 4 t^2-4 t 2 − 4 on each principal series component with inversion symmetry, where t = z + z − 1 t=z+z^{-1} t = z + z − 1 , and is 1 on the other components.
Acknowledgements
We thank R. Bezrukavnikov and M. Finkelberg for helpful discussions. During the preparation of this paper, D.K. was partially supported by ERC grant no. 101142781. A.B. was partially supported by NSERC. This work was done while A.B. was visiting the Simons Laufer Mathematical Institute.
The spectral Bernstein center
Fix a prime ℓ ≠ p \ell\ne p ℓ = p and put Λ = Q ‾ ℓ \Lambda=\overline{\mathbb{Q}}_\ell Λ = Q ℓ . Let G ^ \widehat{G} G be the Langlands dual group over Λ \Lambda Λ , equipped with the pinned action of W F W_F W F . Fix an isomorphism ι : Λ ≃ C \iota:\Lambda\simeq\mathbb{C} ι : Λ ≃ C and a square root of q q q used in the Fargues–Scholze normalization. All comparisons with complex distributions use these choices.
The parameter scheme
Fargues–Scholze construct a scheme Z 1 ( W F , G ^ ) Λ Z^1(W_F,\widehat{G})_\Lambda Z 1 ( W F , G ) Λ representing condensed 1-cocycles; see [3 Definition VIII.1.1 and Theorem VIII.1.3] . A cocycle c c c satisfies
c ( w w ′ ) = c ( w ) w ( c ( w ′ ) ) , w , w ′ ∈ W F , c(ww')=c(w)w(c(w')), \qquad w,w'\in W_F, c ( w w ′ ) = c ( w ) w ( c ( w ′ )) , w , w ′ ∈ W F ,
together with the condensed continuity condition. Equivalently, it defines a section
φ c : W F ⟶ G ^ ⋊ W F , w ⟼ ( c ( w ) , w ) \varphi_c:W_F\longrightarrow\widehat{G}\rtimes W_F,\qquad w\longmapsto(c(w),w) φ c : W F ⟶ G ⋊ W F , w ⟼ ( c ( w ) , w )
of the projection to W F W_F W F . For split G G G , the Weil action is trivial and these are continuous homomorphisms into G ^ \widehat{G} G . The conjugation action is
( g ⋅ c ) ( w ) = g c ( w ) w ( g ) − 1 . (g\cdot c)(w)=gc(w)w(g)^{-1}. ( g ⋅ c ) ( w ) = g c ( w ) w ( g ) − 1 .
The parameter stack and spectral Bernstein center are
Loc G = [ Z 1 ( W F , G ^ ) Λ / G ^ ] , A G : = Z G s p e c = Γ ( Loc G , O ) . \operatorname{Loc}_G=[Z^1(W_F,\widehat{G})_\Lambda/\widehat{G}],\qquad A_G:=\mathcal{Z}^{\mathrm{spec}}_G=\Gamma(\operatorname{Loc}_G,\mathcal{O}). Loc G = [ Z 1 ( W F , G ) Λ / G ] , A G := Z G spec = Γ ( Loc G , O ) .
Thus
A G = O ( Z 1 ( W F , G ^ ) Λ ) G ^ . A_G=\mathcal{O}(Z^1(W_F,\widehat{G})_\Lambda)^{\widehat{G}}. A G = O ( Z 1 ( W F , G ) Λ ) G .
The full parameter space is handled using finite-type pieces with bounded wild ramification. On characteristic-zero affine quotients, closed geometric points correspond to semisimple parameter classes. In particular, for an invariant regular function a a a ,
a ( φ ) = a ( φ s s ) . a(\varphi)=a(\varphi^{\mathrm{ss}}). a ( φ ) = a ( φ ss ) .
The map to the ordinary Bernstein center
Theorem 2.1 (Fargues–Scholze). With the coefficient and normalization choices above, there is a ring homomorphism
Ψ G : A G ⟶ Z G . \Psi_G : \mathcal{A}_G \longrightarrow\mathcal{Z}_G. Ψ G : A G ⟶ Z G . The construction is given in [3 Proposition I.9.3, Theorem IX.5.2, and Definition IX.7.1] .
Stable and very stable centers
We spell out the harmonic-analytic definitions used in [4 Section 1.1] and [5 Definition 1.1.1] . Two strongly regular semisimple elements of G ( F ) G(F) G ( F ) are stably conjugate if they are conjugate under G ( F ‾ ) G(\overline{F}) G ( F ) . For such an element γ ∈ G ( F ) \gamma\in G(F) γ ∈ G ( F ) , let G γ G_\gamma G γ denote its algebraic centralizer, which is a torus. For f ∈ H G f \in\mathcal{H}_G f ∈ H G , put
O γ ( f ) = ∫ G γ ( F ) \ G ( F ) f ( x − 1 γ x ) d x . O_\gamma(f) = \int_{G_\gamma(F)\backslash G(F)} f(x^{-1}\gamma x)\,dx. O γ ( f ) = ∫ G γ ( F ) \ G ( F ) f ( x − 1 γ x ) d x .
Choose Haar measures on the centralizer tori compatibly under stable conjugacy. The stable orbital integral is
S O γ ( f ) = ∑ [ γ ′ ] ⊂ [ γ ] s t O γ ′ ( f ) , SO_\gamma(f) = \sum_{[\gamma']\subset[\gamma]_{\mathrm{st}}} O_{\gamma'}(f), S O γ ( f ) = [ γ ′ ] ⊂ [ γ ] st ∑ O γ ′ ( f ) ,
where the finite sum runs over the G ( F ) G(F) G ( F ) -conjugacy classes in the stable conjugacy class of γ \gamma γ that occur in G ( F ) G(F) G ( F ) . Changing the compatible measures multiplies this functional by a nonzero scalar and does not change its kernel.
Let
H G u n s t = { f ∈ H G : S O γ ( f ) = 0 for every strongly regular semisimple γ ∈ G ( F ) } . \mathcal{H}^{\mathrm{unst}}_G = \{f \in\mathcal{H}_G : SO_\gamma(f) = 0 \text{ for every strongly regular semisimple } \gamma\in G(F)\}. H G unst = { f ∈ H G : S O γ ( f ) = 0 for every strongly regular semisimple γ ∈ G ( F )} .
An invariant distribution D D D on G ( F ) G(F) G ( F ) is stable if D ( f ) = 0 D(f) = 0 D ( f ) = 0 for every f ∈ H G u n s t f \in\mathcal{H}^{\mathrm{unst}}_G f ∈ H G unst . In particular, the stable Bernstein center is the vector space
Z G s t = { z ∈ Z G : z ( f ) = 0 for all f ∈ H G u n s t } . \mathcal{Z}^{\mathrm{st}}_G = \{z \in\mathcal{Z}_G : z(f) = 0 \text{ for all } f \in\mathcal{H}^{\mathrm{unst}}_G\}. Z G st = { z ∈ Z G : z ( f ) = 0 for all f ∈ H G unst } .
Here stability is a condition on z z z viewed as a distribution; closure of this space under convolution is not part of the definition.
The very stable Bernstein center is
Z G v s t = { z ∈ Z G : z ∗ H G u n s t ⊂ H G u n s t } . \mathcal{Z}^{\mathrm{vst}}_G = \{z \in\mathcal{Z}_G : z * \mathcal{H}^{\mathrm{unst}}_G \subset\mathcal{H}^{\mathrm{unst}}_G\}. Z G vst = { z ∈ Z G : z ∗ H G unst ⊂ H G unst } .
Thus z z z is very stable precisely when convolution by z z z descends to an operator on the stable quotient H G / H G u n s t \mathcal{H}_G/\mathcal{H}^{\mathrm{unst}}_G H G / H G unst . Unlike the definition of stability, this condition concerns the action on all unstable test functions. It immediately makes Z G v s t \mathcal{Z}^{\mathrm{vst}}_G Z G vst a subalgebra of Z G \mathcal{Z}_G Z G .
For p p p -adic F F F , one has Z G v s t ⊂ Z G s t \mathcal{Z}^{\mathrm{vst}}_G \subset\mathcal{Z}^{\mathrm{st}}_G Z G vst ⊂ Z G st ; see [5 Remark 1.1.2] . One way to see the implication uses the stability of the distribution f ↦ f ( 1 ) f \mapsto f(1) f ↦ f ( 1 ) , together with z ( f ) = ( z ∗ f ∨ ) ( 1 ) z(f) = (z * f^\vee)(1) z ( f ) = ( z ∗ f ∨ ) ( 1 ) and the preservation of H G u n s t \mathcal{H}^{\mathrm{unst}}_G H G unst by inversion. Moreover, Z G s t \mathcal{Z}^{\mathrm{st}}_G Z G st is a module over Z G v s t \mathcal{Z}^{\mathrm{vst}}_G Z G vst under convolution.
Remark 2.2. For p p p -adic F F F , the equality
Z G s t = Z G v s t \mathcal{Z}^{\mathrm{st}}_G = \mathcal{Z}^{\mathrm{vst}}_G Z G st = Z G vst
is conjectured in general; see [5 Conjecture 1.1.4] and [4 Section 1.1] . It is stronger than merely asserting that stable central distributions form a subalgebra. Varma proves the equality under the hypotheses on tempered L L L -packets specified in [5 Hypothesis 2.5.1 and Theorem 4.4.2] . Hansen’s theorem below gives an unconditional inclusion for the image of the spectral center; it does not assert equality of the stable and very stable centers for arbitrary G G G .
Theorem 2.3 (Hansen). If F F F is a finite extension of Q p \mathbb{Q}_p Q p , then
Im ( Ψ G ) ⊂ Z G v s t ⊂ Z G s t . \operatorname{Im}(\Psi_G) \subset\mathcal{Z}^{\mathrm{vst}}_G \subset\mathcal{Z}^{\mathrm{st}}_G. Im ( Ψ G ) ⊂ Z G vst ⊂ Z G st . The assertion is [4 Theorem 1.1] .
Remark 2.4. The cited theorem of Hansen assumes characteristic zero. No stability assertion for fields of positive characteristic is deduced from that reference here. The definitions and conjectures in this article are nevertheless formulated for arbitrary non-Archimedean local fields.
A general algebraic construction
We first give a construction for ordinary representation varieties. Its adaptation to endoscopic parameter spaces is described in § 4 .
Setup
Let k k k be a field of characteristic zero, let G ^ \widehat{G} G be a reductive algebraic group over k k k , and let i : H ^ ↪ G ^ i:\widehat{H}\hookrightarrow\widehat{G} i : H ↪ G be a closed reductive subgroup. Let Γ \Gamma Γ be a finitely generated discrete group. Write
A ( K ) = Hom ( Γ , K ) , B ( K ) = A ( K ) / / K = Spec k [ A ( K ) ] K , K = H ^ , G ^ . A(K)=\operatorname{Hom}(\Gamma,K),\qquad B(K)=A(K)//K=\operatorname{Spec} k[A(K)]^{K},\qquad K=\widehat{H},\widehat{G}. A ( K ) = Hom ( Γ , K ) , B ( K ) = A ( K ) // K = Spec k [ A ( K ) ] K , K = H , G .
We use reduced representation varieties; the same arguments also apply to the affine representation schemes. The induced morphism
f : B ( H ^ ) ⟶ B ( G ^ ) f:B(\widehat{H})\longrightarrow B(\widehat{G}) f : B ( H ) ⟶ B ( G )
is finite by the Vinberg finiteness theorem (see [6 Theorem 1.1] ). In particular, its image C C C is closed.
For a geometric point c ′ ∈ B ( H ^ ) c'\in B(\widehat{H}) c ′ ∈ B ( H ) , choose a representation ρ : Γ → H ^ \rho:\Gamma\to\widehat{H} ρ : Γ → H in the unique closed H ^ \widehat{H} H -orbit over c ′ c' c ′ , and put
d ( c ′ ) = dim Z G ^ ( i ∘ ρ ) − dim Z H ^ ( ρ ) . d(c')=\dim Z_{\widehat{G}}(i\circ\rho)-\dim Z_{\widehat{H}}(\rho). d ( c ′ ) = dim Z G ( i ∘ ρ ) − dim Z H ( ρ ) .
Here Z K ( ρ ) Z_K(\rho) Z K ( ρ ) denotes the centralizer of the image of the representation, equivalently its stabilizer under conjugation. This integer is independent of the choice within that closed orbit. For r ≥ 1 r\geq1 r ≥ 1 , define
Z r = { c ∈ C : max c ′ ∈ f − 1 ( c ) d ( c ′ ) ≥ r } , Z = Z 1 . Z_r=\left\{c\in C:\max_{c'\in f^{-1}(c)}d(c')\geq r\right\},\qquad Z=Z_1. Z r = { c ∈ C : c ′ ∈ f − 1 ( c ) max d ( c ′ ) ≥ r } , Z = Z 1 .
Fibers and maxima are interpreted over an algebraic closure of the residue field. The fibers are finite and nonempty for c ∈ C c\in C c ∈ C .
Closedness of the exceptional locus
Proposition 3.1. For every integer r ≥ 1 r\geq1 r ≥ 1 , the subset Z r Z_r Z r is Zariski closed in B ( G ^ ) B(\widehat{G}) B ( G ) .
Proof. We first work over an algebraically closed field. Set g = Lie ( G ^ ) \mathfrak{g}=\operatorname{Lie}(\widehat{G}) g = Lie ( G ) and h = Lie ( H ^ ) \mathfrak{h}=\operatorname{Lie}(\widehat{H}) h = Lie ( H ) . Since H ^ \widehat{H} H is linearly reductive in characteristic zero, the exact sequence of H ^ \widehat{H} H -modules
0 ⟶ h ⟶ g ⟶ V ⟶ 0 , V = g / h , 0\longrightarrow\mathfrak{h}\longrightarrow\mathfrak{g}\longrightarrow V\longrightarrow0,\qquad V=\mathfrak{g}/\mathfrak{h}, 0 ⟶ h ⟶ g ⟶ V ⟶ 0 , V = g / h ,
splits H ^ \widehat{H} H -equivariantly. Centralizers are smooth in characteristic zero, and their Lie algebras are the corresponding fixed subspaces. Hence, for every ρ ∈ A ( H ^ ) \rho\in A(\widehat{H}) ρ ∈ A ( H ) , including representations with nonclosed orbits,
dim Z G ^ ( i ∘ ρ ) − dim Z H ^ ( ρ ) = dim g ρ ( Γ ) − dim h ρ ( Γ ) = dim V ρ ( Γ ) . (3) \dim Z_{\widehat{G}}(i\circ\rho)-\dim Z_{\widehat{H}}(\rho)=\dim\mathfrak{g}^{\rho(\Gamma)}-\dim\mathfrak{h}^{\rho(\Gamma)}=\dim V^{\rho(\Gamma)}. \tag*{(3)} dim Z G ( i ∘ ρ ) − dim Z H ( ρ ) = dim g ρ ( Γ ) − dim h ρ ( Γ ) = dim V ρ ( Γ ) . (3)
Choose generators γ 1 , … , γ m \gamma_1,\ldots,\gamma_m γ 1 , … , γ m of Γ \Gamma Γ . The fixed subspace in (3) is the kernel of
T ρ : V ⟶ V ⊕ m , v ⟼ ( ( ρ ( γ j ) − 1 ) v ) j = 1 m , T_\rho:V\longrightarrow V^{\oplus m},\qquad v\longmapsto\bigl((\rho(\gamma_j)-1)v\bigr)_{j=1}^{m}, T ρ : V ⟶ V ⊕ m , v ⟼ ( ( ρ ( γ j ) − 1 ) v ) j = 1 m ,
where H ^ \widehat{H} H acts on V V V through the adjoint action. Its matrix entries depend regularly on ρ \rho ρ . Thus
D r = { ρ ∈ A ( H ^ ) : dim ker T ρ ≥ r } D_r=\{\rho\in A(\widehat{H}):\dim\ker T_\rho\geq r\} D r = { ρ ∈ A ( H ) : dim ker T ρ ≥ r }
is a closed, H ^ \widehat{H} H -invariant determinantal subset. Let q H ^ : A ( H ^ ) → B ( H ^ ) q_{\widehat{H}}: A(\widehat{H}) \to B(\widehat{H}) q H : A ( H ) → B ( H ) be the quotient map. An affine reductive quotient sends invariant closed subsets to closed subsets; see [7 Lemma 2.3] . Moreover,
q H ^ ( D r ) = { c ′ ∈ B ( H ^ ) : d ( c ′ ) ≥ r } . (4) q_{\widehat{H}}(D_r)=\{c' \in B(\widehat{H}) : d(c') \ge r\}. \tag*{(4)} q H ( D r ) = { c ′ ∈ B ( H ) : d ( c ′ ) ≥ r } . (4)
Indeed, if a quotient fiber meets D r D_r D r , then the closure of the orbit of such a point is contained in D r D_r D r . That orbit closure contains the unique closed orbit of the fiber. Conversely, a closed-orbit representative with d ( c ′ ) ≥ r d(c') \ge r d ( c ′ ) ≥ r belongs to D r D_r D r by (3). Since f f f is finite and hence closed, (4) gives
Z r = f ( q H ^ ( D r ) ) , Z_r=f(q_{\widehat{H}}(D_r)), Z r = f ( q H ( D r )) ,
which is closed in B ( G ^ ) B(\widehat{G}) B ( G ) . All constructions are defined over k k k ; invariants commute with field extension in characteristic zero. Thus the argument gives the stated closed locus over the original field as well.
A canonical coherent sheaf
The following construction provides a natural scheme structure obtained from a fixed-vector rank condition.
Put
R = k [ A ( H ^ ) ] , S = R H ^ , V = g / h . R=k[A(\widehat{H})], \qquad S=R^{\widehat{H}}, \qquad V=\mathfrak{g}/\mathfrak{h}. R = k [ A ( H )] , S = R H , V = g / h .
The universal representation ρ u n i v : Γ → H ^ ( R ) \rho_{\mathrm{univ}}:\Gamma\to\widehat{H}(R) ρ univ : Γ → H ( R ) acts on N = R ⊗ k V ∨ N=R\otimes_k V^\vee N = R ⊗ k V ∨ . Form its module of coinvariants
M = N Γ = N / ⟨ γ u − u : γ ∈ Γ , u ∈ N ⟩ R . (5) M=N_{\Gamma}=N/\langle\gamma u-u:\gamma\in\Gamma,\ u\in N\rangle_R. \tag*{(5)} M = N Γ = N / ⟨ γ u − u : γ ∈ Γ , u ∈ N ⟩ R . (5)
This definition is canonical: neither generators of Γ \Gamma Γ nor a complement to h \mathfrak{h} h in g \mathfrak{g} g are part of the construction. The module is finitely presented. Indeed, for any finite generating set γ 1 , … , γ m \gamma_1,\ldots,\gamma_m γ 1 , … , γ m , it is the cokernel of
N ⊕ m ⟶ N , ( u 1 , … , u m ) ⟼ ∑ j = 1 m ( γ j − 1 ) u j . N^{\oplus m}\longrightarrow N,\qquad(u_1,\ldots,u_m)\longmapsto\sum_{j=1}^{m}(\gamma_j-1)u_j. N ⊕ m ⟶ N , ( u 1 , … , u m ) ⟼ j = 1 ∑ m ( γ j − 1 ) u j .
At a geometric point ρ \rho ρ , coinvariants commute with specialization and finite-dimensional duality gives
M ⊗ R k ( ρ ) ≃ ( V ∨ ⊗ k k ( ρ ) ) ρ ( Γ ) ≃ ( ( V ⊗ k k ( ρ ) ) ρ ( Γ ) ) ∨ . (6) M\otimes_R k(\rho)\simeq(V^\vee\otimes_k k(\rho))_{\rho(\Gamma)}\simeq((V\otimes_k k(\rho))^{\rho(\Gamma)})^\vee. \tag*{(6)} M ⊗ R k ( ρ ) ≃ ( V ∨ ⊗ k k ( ρ ) ) ρ ( Γ ) ≃ (( V ⊗ k k ( ρ ) ) ρ ( Γ ) ) ∨ . (6)
Choose generators γ 1 , … , γ m \gamma_1,\ldots,\gamma_m γ 1 , … , γ m of Γ \Gamma Γ and a basis of V V V . The relations in (5) are represented by an n × m n n\times mn n × mn matrix, where n = dim V n=\dim V n = dim V ; its blocks are the matrices of γ j − 1 \gamma_j-1 γ j − 1 acting on V ∨ V^\vee V ∨ . Let I I I be the ideal generated by all its n × n n\times n n × n minors (with I = R I=R I = R if n = 0 n=0 n = 0 ). Although this description uses choices, the ideal is independent of them: it is the standard zeroth determinantal (or Fitting) ideal of the module M M M . Define
I = Fitt 0 R ( M ) , J = I ∩ S , E = Spec ( S / J ) ⊂ B ( H ^ ) . I=\operatorname{Fitt}^{R}_{0}(M),\qquad J=I\cap S,\qquad E=\operatorname{Spec}(S/J)\subset B(\widehat{H}). I = Fitt 0 R ( M ) , J = I ∩ S , E = Spec ( S / J ) ⊂ B ( H ) .
The universal construction is equivariant for conjugation by H ^ \widehat{H} H , so I I I is H ^ \widehat{H} H -stable. The zero set of I I I is Supp R ( M ) \operatorname{Supp}_R(M) Supp R ( M ) , which by (6) is precisely D 1 D_1 D 1 . Exactness of invariants in characteristic zero gives
S / J ≃ ( R / I ) H ^ . S/J\simeq(R/I)^{\widehat{H}}. S / J ≃ ( R / I ) H .
Consequently E E E is the affine quotient of the canonical determinantal subscheme Spec ( R / I ) \operatorname{Spec}(R/I) Spec ( R / I ) , and its underlying set is
∣ E ∣ = q H ^ ( D 1 ) = { c ′ ∈ B ( H ^ ) : d ( c ′ ) > 0 } . |E|=q_{\widehat{H}}(D_1)=\{c'\in B(\widehat{H}):d(c')>0\}. ∣ E ∣ = q H ( D 1 ) = { c ′ ∈ B ( H ) : d ( c ′ ) > 0 } .
Proposition 3.2. Regard O E \mathcal{O}_E O E as a coherent sheaf on B ( H ^ ) B(\widehat{H}) B ( H ) via its closed inclusion. Then
F = f ∗ O E \mathcal{F}=f_*\mathcal{O}_E F = f ∗ O E is a canonically defined coherent sheaf on B ( G ^ ) B(\widehat{G}) B ( G ) , and Supp ( F ) = Z \operatorname{Supp}(\mathcal{F})=Z Supp ( F ) = Z .
Proof. The sheaf O E \mathcal{O}_E O E is coherent, and pushforward along the finite morphism f f f preserves coherence. For a finite morphism, the support of the pushforward of a coherent sheaf is the image of its support. Thus
Supp ( F ) = f ( ∣ E ∣ ) = Z . \operatorname{Supp}(\mathcal{F})=f(|E|)=Z. Supp ( F ) = f ( ∣ E ∣ ) = Z .
More explicitly, if T = k [ B ( G ^ ) ] T=k[B(\widehat{G})] T = k [ B ( G )] and f ∗ : T → S f^*:T\to S f ∗ : T → S is the restriction map, then F \mathcal{F} F corresponds to the finite T T T -module S / J S/J S / J . Its annihilator is
Ann T ( S / J ) = ( f ∗ ) − 1 ( J ) . \operatorname{Ann}_{T}(S/J)=(f^*)^{-1}(J). Ann T ( S / J ) = ( f ∗ ) − 1 ( J ) .
Hence this construction also defines a canonical closed subscheme of B ( G ^ ) B(\widehat{G}) B ( G ) with underlying set Z Z Z .
Remark 3.3. The complement used in the proof of (3) is not canonical, but the module V = g / h V=\mathfrak{g}/\mathfrak{h} V = g / h , the coinvariant module M M M , its Fitting ideal, and the sheaf F \mathcal{F} F are canonical. Replacing E E E by E r e d E_{\mathrm{red}} E red gives another canonical coherent sheaf with the same support.
Endoscopic parameter spaces and the candidate ideal
A discrete Weil-group replacement
Write P F ⊂ W F P_F\subset W_F P F ⊂ W F for wild inertia. Let U ⊂ P F U\subset P_F U ⊂ P F be open in P F P_F P F and normal in W F W_F W F . The quotient W F / U W_F/U W F / U still contains profinite tame inertia, so it is not a finitely generated discrete group. Choose an arithmetic Frobenius Fr \operatorname{Fr} Fr and a pro-generator s t s_t s t of tame inertia I F / P F I_F/P_F I F / P F . Let W F 0 W_F^0 W F 0 be the inverse image in W F W_F W F of
⟨ Fr , s t ⟩ ≃ Z [ 1 / q ] ⋊ Z ⊂ W F / P F , Fr s t Fr − 1 = s t q . \langle\operatorname{Fr},s_t\rangle\simeq\mathbb{Z}[1/q]\rtimes\mathbb{Z}\subset W_F/P_F,\qquad\operatorname{Fr}s_t\operatorname{Fr}^{-1}=s_t^q. ⟨ Fr , s t ⟩ ≃ Z [ 1/ q ] ⋊ Z ⊂ W F / P F , Fr s t Fr − 1 = s t q .
Endow W F 0 W_F^0 W F 0 with the topology for which P F P_F P F is open with its profinite topology and W F 0 / P F W_F^0/P_F W F 0 / P F is discrete. Then
Γ U = W F 0 / U \Gamma_U=W_F^0/U Γ U = W F 0 / U
is a finitely generated discrete group: it is an extension of the two-generator group ⟨ Fr , s t ⟩ \langle\operatorname{Fr},s_t\rangle ⟨ Fr , s t ⟩ by the finite group P F / U P_F/U P F / U . This is the discretized Weil group used in [8 Introduction and Sections 2–4] . For sufficiently small U U U , the Weil action on the dual group factors through Γ U \Gamma_U Γ U , and the bounded parameter scheme is described by
X G , U = Z 1 ( Γ U , G ^ ) , Y G , U = X G , U / / G ^ , A G , U = O ( Y G , U ) . X_{G,U}=Z^1(\Gamma_U,\widehat{G}),\qquad Y_{G,U}=X_{G,U}//\widehat{G},\qquad A_{G,U}=\mathcal{O}(Y_{G,U}). X G , U = Z 1 ( Γ U , G ) , Y G , U = X G , U // G , A G , U = O ( Y G , U ) .
These are cocycle schemes, not ordinary homomorphism schemes unless the action is trivial. The characteristic-zero comparison with the condensed parameter construction is used to regard these quotients as bounded pieces of the spectral parameter space; see [3 Section VIII.1] and [8 ] .
Endoscopic data
Fix an endoscopic datum ϵ = ( H , H ′ , s e , η ) \epsilon=(H,H',s_e,\eta) ϵ = ( H , H ′ , s e , η ) , where H H H is quasi-split and H ′ H' H ′ fits into an exact sequence
1 ⟶ H ^ ⟶ H ′ → p e W F ⟶ 1. 1\longrightarrow\widehat{H}\longrightarrow H'\xrightarrow{p_e}W_F\longrightarrow1. 1 ⟶ H ⟶ H ′ p e W F ⟶ 1.
The embedding
η : H ′ ⟶ L G , η ( H ^ ) = Z G ^ ( s e ) ∘ , \eta:H'\longrightarrow{}^LG,\qquad\eta(\widehat{H})=Z_{\widehat{G}}(s_e)^\circ, η : H ′ ⟶ L G , η ( H ) = Z G ( s e ) ∘ ,
is compatible with the projections to W F W_F W F . Here s e ∈ G ^ s_e\in\widehat{G} s e ∈ G is semisimple, and elements of η ( H ′ ) \eta(H') η ( H ′ ) conjugate s e s_e s e to itself modulo Z ( G ^ ) Z(\widehat{G}) Z ( G ) . When an L L L -embedding L H → L G {}^LH\to{}^LG L H → L G has been chosen, one can use H ′ = L H H'={}^LH H ′ = L H . In general the extension H ′ H' H ′ must be retained; a bare inclusion H ^ ⊂ G ^ \widehat{H}\subset\widehat{G} H ⊂ G does not specify parameter transfer. Let φ H : W F → H ′ \varphi_H:W_F\to H' φ H : W F → H ′ be a continuous homomorphic section of p e p_e p e . Define
d e ( φ H ) = dim Z G ^ ( η ∘ φ H ) − dim Z H ^ ( φ H ) . d_e(\varphi_H)=\dim Z_{\widehat{G}}(\eta\circ\varphi_H)-\dim Z_{\widehat{H}}(\varphi_H). d e ( φ H ) = dim Z G ( η ∘ φ H ) − dim Z H ( φ H ) .
The centralizers use the full L L L -parameters, including their Weil action. Put g = Lie ( G ^ ) \mathfrak{g}=\operatorname{Lie}(\widehat{G}) g = Lie ( G ) , h = Lie ( H ^ ) \mathfrak{h}=\operatorname{Lie}(\widehat{H}) h = Lie ( H ) , and V e = g / h V_e=\mathfrak{g}/\mathfrak{h} V e = g / h . Since
h = g Ad ( s e ) , \mathfrak{h}=\mathfrak{g}^{\operatorname{Ad}(s_e)}, h = g Ad ( s e ) ,
the eigenspace decomposition of Ad ( s e ) \operatorname{Ad}(s_e) Ad ( s e ) gives a complement to h \mathfrak{h} h preserved by η ( H ′ ) \eta(H') η ( H ′ ) . Indeed, conjugating s e s_e s e by such an element changes it only by a central factor, which acts trivially on g \mathfrak{g} g . Consequently the splitting is compatible with every φ H \varphi_H φ H , and
d e ( φ H ) = dim V e Ad ( η ∘ φ H ) ( W F ) . (7) d_e(\varphi_H)=\dim V_e^{\operatorname{Ad}(\eta\circ\varphi_H)(W_F)}. \tag*{(7)} d e ( φ H ) = dim V e Ad ( η ∘ φ H ) ( W F ) . (7)
This supplies the analogue of (3) in the endoscopic setting.
The ideal on a bounded parameter piece
Choose compatible wild-ramification bounds for the source and target, so that parameter transfer gives an algebraic morphism
f e , U : Y e , U ⟶ Y G , U . f_{e,U}:Y_{e,U}\longrightarrow Y_{G,U}. f e , U : Y e , U ⟶ Y G , U .
Here X e , U X_{e,U} X e , U is the affine scheme of sections into the endoscopic extension with the chosen bounds, and Y e , U = X e , U / / H ^ Y_{e,U}=X_{e,U}//\widehat{H} Y e , U = X e , U // H . The choices must ensure that the universal adjoint action on V e V_e V e factors through a finitely generated discrete Weil-group quotient. Write this quotient as Γ \Gamma Γ .
Remark 4.1. The coherent-pushforward construction below requires f e , U f_{e,U} f e , U to be finite. Section 3 proves this in the ordinary representation-variety situation by Cotner’s theorem. Its use for an endoscopic cocycle model requires the corresponding finiteness result for that model. This is an explicit requirement of the endoscopic construction, rather than a consequence asserted solely from H ^ ⊂ G ^ \widehat{H}\subset\widehat{G} H ⊂ G .
Assume this finiteness condition, and put
R e , U = O ( X e , U ) , S e , U = R e , U H ^ . R_{e,U}=\mathcal{O}(X_{e,U}),\qquad S_{e,U}=R_{e,U}^{\widehat{H}}. R e , U = O ( X e , U ) , S e , U = R e , U H .
Using the universal adjoint action, define
M e , U = ( R e , U ⊗ ⋀ V e ∨ ) Γ , K e , U = Fitt 0 R e , U ( M e , U ) , J e , U = K e , U ∩ S e , U . M_{e,U}=(R_{e,U}\otimes\bigwedge V_e^\vee)^\Gamma,\qquad
K_{e,U}=\operatorname{Fitt}_{0}^{R_{e,U}}(M_{e,U}),\qquad
J_{e,U}=K_{e,U}\cap S_{e,U}. M e , U = ( R e , U ⊗ ⋀ V e ∨ ) Γ , K e , U = Fitt 0 R e , U ( M e , U ) , J e , U = K e , U ∩ S e , U .
Set
E e , U = Spec ( S e , U / J e , U ) , F e , U = ( f e , U ) ∗ O E e , U . E_{e,U}=\operatorname{Spec}(S_{e,U}/J_{e,U}),\qquad
F_{e,U}=(f_{e,U})_*\mathcal{O}_{E_{e,U}}. E e , U = Spec ( S e , U / J e , U ) , F e , U = ( f e , U ) ∗ O E e , U .
The splitting in (7) and the same determinantal and quotient arguments as in Section 3 show that E e , U E_{e,U} E e , U has underlying set
{ [ φ H ] : d e ( φ H ) > 0 } , \{[\varphi_H]:d_e(\varphi_H)>0\}, {[ φ H ] : d e ( φ H ) > 0 } ,
where closed-orbit representatives are used. Define the spectral ideal
I G , U e = Ann A G , U ( F e , U ) = ( f e , U ) − 1 ( J e , U ) ⊂ A G , U . (8) I^e_{G,U}=\operatorname{Ann}_{A_{G,U}}(F_{e,U})=(f_{e,U})^{-1}(J_{e,U})\subset A_{G,U}. \tag*{(8)} I G , U e = Ann A G , U ( F e , U ) = ( f e , U ) − 1 ( J e , U ) ⊂ A G , U . (8)
Its zero set is the image of the exceptional locus. Formula (8) specifies the scheme structure, not just its underlying set.
Remark 4.2. Replacing this annihilator by its radical, or by the zeroth Fitting ideal of the pushforward, can change the ideal and hence the conjecture. We use exactly the annihilator in (8).
The intersection ideal
Use ordinary untwisted endoscopic data, with their realized parameter-transfer maps, taken up to isomorphism and conjugacy over the target. The usual untwisted endoscopic condition is part of this convention; transfers with a nontrivial twisting character are not included. For a collection E U \mathcal{E}_U E U on a bounded piece, put
I G , U ∩ = ⋂ e ∈ E U I G , U e . (9) I^\cap_{G,U}=\bigcap_{e\in\mathcal{E}_U}I^e_{G,U}. \tag*{(9)} I G , U ∩ = e ∈ E U ⋂ I G , U e . (9)
The empty intersection is the unit ideal. For a finite collection, this can also be written
I G , U ∩ = Ann A G , U ( ⨁ e ∈ E U F e , U ) . I^\cap_{G,U}=\operatorname{Ann}_{A_{G,U}}\left(\bigoplus_{e\in\mathcal{E}_U}F_{e,U}\right). I G , U ∩ = Ann A G , U ( e ∈ E U ⨁ F e , U ) .
In particular, repeating an equivalent transfer does not change the ideal. Distinct transfers with the same annihilator also impose just one ideal-theoretic condition.
Here is a formulation that does not assume that an arbitrary intersection commutes with localization. On every bounded target piece consider all compatible bounded source pieces and all endoscopic transfers as above. Define
I G Γ = { a ∈ A G : f e , U ∗ ( a ∣ Y G , U ) ∈ J e , U for every such ( e , U ) } . (10) \mathcal{I}^{\Gamma}_{G}=\left\{a\in A_G:f^{*}_{e,U}(a|_{Y_{G,U}})\in J_{e,U}\text{ for every such }(e,U)\right\}. \tag*{(10)} I G Γ = { a ∈ A G : f e , U ∗ ( a ∣ Y G , U ) ∈ J e , U for every such ( e , U ) } . (10)
This is an ideal, defined by simultaneous pullback conditions. Equivalently, its restrictions satisfy all the annihilator conditions in (8). The pullback definition makes sense even before finiteness of the transfers is established; when they are finite it has the coherent-sheaf interpretation given above. For a finite collection on a finite-type piece, the local construction is exactly (9). Compatibility and independence of the bounded parameter models are included in the geometric Conjecture 1.5. They hold in the general-linear computation in Section 6. This completes the definition of the ideal appearing in Conjecture 1.5. Its proposed image in the Bernstein center is the common-denominator ideal J G \mathcal{J}_G J G .
Proof of the theorem for general linear groups
Some notation
In this section G = GL ( N ) G=\operatorname{GL}(N) G = GL ( N ) and L = ∏ j = 1 s GL ( n j ) L=\prod_{j=1}^{s}\operatorname{GL}(n_j) L = ∏ j = 1 s GL ( n j ) . We fix Haar measures on L ( F ) L(F) L ( F ) and on M ( F ) M(F) M ( F ) for every Levi subgroup M M M of L L L . Write
H L = C c ∞ ( L ( F ) ) , H ‾ L = H L / [ H L , H L ] , \mathcal{H}_L=C^\infty_c(L(F)),\qquad\overline{\mathcal{H}}_L=\mathcal{H}_L/[\mathcal{H}_L,\mathcal{H}_L], H L = C c ∞ ( L ( F )) , H L = H L / [ H L , H L ] ,
and let Z L \mathcal{Z}_L Z L be the Bernstein center. For a compact open subgroup K ⊂ L ( F ) K\subset L(F) K ⊂ L ( F ) , put
e K = vol ( K ) − 1 1 K , H ( L ( F ) , K ) = e K ∗ H L ∗ e K . e_K=\operatorname{vol}(K)^{-1}1_K,\qquad\mathcal{H}(L(F),K)=e_K*\mathcal{H}_L*e_K. e K = vol ( K ) − 1 1 K , H ( L ( F ) , K ) = e K ∗ H L ∗ e K .
An invariant distribution is a functional on H ‾ L \overline{\mathcal{H}}_L H L . We regard every b ∈ Z L b\in\mathcal{Z}_L b ∈ Z L as an invariant distribution and write b ( h ) b(h) b ( h ) for its value on a test function, or on its class in the cocenter. We also use the realization of b b b as a regular function on the Bernstein variety. These descriptions are related by the Plancherel formula
b ( h ) = ∫ L ( F ) t e m p ∧ b ( π ) tr π ( h ) d ν L ( π ) . b(h)=\int_{L(F)^\wedge_{\mathrm{temp}}}b(\pi)\operatorname{tr}\pi(h)\,d\nu_L(\pi). b ( h ) = ∫ L ( F ) temp ∧ b ( π ) tr π ( h ) d ν L ( π ) .
Here ν L \nu_L ν L is the Plancherel measure. Essential compactness means that b ∗ e K ∈ H ( L ( F ) , K ) b*e_K\in\mathcal{H}(L(F),K) b ∗ e K ∈ H ( L ( F ) , K ) for every K K K .
Bounded functionals and maps attached to Levi subgroups
We say that a subset of L ( F ) L(F) L ( F ) is bounded modulo conjugation if it is contained in
L ( F ) ⋅ C = { g c g − 1 : g ∈ L ( F ) , c ∈ C } L(F)\cdot C=\{gcg^{-1}:g\in L(F),\ c\in C\} L ( F ) ⋅ C = { g c g − 1 : g ∈ L ( F ) , c ∈ C }
for some compact subset C ⊂ L ( F ) C\subset L(F) C ⊂ L ( F ) .
Here is a convenient family of sets exhausting all such supports. Extend the absolute value of F F F to an algebraic closure. For an integer R ≥ 0 R\geq0 R ≥ 0 , let
B R ( L ( F ) ) = { g = ( g j ) ∈ L ( F ) : q − R ≤ ∣ α ∣ ≤ q R for every eigenvalue α of every g j } . (11) B_R(L(F))=\{g=(g_j)\in L(F):q^{-R}\leq|\alpha|\leq q^R\text{ for every eigenvalue }\alpha\text{ of every }g_j\}. \tag*{(11)} B R ( L ( F )) = { g = ( g j ) ∈ L ( F ) : q − R ≤ ∣ α ∣ ≤ q R for every eigenvalue α of every g j } . (11)
These sets are conjugation-invariant, open and closed. Every compact subset of L ( F ) L(F) L ( F ) is contained in some B R ( L ( F ) ) B_R(L(F)) B R ( L ( F )) , by bounding the matrix entries of g j g_j g j and g j − 1 g_j^{-1} g j − 1 . Conversely, each B R ( L ( F ) ) B_R(L(F)) B R ( L ( F )) is bounded modulo conjugation. Indeed, put a matrix in rational canonical form. The coefficients of each invariant-factor polynomial are elementary symmetric functions of a subset of its eigenvalues, hence are uniformly bounded; the constant coefficients are uniformly bounded away from zero. The corresponding companion matrices and their inverses therefore lie in a fixed compact set. There are only finitely many possible block sizes.
We also use the set L ( F ) c m L(F)_{\mathrm{cm}} L ( F ) cm of elements compact modulo the center. For GL ( n , F ) \operatorname{GL}(n,F) GL ( n , F ) this means that all eigenvalues of g g g have the same absolute value. Put
d ( g ) = ( v F ( det g 1 ) , … , v F ( det g s ) ) , L ( F ) c m , k = L ( F ) c m ∩ d − 1 ( k ) , k ∈ Z s . d(g)=(v_F(\det g_1),\ldots,v_F(\det g_s)),\qquad L(F)_{\mathrm{cm},k}=L(F)_{\mathrm{cm}}\cap d^{-1}(k),\quad k\in\mathbb{Z}^s. d ( g ) = ( v F ( det g 1 ) , … , v F ( det g s )) , L ( F ) cm , k = L ( F ) cm ∩ d − 1 ( k ) , k ∈ Z s .
These sets are open and closed. Each L ( F ) c m , k L(F)_{\mathrm{cm},k} L ( F ) cm , k is bounded modulo conjugation: in its j j j th factor all eigenvalues have absolute value q − k j / n j q^{-k_j/n_j} q − k j / n j . For a conjugation-invariant open-and-closed subset U ⊂ L ( F ) U \subset L(F) U ⊂ L ( F ) , multiplication by 1 U \mathbf{1}_U 1 U induces a projector on H ‾ L \overline{\mathcal{H}}_L H L . Denote the projectors for B R ( L ( F ) ) B_R(L(F)) B R ( L ( F )) , L ( F ) c m L(F)_{\mathrm{cm}} L ( F ) cm , and L ( F ) c m , k L(F)_{\mathrm{cm},k} L ( F ) cm , k by p R p_R p R , p c m p_{\mathrm{cm}} p cm , and p c m , k p_{\mathrm{cm},k} p cm , k , respectively.
Fix the standard Iwahori subgroup I ⊂ GL ( n , F ) I \subset\operatorname{GL}(n,F) I ⊂ GL ( n , F ) , and put
A = { a ∈ M n ( O F ) : a m o d ϖ is upper triangular } , I m = 1 + ϖ m A ( m ≥ 1 ) . \mathfrak{A}=\{a\in M_n(\mathcal{O}_F):a\bmod\varpi\text{ is upper triangular}\},\qquad I_m=1+\varpi^m\mathfrak{A}\quad(m\geq1). A = { a ∈ M n ( O F ) : a mod ϖ is upper triangular } , I m = 1 + ϖ m A ( m ≥ 1 ) .
Thus the diagonal entries of an element of I m I_m I m lie in 1 + ϖ m O F 1+\varpi^m\mathcal{O}_F 1 + ϖ m O F , its entries above the diagonal lie in ϖ m O F \varpi^m\mathcal{O}_F ϖ m O F , and those below the diagonal lie in ϖ m + 1 O F \varpi^{m+1}\mathcal{O}_F ϖ m + 1 O F . These are the integer-depth Moy–Prasad subgroups at the barycenter of the standard alcove. For the group of points L ( F ) L(F) L ( F ) of a product L L L , take products of these subgroups. They form a basis of neighbourhoods of 1 1 1 .
For such a subgroup K K K , let V K V_K V K be the image of H ( L ( F ) , K ) \mathcal{H}(L(F),K) H ( L ( F ) , K ) in H ‾ L \overline{\mathcal{H}}_L H L . All the projectors defined above preserve V K V_K V K ; we recall the precise finite-level statement in the proof below.
Definition 5.1. We say that a functional ℓ \ell ℓ on V K V_K V K has bounded support modulo conjugation if
ℓ = ℓ ∘ p R for some R . (12) \ell=\ell\circ p_R\quad\text{for some }R. \tag*{(12)} ℓ = ℓ ∘ p R for some R . (12)
In particular, the restriction of any invariant distribution supported in the conjugates of a compact set has this property.
Induction and restriction
Choose representatives L ( L ) \mathcal{L}(L) L ( L ) for the L ( F ) L(F) L ( F ) -conjugacy classes of Levi F F F -subgroups of L L L , including L L L itself. For M ∈ L ( L ) M\in\mathcal{L}(L) M ∈ L ( L ) , choose a parabolic subgroup P = M U P=MU P = M U of L L L . Then we have the restriction map
r ‾ M L : H ‾ L ⟶ H ‾ M \overline{r}^{L}_{M}:\overline{\mathcal{H}}_L\longrightarrow\overline{\mathcal{H}}_M r M L : H L ⟶ H M
adjoint to normalized parabolic induction. More precisely, for every smooth admissible representation σ \sigma σ of M ( F ) M(F) M ( F ) ,
tr σ ( r ‾ M L h ) = tr ( i P ( F ) L ( F ) σ ) ( h ) . (13) \operatorname{tr}\sigma\left(\overline{r}^{L}_{M}h\right)=\operatorname{tr}\left(i^{L(F)}_{P(F)}\sigma\right)(h). \tag*{(13)} tr σ ( r M L h ) = tr ( i P ( F ) L ( F ) σ ) ( h ) . (13)
This map does not depend on P P P : the character on L ( F ) L(F) L ( F ) of normalized induction from M ( F ) M(F) M ( F ) is independent of the parabolic F F F -subgroup P P P with Levi factor M M M .
In what follows we need a geometric description of this. Choose the Levi representatives to be standard, and choose each P = M U P=MU P = M U to be standard. Let K 0 = L ( O F ) K_0=L(\mathcal{O}_F) K 0 = L ( O F ) be the standard maximal compact subgroup of L ( F ) L(F) L ( F ) , endowed with probability Haar measure d k dk d k . Normalize the Haar measure d u du d u on U ( F ) U(F) U ( F ) by
vol U ( F ) ( K 0 ∩ U ( F ) ) = vol L ( F ) ( K 0 ) vol M ( F ) ( K 0 ∩ M ( F ) ) . \operatorname{vol}_{U(F)}(K_0\cap U(F))=
\frac{\operatorname{vol}_{L(F)}(K_0)}
{\operatorname{vol}_{M(F)}(K_0\cap M(F))}. vol U ( F ) ( K 0 ∩ U ( F )) = vol M ( F ) ( K 0 ∩ M ( F )) vol L ( F ) ( K 0 ) .
The class r ‾ M L h \overline{r}^{L}_{M}h r M L h is represented by the normalized, compact-averaged constant term
h P , K 0 ( m ) = ∣ δ P ( m ) ∣ 1 / 2 ∫ K 0 ∫ U ( F ) h ( k − 1 m u k ) d u d k , m ∈ M ( F ) . (14) h^{P,K_0}(m)=\lvert\delta_P(m)\rvert^{1/2}
\int_{K_0}\int_{U(F)}h(k^{-1}muk)\,du\,dk,\qquad m\in M(F). \tag*{(14)} h P , K 0 ( m ) = ∣ δ P ( m ) ∣ 1/2 ∫ K 0 ∫ U ( F ) h ( k − 1 m u k ) d u d k , m ∈ M ( F ) . (14)
Here δ P \delta_P δ P is the character of M M M given by the determinant of the adjoint action on the Lie algebra of U U U . We shall abbreviate r ‾ M L \overline{r}^{L}_{M} r M L to r ‾ M \overline{r}_M r M when the ambient group is clear.
Thus an invariant distribution α M \alpha_M α M on M ( F ) M(F) M ( F ) induces the invariant distribution α M ∘ r ‾ M \alpha_M\circ\overline{r}_M α M ∘ r M on L ( F ) L(F) L ( F ) . In particular, for b ∈ Z M b\in\mathcal{Z}_M b ∈ Z M the induced distribution is
F M , b ( h ) = b ( r ‾ M L ( h ) ) . F_{M,b}(h)=b\left(\overline{r}^{L}_{M}(h)\right). F M , b ( h ) = b ( r M L ( h ) ) .
Inversion commutes with restriction on cocenters: r ‾ M L ( h ∨ ) = ( r ‾ M L h ) ∨ \overline{r}^{L}_{M}(h^\vee)=\left(\overline{r}^{L}_{M}h\right)^\vee r M L ( h ∨ ) = ( r M L h ) ∨ . This follows from (13), compatibility of normalized induction with contragredients, and the density theorem for characters.
There is also a map of Bernstein centers res M : Z L → Z M \operatorname{res}_M:\mathcal{Z}_L\to\mathcal{Z}_M res M : Z L → Z M , characterized by
r ‾ M ( b ∗ h ) = res M ( b ) ∗ r ‾ M ( h ) . (15) \overline{r}_M(b*h)=\operatorname{res}_M(b)*\overline{r}_M(h). \tag*{(15)} r M ( b ∗ h ) = res M ( b ) ∗ r M ( h ) . (15)
The discriminant and sufficiency
Consider a Bernstein component of G ( F ) = G L ( N , F ) G(F)=\mathrm{GL}(N,F) G ( F ) = GL ( N , F ) of the form
s = [ ∏ a = 1 r G L ( d a , F ) m a , ⨂ a = 1 r σ a ⊗ m a ] , ∑ a d a m a = N , (16) \mathfrak{s}=\left[\prod_{a=1}^{r}\mathrm{GL}(d_a,F)^{m_a},\ \bigotimes_{a=1}^{r}\sigma_a^{\otimes m_a}\right],\qquad\sum_a d_a m_a=N, \tag*{(16)} s = [ a = 1 ∏ r GL ( d a , F ) m a , a = 1 ⨂ r σ a ⊗ m a ] , a ∑ d a m a = N , (16)
where σ a \sigma_a σ a is a unitary cuspidal representation of G L ( d a , F ) \mathrm{GL}(d_a,F) GL ( d a , F ) , and these representations are pairwise inequivalent up to unramified twist. Let
t a = # { χ unramified on F ∗ : σ a ⊗ ( χ ∘ det ) ≃ σ a } . t_a=\#\{\chi\ \text{unramified on }F^\ast:\sigma_a\otimes(\chi\circ\det)\simeq\sigma_a\}. t a = # { χ unramified on F ∗ : σ a ⊗ ( χ ∘ det ) ≃ σ a } .
Coordinates on the torus of distinct unramified twists are given by
z a , j = χ a , j ( ϖ ) t a , 1 ≤ j ≤ m a . z_{a,j}=\chi_{a,j}(\varpi)^{t_a},\qquad1\leq j\leq m_a. z a , j = χ a , j ( ϖ ) t a , 1 ≤ j ≤ m a .
Thus
Z s = ⨂ a C [ z a , 1 ± 1 , … , z a , m a ± 1 ] S m a . Z_{\mathfrak{s}}=\bigotimes_a\mathbb{C}[z_{a,1}^{\pm1},\ldots,z_{a,m_a}^{\pm1}]^{S_{m_a}}. Z s = a ⨂ C [ z a , 1 ± 1 , … , z a , m a ± 1 ] S m a .
Define
Δ s = ∏ a ∏ 1 ≤ i , j ≤ m a i ≠ j ( 1 − z a , i z a , j ) . (17) \Delta_{\mathfrak{s}}=\prod_a\prod_{\substack{1\leq i,j\leq m_a\\i\ne j}}\left(1-\frac{z_{a,i}}{z_{a,j}}\right). \tag*{(17)} Δ s = a ∏ 1 ≤ i , j ≤ m a i = j ∏ ( 1 − z a , j z a , i ) . (17)
Changing σ a \sigma_a σ a by an unramified twist multiplies all the coordinates in its group by the same constant. Consequently (17) is independent of these choices. As usual we set the empty product to be 1. These componentwise elements define
Δ G = ( Δ s ) s ∈ Z G = ∏ s Z s . \Delta_G=(\Delta_{\mathfrak{s}})_{\mathfrak{s}}\in Z_G=\prod_{\mathfrak{s}}Z_{\mathfrak{s}}. Δ G = ( Δ s ) s ∈ Z G = s ∏ Z s .
For L = ∏ j G L ( n j ) L=\prod_j\mathrm{GL}(n_j) L = ∏ j GL ( n j ) , define Δ L ∈ Z L \Delta_L\in Z_L Δ L ∈ Z L by taking the product of these discriminants for its factors.
Theorem 5.2. For every product L L L of general linear groups,
Δ L ∗ D L c ⊂ Z L . \Delta_L*D_L^c\subset Z_L. Δ L ∗ D L c ⊂ Z L . Moreover, Δ L \Delta_L Δ L is a non-zero-divisor in Z L Z_L Z L . Consequently Conjecture 1.1 holds for L L L , and
( Δ L ) ⊂ J L , D L c ⊂ R L . (\Delta_L)\subset J_L,\qquad D_L^c\subset R_L. ( Δ L ) ⊂ J L , D L c ⊂ R L .
The formula for Δ L \Delta_L Δ L is unchanged by contragredience: dualizing the cuspidal representatives and inverting their twist coordinates permutes its factors. In particular, Δ L ∨ = Δ L \Delta_L^\vee=\Delta_L Δ L ∨ = Δ L . We prove the theorem using the following two statements.
Theorem 5.3. Let L L L be a product of general linear groups.
(Decomposition over Levi subgroups.) Let K ⊂ L ( F ) K\subset L(F) K ⊂ L ( F ) be one of the congruence subgroups specified above, and let ℓ : V K → C \ell:V_K\to\mathbb{C} ℓ : V K → C have bounded support modulo conjugation. There are regular central elements b M ∈ Z M b_M\in Z_M b M ∈ Z M , each supported on finitely many Bernstein components, such that
ℓ ( h ) = ∑ M ∈ L ( L ) b M ( r ˉ M L ( h ) ) , h ∈ H ( L ( F ) , K ) . (18) \ell(h)=\sum_{M\in\mathcal{L}(L)}b_M(\bar r_M^L(h)),\qquad h\in\mathcal{H}(L(F),K). \tag*{(18)} ℓ ( h ) = M ∈ L ( L ) ∑ b M ( r ˉ M L ( h )) , h ∈ H ( L ( F ) , K ) . (18) The term M = L M=L M = L is allowed. No discriminant occurs in this decomposition.
Let M M M be a Levi subgroup of L L L and let b ∈ Z M b\in Z_M b ∈ Z M have finite Bernstein support. For the induced distribution
F M , b ( h ) = b ( r ˉ M L ( h ) ) F_{M,b}(h)=b(\bar r_M^L(h)) F M , b ( h ) = b ( r ˉ M L ( h )) there is a regular central element A M , b ∈ Z L A_{M,b}\in Z_L A M , b ∈ Z L , with finite Bernstein support, such that
F M , b ( Δ L ∗ h ) = A M , b ( h ) , h ∈ H L . (19) F_{M,b}(\Delta_L*h)=A_{M,b}(h),\qquad h\in\mathcal{H}_L. \tag*{(19)} F M , b ( Δ L ∗ h ) = A M , b ( h ) , h ∈ H L . (19)
Corollary 5.4. Let L L L be a product of general linear groups and let K ⊂ L ( F ) K \subset L(F) K ⊂ L ( F ) be a congruence subgroup as above. For every functional ℓ : V K → C \ell: V_K \to\mathbb{C} ℓ : V K → C with bounded support modulo conjugation, the functional Δ L ℓ \Delta_L\ell Δ L ℓ is regular: there exists an element a ℓ ∈ Z L a_\ell\in\mathcal{Z}_L a ℓ ∈ Z L , with finite Bernstein support, such that
( Δ L ℓ ) ( h ) : = ℓ ( Δ L ∗ h ) = a ℓ ( h ) , h ∈ H ( L ( F ) , K ) . (20) (\Delta_L\ell)(h) := \ell(\Delta_L * h) = a_\ell(h), \qquad h \in\mathcal{H}(L(F),K). \tag*{(20)} ( Δ L ℓ ) ( h ) := ℓ ( Δ L ∗ h ) = a ℓ ( h ) , h ∈ H ( L ( F ) , K ) . (20) Thus multiplication by Δ L \Delta_L Δ L makes every such functional come from an element of the Bernstein center, viewed as a distribution.
Proof. Apply statement (1) of Theorem 5.3 to ℓ \ell ℓ , and statement (2) to each summand of (18). Central operators preserve H ( L ( F ) , K ) \mathcal{H}(L(F),K) H ( L ( F ) , K ) , so for every h h h in this algebra,
ℓ ( Δ L ∗ h ) = ∑ M ∈ L ( L ) A M , b M ( h ) = ( ∑ M ∈ L ( L ) A M , b M ) ( h ) . \ell(\Delta_L * h) = \sum_{M\in\mathcal{L}(L)} A_{M,b_M}(h) = \left(\sum_{M\in\mathcal{L}(L)} A_{M,b_M}\right)(h). ℓ ( Δ L ∗ h ) = M ∈ L ( L ) ∑ A M , b M ( h ) = M ∈ L ( L ) ∑ A M , b M ( h ) .
Thus a ℓ = ∑ M A M , b M a_\ell= \sum_M A_{M,b_M} a ℓ = ∑ M A M , b M has the required properties. □
The proof of statement (2) uses the type A A A Plancherel formulas. Statement (1) uses separation of tempered representations by cuspidal support, which is special to general linear groups. Neither argument is asserted here for an arbitrary reductive group.
Proof of Theorem 5.3(1)
We use the following consequences of the finite-level Newton decomposition; see [11 Theorem 11 and Corollary 9] , with the numbering of the published version:
The projections p R p_R p R , p c m p_{\mathrm{cm}} p cm , and p c m , k p_{\mathrm{cm},k} p cm , k preserve V K V_K V K . Their images, and the images of their complementary projectors, have representatives in H ( L ( F ) , K ) \mathcal{H}(L(F),K) H ( L ( F ) , K ) supported in the indicated subsets.
The spaces p R V K p_RV_K p R V K and p c m , k V K p_{\mathrm{cm},k}V_K p cm , k V K are finite dimensional, and
V K = ⋃ R ≥ 0 p R V K , p c m V K = ⨁ k ∈ Z s p c m , k V K . V_K = \bigcup_{R\geq0} p_RV_K, \qquad p_{\mathrm{cm}}V_K = \bigoplus_{k\in\mathbb{Z}^s} p_{\mathrm{cm},k}V_K. V K = R ≥ 0 ⋃ p R V K , p cm V K = k ∈ Z s ⨁ p cm , k V K .
For the first finite-dimensionality assertion one can equivalently apply Howe’s finiteness theorem to a compact set whose conjugates cover B R ( L ( F ) ) B_R(L(F)) B R ( L ( F )) ; see [11 Theorem 18] . The assertion for L ( F ) c m , k L(F)_{\mathrm{cm},k} L ( F ) cm , k follows by its containment in some B R ( L ( F ) ) B_R(L(F)) B R ( L ( F )) .
The statement below is crucial for our argument (and this is where we use heavily the fact that L L L is a product of general linear groups).
Proposition 5.5. The functionals
V K ⟶ C , h ⟼ b ( h ) , b ∈ Z L , V_K \longrightarrow\mathbb{C}, \qquad h \longmapsto b(h), \qquad b \in\mathcal{Z}_L, V K ⟶ C , h ⟼ b ( h ) , b ∈ Z L , separate V K V_K V K ; in other words if h ∈ V K h \in V_K h ∈ V K is such that b ( h ) = 0 b(h)=0 b ( h ) = 0 for all b ∈ Z L b \in\mathcal{Z}_L b ∈ Z L , then h = 0 h=0 h = 0 .
Corollary 5.6. Let W ⊂ V K W \subset V_K W ⊂ V K be a finite-dimensional subspace of V K V_K V K . Then the map Z L → W ∗ \mathcal{Z}_L \to W^* Z L → W ∗ given by b ↦ ( h ↦ b ( h ) ) b \mapsto(h \mapsto b(h)) b ↦ ( h ↦ b ( h )) is surjective. The element b b b may be chosen to be supported on the finitely many Bernstein components visible at level K K K .
The corollary follows by finite-dimensional linear algebra. A central element may be multiplied by the sum of the finitely many Bernstein projectors visible at level K K K without changing its restriction to V K V_K V K . We now prove Proposition 5.5.
Proof. The key observation is that an irreducible tempered representation of a group G L ( n , F ) \mathrm{GL}(n,F) GL ( n , F ) is determined by its cuspidal support. To see this, work separately in each factor G L ( n j , F ) \mathrm{GL}(n_j,F) GL ( n j , F ) of L ( F ) L(F) L ( F ) , and group the cuspidal constituents of the cuspidal support by inertial equivalence. For each class choose a unitary representative σ a \sigma_a σ a , and use the coordinates z = χ ( ϖ ) t a z=\chi(\varpi)^{t_a} z = χ ( ϖ ) t a introduced above, where t a t_a t a is the number of unramified self-twists of σ a \sigma_a σ a . The parameters in this class form a multiset union of strings
u Q ( ℓ − 1 ) / 2 , u Q ( ℓ − 3 ) / 2 , … , u Q − ( ℓ − 1 ) / 2 , ∣ u ∣ = 1 , Q = q t a . (21) uQ^{(\ell-1)/2},\ uQ^{(\ell-3)/2},\ \ldots,\ uQ^{-(\ell-1)/2}, \qquad|u|=1,\qquad Q=q^{t_a}. \tag*{(21)} u Q ( ℓ − 1 ) /2 , u Q ( ℓ − 3 ) /2 , … , u Q − ( ℓ − 1 ) /2 , ∣ u ∣ = 1 , Q = q t a . (21)
The entries of largest absolute value determine the longest strings and their unitary parameters. Removing these strings and repeating recovers the complete multiset of strings. The classification of tempered representations of GL ( n , F ) \operatorname{GL}(n,F) GL ( n , F ) , including irreducibility of normalized induction of irreducible tempered representations, therefore recovers the representation from its cuspidal support.
Let X K X_K X K be the set of irreducible tempered representations of L ( F ) L(F) L ( F ) with a nonzero K K K -fixed vector. Only finitely many Bernstein components are visible at level K K K . Within each of them there are finitely many tempered induction families, parametrized by compact tori modulo finite groups. The function π ↦ dim π K \pi\mapsto\dim\pi^K π ↦ dim π K is constant on each such family, so X K X_K X K is a finite union of these compact families. The cuspidal-support map identifies X K X_K X K with a compact subset of the Bernstein variety. In particular, X K X_K X K is Hausdorff and regular central functions separate its points.
Their restrictions to X K X_K X K are also closed under complex conjugation. Indeed, every multiset of strings in (21) is invariant under z ↦ z ˉ − 1 z \mapsto\bar{z}^{-1} z ↦ z ˉ − 1 . Thus, for a symmetric Laurent polynomial b b b , conjugating its coefficients and replacing every coordinate by its inverse gives a regular function whose restriction to X K X_K X K is b ˉ \bar{b} b ˉ .
The complex Stone–Weierstrass theorem says that a subalgebra of C ( X ) C(X) C ( X ) on a compact Hausdorff space X X X is uniformly dense if it contains the constant functions, separates points, and is closed under complex conjugation. These conditions hold for the restrictions of Z L \mathcal{Z}_L Z L to X K X_K X K .
Now let h ∈ H ( L ( F ) , K ) h \in\mathcal{H}(L(F),K) h ∈ H ( L ( F ) , K ) represent a class annihilated by all b ∈ Z L b \in\mathcal{Z}_L b ∈ Z L , and put f h ( π ) = tr π ( h ∨ ) f_h(\pi)=\operatorname{tr}\pi(h^\vee) f h ( π ) = tr π ( h ∨ ) . This is a continuous function on X K X_K X K , and it vanishes outside X K X_K X K . Plancherel gives
0 = ∫ X K b ( π ) f h ( π ) d ν L ( π ) ( b ∈ Z L ) . 0=\int_{X_K} b(\pi)f_h(\pi)\,d\nu_L(\pi) \qquad(b\in\mathcal{Z}_L). 0 = ∫ X K b ( π ) f h ( π ) d ν L ( π ) ( b ∈ Z L ) .
The measure of X K X_K X K is finite, since
∫ X K dim π K d ν L ( π ) = e K ( 1 ) < ∞ . \int_{X_K}\dim\pi^K\,d\nu_L(\pi)=e_K(1)<\infty. ∫ X K dim π K d ν L ( π ) = e K ( 1 ) < ∞.
Uniformly approximate f h ‾ \overline{f_h} f h by regular central functions. Passing to the limit yields
∫ X K ∣ f h ( π ) ∣ 2 d ν L ( π ) = 0. \int_{X_K}|f_h(\pi)|^2\,d\nu_L(\pi)=0. ∫ X K ∣ f h ( π ) ∣ 2 d ν L ( π ) = 0.
The Plancherel measure has full support on every tempered induction family in X K X_K X K . Continuity therefore gives f h ( π ) = 0 f_h(\pi)=0 f h ( π ) = 0 at every point of X K X_K X K , and hence on the entire tempered dual. To pass from tempered traces to all admissible traces, fix a tempered induction datum and vary its unramified characters. The trace of h ∨ h^\vee h ∨ is a Laurent polynomial and vanishes on the unitary torus, so it vanishes on the whole complex torus. The Langlands classification expresses every irreducible character as a linear combination of such standard-module characters. Spectral density, in the form of [10 Theorem 2.8] , now implies that the class of h ∨ h^\vee h ∨ , and hence of h h h , in H L \mathcal{H}_L H L is zero. This form applies over non-Archimedean local fields of either characteristic. For the split groups used here, see also [12 Theorem 0] in characteristic zero and [13 ] for passage to positive characteristic.
We prove Theorem 5.3(1) by induction on the semisimple rank of L L L . The following statement will be needed in this induction argument.
Lemma 5.7. For every b ∈ Z L b\in\mathcal{Z}_L b ∈ Z L , its restriction to V K V_K V K has bounded support modulo conjugation.
Proof. The function b K = b ∗ e K b_K=b*e_K b K = b ∗ e K is compactly supported. Choose R R R with supp ( b K ) ⊂ B R ( L ( F ) ) \operatorname{supp}(b_K)\subset B_R(L(F)) supp ( b K ) ⊂ B R ( L ( F )) . A class in ( 1 − p R ) V K (1-p_R)V_K ( 1 − p R ) V K has a representative h ∈ H ( L ( F ) , K ) h\in\mathcal{H}(L(F),K) h ∈ H ( L ( F ) , K ) supported outside B R ( L ( F ) ) B_R(L(F)) B R ( L ( F )) . For this representative,
b ( h ) = ∫ L ( F ) b K ( g ) h ( g ) d g = 0. b(h)=\int_{L(F)}b_K(g)h(g)\,dg=0. b ( h ) = ∫ L ( F ) b K ( g ) h ( g ) d g = 0.
Thus b ∣ V K = ( b ∣ V K ) ∘ p R b|_{V_K}=(b|_{V_K})\circ p_R b ∣ V K = ( b ∣ V K ) ∘ p R .
:::
We now use the following cocenter theorem of Ciubotaru–He:
⋂ M ⊊ L ker r ‾ M ⊂ H ‾ L , p c m (22) \bigcap_{M\subsetneq L}\ker\overline{r}_M\subset\overline{\mathcal{H}}_{L,\mathrm{pcm}} \tag*{(22)} M ⊊ L ⋂ ker r M ⊂ H L , pcm (22)
where M M M ranges over L ( L ) ∖ { L } \mathcal{L}(L)\setminus\{L\} L ( L ) ∖ { L } . In words, a cocenter class whose restrictions to all proper Levi subgroups vanish is represented by functions supported on elements compact modulo the center; see [10 Theorem 4.4] .
There is also the elementary compatibility
r ‾ M p R L = p R M r ‾ M . (23) \overline{r}_M p_R^L=p_R^M\overline{r}_M. \tag*{(23)} r M p R L = p R M r M . (23)
Indeed, the compact-averaged constant-term formula defining r ‾ M \overline{r}_M r M evaluates h h h on conjugates of m u mu m u , with m ∈ M ( F ) m\in M(F) m ∈ M ( F ) and u ∈ U ( F ) u\in U(F) u ∈ U ( F ) , where U U U is the unipotent radical of P P P . A block upper-triangular matrix m u mu m u has precisely the eigenvalues of its diagonal blocks m m m , so m u ∈ B R ( L ( F ) ) mu\in B_R(L(F)) m u ∈ B R ( L ( F )) if and only if m ∈ B R ( M ( F ) ) m\in B_R(M(F)) m ∈ B R ( M ( F )) .
Lemma 5.8. Let ℓ \ell ℓ be a functional on V K V_K V K with bounded support modulo conjugation, and suppose that ℓ \ell ℓ vanishes on p c m V K p_{\mathrm{cm}}V_K p cm V K . Then, for suitable congruence subgroups K M ⊂ M ( F ) K_M\subset M(F) K M ⊂ M ( F ) , there are functionals ℓ M \ell_M ℓ M on V K M V_{K_M} V K M , each with bounded support modulo conjugation, such that
ℓ = ∑ M ∈ L ( L ) , M ≠ L ℓ M ∘ r ‾ M on V K . (24) \ell=\sum_{\substack{M\in\mathcal{L}(L),\,M\ne L}}\ell_M\circ\overline{r}_M\quad\text{on }V_K. \tag*{(24)} ℓ = M ∈ L ( L ) , M = L ∑ ℓ M ∘ r M on V K . (24)
Proof. Choose R R R such that ℓ = ℓ ∘ p R \ell=\ell\circ p_R ℓ = ℓ ∘ p R . Choose K M K_M K M sufficiently small that r ‾ M ( V K ) ⊂ V K M \overline{r}_M(V_K)\subset V_{K_M} r M ( V K ) ⊂ V K M for each M M M . A uniform level exists by the compact-averaged constant-term formula: first pass to a principal congruence subgroup normal in a maximal compact subgroup and then to a smaller subgroup of the prescribed form in M ( F ) M(F) M ( F ) .
Consider the map of finite-dimensional spaces
p R L V K ⟶ ⨁ M ⊊ L p R M V K M , h ⟼ ( r ‾ M h ) M . p_R^L V_K\longrightarrow\bigoplus_{M\subsetneq L}p_R^M V_{K_M},\qquad h\longmapsto(\overline{r}_M h)_M. p R L V K ⟶ M ⊊ L ⨁ p R M V K M , h ⟼ ( r M h ) M .
By (22), its kernel is contained in p c m V K p_{\mathrm{cm}}V_K p cm V K , so ℓ \ell ℓ annihilates that kernel. Thus ℓ \ell ℓ factors through the image of this map. Extend the resulting functional to the displayed direct sum and write it as a sum of functionals λ M \lambda_M λ M on the individual factors. Set ℓ M = λ M ∘ p R M \ell_M=\lambda_M\circ p_R^M ℓ M = λ M ∘ p R M on V K M V_{K_M} V K M . Equation (23) gives, for every h ∈ V K h\in V_K h ∈ V K ,
ℓ ( h ) = ℓ ( p R L h ) = ∑ M λ M ( r ‾ M p R L h ) = ∑ M ℓ M ( r ‾ M h ) . \ell(h)=\ell(p_R^Lh)=\sum_M\lambda_M(\overline{r}_Mp_R^Lh)=\sum_M\ell_M(\overline{r}_Mh). ℓ ( h ) = ℓ ( p R L h ) = M ∑ λ M ( r M p R L h ) = M ∑ ℓ M ( r M h ) .
Each ℓ M \ell_M ℓ M satisfies (12), as required. □
End of the proof of Theorem 5.3(1)
We use induction on the semisimple rank of L L L . If L L L is a torus, the bounded-support condition means that the functional on H ( L ( F ) , K ) \mathcal{H}(L(F),K) H ( L ( F ) , K ) is supported on finitely many cosets in L ( F ) / K L(F)/K L ( F ) / K . It is therefore represented, using the fixed Haar measure, by a compactly supported K K K -invariant function. As L ( F ) L(F) L ( F ) is abelian, this defines a central distribution; its Bernstein support is contained in the finitely many components visible at level K K K . This proves the base case.
Suppose now that the assertion holds for proper Levi subgroups of L L L . We first match ℓ \ell ℓ on all elements compact modulo the center. If ℓ = ℓ ∘ p R \ell=\ell\circ p_R ℓ = ℓ ∘ p R , its restriction to p c m , k V K p_{\mathrm{cm},k}V_K p cm , k V K can be nonzero only for the finitely many multidegrees satisfying ∣ k j ∣ ≤ n j R |k_j|\le n_jR ∣ k j ∣ ≤ n j R for every j j j . For each such k k k , Corollary 5.6 applied to W = p c m , k V K W=p_{\mathrm{cm},k}V_K W = p cm , k V K implies that there exists some b k ∈ Z L b_k\in\mathcal{Z}_L b k ∈ Z L with finite Bernstein support such that
b k ( h ) = ℓ ( h ) , h ∈ p c m , k V K . b_k(h)=\ell(h),\qquad h\in p_{\mathrm{cm},k}V_K. b k ( h ) = ℓ ( h ) , h ∈ p cm , k V K .
Retain only the homogeneous component of b k b_k b k of determinant multidegree k k k . More explicitly, for c ∈ ( C × ) s \mathbf{c}\in(\mathbb{C}^{\times})^s c ∈ ( C × ) s put
ν c ( g ) = ∏ j c j v F ( det g j ) , ( T c b ) ( h ) = b ( ν c h ) . \nu_{\mathbf{c}}(g)=\prod_j c_j^{v_F(\det g_j)},\qquad(T_{\mathbf{c}}b)(h)=b(\nu_{\mathbf{c}}h). ν c ( g ) = j ∏ c j v F ( d e t g j ) , ( T c b ) ( h ) = b ( ν c h ) .
Its spectral function is ( T c b ) ( π ) = b ( π ⊗ ν c ) (T_c b)(\pi)=b(\pi\otimes\nu_c) ( T c b ) ( π ) = b ( π ⊗ ν c ) . For a regular central element b b b with finite Bernstein support the algebraic twisting action has a finite Laurent expansion
T c b = ∑ k ∈ Z s c k b [ k ] , b [ k ] ∈ Z L . T_c b=\sum_{k\in\mathbb{Z}^s}\mathbf{c}^k b^{[k]},\qquad b^{[k]}\in\mathcal{Z}_L. T c b = k ∈ Z s ∑ c k b [ k ] , b [ k ] ∈ Z L .
If h h h is supported in determinant multidegree j j j , then ( T c b ) ( h ) = c j b ( h ) (T_c b)(h)=\mathbf{c}^j b(h) ( T c b ) ( h ) = c j b ( h ) . Comparison of coefficients gives b [ k ] ( h ) = δ k , j b ( h ) b^{[k]}(h)=\delta_{k,j}b(h) b [ k ] ( h ) = δ k , j b ( h ) . Thus replacing b k b_k b k by b [ k ] b^{[k]} b [ k ] preserves the displayed equality, retains finite Bernstein support, and makes the distribution vanish on every other determinant multidegree.
Set b L = ∑ k b k b_L=\sum_k b_k b L = ∑ k b k . Then
b L ( h ) = ℓ ( h ) , h ∈ P c m V K . b_L(h)=\ell(h),\qquad h\in\mathcal{P}_{\mathrm{cm}}V_K. b L ( h ) = ℓ ( h ) , h ∈ P cm V K .
Consequently η ( h ) = ℓ ( h ) − b L ( h ) \eta(h)=\ell(h)-b_L(h) η ( h ) = ℓ ( h ) − b L ( h ) vanishes on P c m V K \mathcal{P}_{\mathrm{cm}}V_K P cm V K . It still has bounded support modulo conjugation, by Lemma 5.7. Lemma 5.8 gives
η = ∑ M ∈ L ( L ) , M ≠ L ℓ M ∘ r ˉ M − L . \eta=\sum_{\substack{M\in\mathcal{L}(L),\ M\ne L}}\ell_M\circ\bar r_M^{-L}. η = M ∈ L ( L ) , M = L ∑ ℓ M ∘ r ˉ M − L .
where ℓ M \ell_M ℓ M has bounded support on a suitable V K M V_{K_M} V K M . Apply the induction hypothesis to every ℓ M \ell_M ℓ M . Each is a sum of distributions induced from Bernstein center elements of Levi subgroups D D D of M M M . Transitivity,
r ˉ D M ∘ r ˉ M L = r ˉ D L , \bar r_D^M\circ\bar r_M^L=\bar r_D^L, r ˉ D M ∘ r ˉ M L = r ˉ D L ,
turns these into distributions induced from Bernstein center elements of Levi subgroups of L L L . Transporting conjugate Levi subgroups to the chosen representatives and adding their central elements gives (18). There are only finitely many terms, and every central element obtained has finite Bernstein support which finishes the proof.
Proof of Theorem 5.3(2)
We use an explicit calculation with the Plancherel formula. First observe that the quotient
C M = res M ( Δ L ) Δ M (25) C_M=\frac{\operatorname{res}_M(\Delta_L)}{\Delta_M} \tag*{(25)} C M = Δ M res M ( Δ L ) (25)
is a regular element of Z M \mathcal{Z}_M Z M . Since the discriminants are invariant under inversion and restriction commutes with inversion, C M ∨ = C M C_M^\vee=C_M C M ∨ = C M . We now claim
Lemma 5.9. For a ∈ Z M a\in\mathcal{Z}_M a ∈ Z M supported on finitely many Bernstein components, there is T M ( a ) ∈ Z L T_M(a)\in\mathcal{Z}_L T M ( a ) ∈ Z L , also supported on finitely many components, such that
( a ∗ C M ) ( r ˉ M ( h ) ) = T M ( a ) ( h ) , h ∈ H L . (26) (a*C_M)(\bar r_M(h))=T_M(a)(h),\qquad h\in\mathcal{H}_L. \tag*{(26)} ( a ∗ C M ) ( r ˉ M ( h )) = T M ( a ) ( h ) , h ∈ H L . (26)
Proof. Fix a component t \mathfrak{t} t of M ( F ) M(F) M ( F ) mapping to a component s \mathfrak{s} s of L ( F ) L(F) L ( F ) . For a product L L L , the following indices and cross pairs are taken separately within each of its general linear factors. For each inertial class represented by σ a \sigma_a σ a , write m a , b m_{a,b} m a , b for the number of its occurrences assigned to the b b b th M M M -factor. Put
Q a = q t a , W s = ∏ a S m a , W t = ∏ a , b S m a , b , Q_a=q^{t_a},\qquad W_{\mathfrak{s}}=\prod_a S_{m_a},\qquad W_{\mathfrak{t}}=\prod_{a,b}S_{m_{a,b}}, Q a = q t a , W s = a ∏ S m a , W t = a , b ∏ S m a , b ,
and define the regular function
P s / t = ∏ a ∏ i , j assigned to different M -factors ( 1 − Q a − 1 z a , i z a , j ) . (27) P_{\mathfrak{s}/\mathfrak{t}}=\prod_a\prod_{\substack{i,j\ \text{assigned to}\\\text{different }M\text{-factors}}}
\left(1-Q_a^{-1}\frac{z_{a,i}}{z_{a,j}}\right). \tag*{(27)} P s / t = a ∏ i , j assigned to different M -factors ∏ ( 1 − Q a − 1 z a , j z a , i ) . (27)
Both orientations of every cross pair occur in this product.
For each type A affine Hecke algebra use the canonical trace τ ( T w ) = δ w , 1 \tau(T_w)=\delta_{w,1} τ ( T w ) = δ w , 1 in its standard basis, and probability Haar measure on its compact unramified-parameter torus. The Plancherel measure of the component s \mathfrak{s} s is a constant κ L , s > 0 \kappa_{L,\mathfrak{s}}>0 κ L , s > 0 times the product of these normalized type A affine-Hecke Plancherel measures with parameters Q a Q_a Q a . The same holds for ( M , t ) (M,\mathfrak{t}) ( M , t ) , with constant κ M , t > 0 \kappa_{M,\mathfrak{t}}>0 κ M , t > 0 . These comparison constants incorporate the fixed Haar measures on the groups. This includes all tempered induction families, by [9 Theorems 5.7 and 6.12] .
Set
c t = κ M , t κ L , s ∏ a Q a ( m a 2 ) − ∑ b ( m a , b 2 ) . c_t=\frac{\kappa_{M,t}}{\kappa_{L,s}}\prod_a Q_a^{\binom{m^a}{2}-\sum_b\binom{m^{a,b}}{2}}. c t = κ L , s κ M , t a ∏ Q a ( 2 m a ) − ∑ b ( 2 m a , b ) .
The contribution of a t a_t a t to T M ( a ) ∣ s \mathcal{T}_M(a)|_s T M ( a ) ∣ s is
c t ∑ w ∈ W s / W t w ( a t P s / t ) . (28) c_t\sum_{w\in W_s/W_t}w(a_tP_{s/t}). \tag*{(28)} c t w ∈ W s / W t ∑ w ( a t P s / t ) . (28)
It is regular and W s W_s W s -invariant.
We verify the distribution identity by applying Plancherel to h ∨ h^\vee h ∨ and using r M ( h ∨ ) ‾ = ( r M h ‾ ) ∨ \overline{r_M(h^\vee)}=(\overline{r_Mh})^\vee r M ( h ∨ ) = ( r M h ) ∨ . On each tempered family, the parameters are unions of the strings (21) . If a placement of the coordinates splits a string between two M M M -factors, some adjacent pair in that string lies in different factors. Their ratio is Q a Q_a Q a in one orientation, so the corresponding factor in (27) vanishes. Thus only placements of entire strings in M M M -factors contribute.
For a surviving placement, pass to a common parameter cover of the tempered family on the two sides. The relative type A A A Plancherel factor is
( ∏ a Q a − ( m a 2 ) + ∑ b ( m a , b 2 ) ) C M P s / t . \left(\prod_a Q_a^{-\binom{m^a}{2}+\sum_b\binom{m^{a,b}}{2}}\right)\frac{C_M}{P_{s/t}}. ( a ∏ Q a − ( 2 m a ) + ∑ b ( 2 m a , b ) ) P s / t C M .
The discrete-series factors are the same on this common cover. We spell out the finite-group multiplicities, making the following count separately in each general linear factor of L L L . For each a a a , let n a , ℓ n_{a,\ell} n a , ℓ be the number of strings of length ℓ \ell ℓ in the chosen tempered family, and put
A = ∏ a , ℓ n a , ℓ ! . A=\prod_{a,\ell}n_{a,\ell}!. A = a , ℓ ∏ n a , ℓ ! .
Fix an assignment type of entire strings to the M M M -factors, and let n a , b , ℓ n_{a,b,\ell} n a , b , ℓ be the number of length-ℓ \ell ℓ strings of class a a a assigned to factor b b b . Thus
∑ b n a , b , ℓ = n a , ℓ , ∑ ℓ ℓ n a , b , ℓ = m a , b . \sum_b n_{a,b,\ell}=n_{a,\ell},\qquad\sum_\ell\ell n_{a,b,\ell}=m_{a,b}. b ∑ n a , b , ℓ = n a , ℓ , ℓ ∑ ℓ n a , b , ℓ = m a , b .
Put B = ∏ a , b , ℓ n a , b , ℓ ! B=\prod_{a,b,\ell}n_{a,b,\ell}! B = ∏ a , b , ℓ n a , b , ℓ ! . On the ordered torus of string parameters, the Weyl factors in the Plancherel measures of the L L L -family and this M M M -family are respectively 1 / A 1/A 1/ A and 1 / B 1/B 1/ B . Exactly A / B A/B A / B cosets in W s / W t W_s/W_t W s / W t give placements of this assignment type. Permuting strings of equal length within each class a a a carries any such placement to any other. Although the corresponding values of a t a_t a t need not agree pointwise, the integrals of the corresponding summands in (28) agree after this relabelling: the ordered Haar measure, the L L L -Plancherel density, and the induced L L L -character are invariant under these permutations. Consequently the coset sum converts the factor 1 / A 1/A 1/ A into 1 / B 1/B 1/ B , as required.
The power of Q a Q_a Q a is independent of the string partition. Indeed, if the strings in class a a a have lengths ℓ a , i \ell_{a,i} ℓ a , i , the difference between the exponents in the two parabolic Plancherel factors is
( ( m a 2 ) − ∑ i ( ℓ a , i 2 ) ) − ( ∑ b ( ( m a , b 2 ) − ∑ i ( ℓ a , i 2 ) ) ) = ( m a 2 ) − ∑ b ( m a , b 2 ) . \left(\binom{m_a}{2}-\sum_i\binom{\ell_{a,i}}{2}\right)-\left(\sum_b\left(\binom{m_{a,b}}{2}-\sum_i\binom{\ell_{a,i}}{2}\right)\right)=\binom{m_a}{2}-\sum_b\binom{m_{a,b}}{2}. ( ( 2 m a ) − i ∑ ( 2 ℓ a , i ) ) − ( b ∑ ( ( 2 m a , b ) − i ∑ ( 2 ℓ a , i ) ) ) = ( 2 m a ) − b ∑ ( 2 m a , b ) .
This is precisely the exponent occurring in c t c_t c t . The relative density formula used here is the parabolic factorization of Plancherel measure [15 Proposition 3.27 and formula (4.96)] . Together with the componentwise constants κ M , t / κ L , s \kappa_{M,t}/\kappa_{L,s} κ M , t / κ L , s , it shows that multiplication by (28) gives exactly the contribution of ( a ∗ C M ) ( r M h ‾ ) (a\ast C_M)(\overline{r_Mh}) ( a ∗ C M ) ( r M h ) from the chosen assignment type. For a product L L L , take the product of these multiplicity and density comparisons. Summing over assignment types gives the identity on the chosen tempered family. These comparisons are made where the string parameters are in general position; the omitted subset has measure zero in each family. Summing over all tempered families and the finitely many relevant Bernstein components proves (26) .
Proof of Theorem 5.3(2). For b ∈ Z M b \in\mathcal{Z}_M b ∈ Z M as in the statement, set
A M , b = T M ( b Δ M ) ∈ Z L . A_{M,b} = T_M(b\Delta_M) \in\mathcal{Z}_L. A M , b = T M ( b Δ M ) ∈ Z L .
This element is regular and has finite Bernstein support. By (15) , (25) , and Lemma 5.9, we obtain
F M , b ( Δ L ∗ h ) = b ( res M ( Δ L ) ∗ r ‾ M h ) = ( b ∗ Δ M ∗ C M ) ( r ‾ M h ) = T M ( b Δ M ) ( h ) . \begin{aligned}
F_{M,b}(\Delta_L * h) &= b(\operatorname{res}_M(\Delta_L) * \overline{r}_M h) \\
&= (b * \Delta_M * C_M)(\overline{r}_M h) \\
&= T_M(b\Delta_M)(h).
\end{aligned} F M , b ( Δ L ∗ h ) = b ( res M ( Δ L ) ∗ r M h ) = ( b ∗ Δ M ∗ C M ) ( r M h ) = T M ( b Δ M ) ( h ) .
Here we used Δ M ∨ = Δ M \Delta_M^\vee= \Delta_M Δ M ∨ = Δ M and C M ∨ = C M C_M^\vee= C_M C M ∨ = C M to pass from convolution on test functions to convolution on distributions. This is (19) . When M = L M = L M = L , the assertion simply reads A L , b = b Δ L A_{L,b} = b\Delta_L A L , b = b Δ L . For a proper Levi subgroup, it is the Plancherel comparison in Lemma 5.9 that supplies the regular numerator.
Passage from finite levels to invariant distributions
Proof of Theorem 5.2. Let α ∈ D G c \alpha\in\mathcal{D}_G^c α ∈ D G c . Its restriction to every V K V_K V K has bounded support in the sense of (12) . By Corollary 5.4 there is a K ∈ Z L a_K \in\mathcal{Z}_L a K ∈ Z L such that
α ( Δ L ∗ h ) = a K ( h ) , h ∈ H ( L ( F ) , K ) . (29) \alpha(\Delta_L * h) = a_K(h), \qquad h \in\mathcal{H}(L(F),K). \tag*{(29)} α ( Δ L ∗ h ) = a K ( h ) , h ∈ H ( L ( F ) , K ) . (29)
Put β = Δ L ∗ α \beta= \Delta_L * \alpha β = Δ L ∗ α . Since Δ L ∨ = Δ L \Delta_L^\vee= \Delta_L Δ L ∨ = Δ L , the definition of the central action gives
β ( h ) = α ( Δ L ∗ h ) = a K ( h ) , h ∈ H ( L ( F ) , K ) . \beta(h) = \alpha(\Delta_L * h) = a_K(h), \qquad h \in\mathcal{H}(L(F),K). β ( h ) = α ( Δ L ∗ h ) = a K ( h ) , h ∈ H ( L ( F ) , K ) .
The convolution v K = β ∗ e K v_K = \beta* e_K v K = β ∗ e K is a locally constant K K K -bi-invariant function; compact support is not needed to define this convolution with a test function. For h ∈ H ( L ( F ) , K ) h \in\mathcal{H}(L(F),K) h ∈ H ( L ( F ) , K ) , (29) becomes
∫ L ( F ) v K ( g ) h ( g ) d g = ∫ L ( F ) ( a K ∗ e K ) ( g ) h ( g ) d g . \int_{L(F)} v_K(g)h(g)\,\mathrm{d}g = \int_{L(F)} (a_K * e_K)(g)h(g)\,\mathrm{d}g. ∫ L ( F ) v K ( g ) h ( g ) d g = ∫ L ( F ) ( a K ∗ e K ) ( g ) h ( g ) d g .
Characteristic functions of individual K K K -double cosets show that v K = a K ∗ e K v_K = a_K * e_K v K = a K ∗ e K . Thus β ∗ e K \beta* e_K β ∗ e K has compact support. The chosen congruence subgroups form a neighbourhood basis of the identity. Every test function is bi-invariant under one of them, and therefore β ∗ h = ( β ∗ e K ) ∗ h \beta* h = (\beta* e_K) * h β ∗ h = ( β ∗ e K ) ∗ h is compactly supported for every h ∈ H L h \in\mathcal{H}_L h ∈ H L . The distribution β \beta β is invariant and essentially compact, hence belongs to Z L \mathcal{Z}_L Z L .
Finally, each Z s \mathcal{Z}_s Z s is a domain and the displayed product defining Δ s \Delta_s Δ s is not identically zero. Thus Δ L \Delta_L Δ L is a non-zero-divisor in the product ring Z L \mathcal{Z}_L Z L . This proves all the assertions.
Corollary 5.10. Let y ∈ GL ( N , F ) y \in\operatorname{GL}(N,F) y ∈ GL ( N , F ) be regular semisimple and let μ y \mu_y μ y be an invariant measure on its conjugacy class. Then
Δ G ∗ μ y ∈ Z G . \Delta_G * \mu_y \in\mathcal{Z}_G. Δ G ∗ μ y ∈ Z G .
For every Bernstein component s s s there is a regular element A y , s ∈ Z s A_{y,s} \in\mathcal{Z}_s A y , s ∈ Z s such that
μ y ∗ ( Δ s ∗ f ) = A y , s ∗ f , f ∈ ( H G ) s . \mu_y * (\Delta_s * f) = A_{y,s} * f, \qquad f \in(\mathcal{H}_G)_s. μ y ∗ ( Δ s ∗ f ) = A y , s ∗ f , f ∈ ( H G ) s .
Moreover, Δ G \Delta_G Δ G is nonzero at every irreducible representation of G ( F ) G(F) G ( F ) that is square-integrable modulo the center.
Proof. The orbital measure belongs to D G c \mathcal{D}_G^c D G c , so the compactness assertion follows from the theorem. For the final assertion, such a representation is a generalized Steinberg representation St ( σ , m ) \operatorname{St}(\sigma,m) St ( σ , m ) , with N = d m N = dm N = d m . For a unitary σ \sigma σ its cuspidal support has twist coordinates
z j = u Q ( m + 1 − 2 j ) / 2 , Q = q t σ > 1 , ∣ u ∣ = 1. z_j = uQ^{(m+1-2j)/2}, \qquad Q = q^{t_\sigma} > 1, \qquad|u| = 1. z j = u Q ( m + 1 − 2 j ) /2 , Q = q t σ > 1 , ∣ u ∣ = 1.
Consequently
Δ G ( St ( σ , m ) ) = ∏ i ≠ j ( 1 − Q j − i ) ≠ 0. \Delta_G(\operatorname{St}(\sigma,m)) = \prod_{i \ne j}(1-Q^{j-i}) \ne0. Δ G ( St ( σ , m )) = i = j ∏ ( 1 − Q j − i ) = 0.
Unramified twists multiply all these coordinates by the same scalar and leave the product unchanged.
Necessity of the discriminant
Completion of the proof of Theorem 1.4 . We give the proof for G = GL ( N ) G=\operatorname{GL}(N) G = GL ( N ) ; for a product, the same construction is performed in each factor. Fix a Bernstein component s \mathfrak{s} s with cuspidal support Levi
M = ∏ a GL ( d a ) m a , σ = ⨂ a σ a ⊗ m a . M=\prod_a \operatorname{GL}(d_a)^{m_a},\qquad\sigma=\bigotimes_a\sigma_a^{\otimes m_a}. M = a ∏ GL ( d a ) m a , σ = a ⨂ σ a ⊗ m a .
Let t = [ M ( F ) , σ ] M \mathfrak{t}=[M(F),\sigma]_M t = [ M ( F ) , σ ] M be the corresponding cuspidal component of M ( F ) M(F) M ( F ) , and let e t ∈ Z M e_{\mathfrak{t}}\in\mathcal{Z}_M e t ∈ Z M be its Bernstein projector. Define the invariant distribution induced from this projector by
α s ( h ) = e t ( r ‾ M G h ) . \alpha_{\mathfrak{s}}(h)=e_{\mathfrak{t}}(\overline{r}^{G}_{M}h). α s ( h ) = e t ( r M G h ) .
The projector e t e_{\mathfrak{t}} e t is supported on compact elements of M ( F ) M(F) M ( F ) ; see [2 Theorem 1.3 and Section 1.6] . In the constant-term formula for r ‾ M G \overline{r}^{G}_{M} r M G , an element m u mu m u has the same eigenvalues as its block-diagonal part m m m . Hence a test function vanishing on B 0 ( G ( F ) ) B_0(G(F)) B 0 ( G ( F )) has constant term vanishing on B 0 ( M ( F ) ) B_0(M(F)) B 0 ( M ( F )) . It follows that α s \alpha_{\mathfrak{s}} α s is supported in B 0 ( G ( F ) ) B_0(G(F)) B 0 ( G ( F )) . This set is bounded modulo conjugation, as proved after (11). Thus α s ∈ D G c \alpha_{\mathfrak{s}}\in\mathcal{D}^{c}_{G} α s ∈ D G c .
On t \mathfrak{t} t one has Δ M = 1 \Delta_M=1 Δ M = 1 and W t = 1 W_{\mathfrak{t}}=1 W t = 1 . Put
P s = ∏ a ∏ i ≠ j ( 1 − Q a − 1 z a , i z a , j ) , Q a = q t a > 1. P_{\mathfrak{s}}=\prod_a\prod_{i\ne j}\left(1-Q_a^{-1}\frac{z_{a,i}}{z_{a,j}}\right),\qquad Q_a=q^{t_a}>1. P s = a ∏ i = j ∏ ( 1 − Q a − 1 z a , j z a , i ) , Q a = q t a > 1.
Formula (28), with a = e t a=e_{\mathfrak{t}} a = e t , therefore gives
Δ G ∗ α s = c s P s , c s ≠ 0 , as invariant distributions, (30) \Delta_G*\alpha_{\mathfrak{s}}=c_{\mathfrak{s}}P_{\mathfrak{s}},\qquad c_{\mathfrak{s}}\ne0,\qquad\text{as invariant distributions,} \tag*{(30)} Δ G ∗ α s = c s P s , c s = 0 , as invariant distributions, (30)
where the central element on the right is extended by zero off s \mathfrak{s} s . The constant includes ∣ W s ∣ |W_{\mathfrak{s}}| ∣ W s ∣ , since P s P_{\mathfrak{s}} P s is W s W_{\mathfrak{s}} W s -invariant.
Now let z ∈ J G z\in\mathcal{J}_G z ∈ J G . By definition, b = z ∗ α s ∈ Z G b=z*\alpha_{\mathfrak{s}}\in\mathcal{Z}_G b = z ∗ α s ∈ Z G . Commutativity of the central action and (30) imply, on the component s \mathfrak{s} s ,
Δ s b s = c s z s P s . (31) \Delta_{\mathfrak{s}}b_{\mathfrak{s}}=c_{\mathfrak{s}}z_{\mathfrak{s}}P_{\mathfrak{s}}. \tag*{(31)} Δ s b s = c s z s P s . (31)
The ring Z s \mathcal{Z}_{\mathfrak{s}} Z s is a Laurent polynomial ring in elementary symmetric functions, and in particular is a unique factorization domain. The discriminant Δ s \Delta_{\mathfrak{s}} Δ s is squarefree in this ring; its prime factors are the reduced collision divisors, one for each a a a with m a ≥ 2 m_a\ge2 m a ≥ 2 (see Proposition 6.1). At the generic point of any such divisor, precisely one pair of coordinates coincides. Its two factors in P s P_{\mathfrak{s}} P s are 1 − Q a − 1 1-Q_a^{-1} 1 − Q a − 1 , which is nonzero, and the other factors are generically nonzero. Thus P s P_{\mathfrak{s}} P s is relatively prime to Δ s \Delta_{\mathfrak{s}} Δ s . Equation (31) gives Δ s ∣ z s \Delta_{\mathfrak{s}}\mid z_{\mathfrak{s}} Δ s ∣ z s .
This holds on every component. The quotients ( z s / Δ s ) s (z_{\mathfrak{s}}/\Delta_{\mathfrak{s}})_{\mathfrak{s}} ( z s / Δ s ) s define an element of Z G = ∏ s Z s \mathcal{Z}_G=\prod_{\mathfrak{s}}\mathcal{Z}_{\mathfrak{s}} Z G = ∏ s Z s , so z ∈ Δ G Z G z\in\Delta_G\mathcal{Z}_G z ∈ Δ G Z G . The reverse inclusion is Theorem 5.2. For products, take M M M , e t e_{\mathfrak{t}} e t , and P s P_{\mathfrak{s}} P s as the corresponding products; the same coprimality and divisibility argument applies. □
The endoscopic intersection for general linear groups
The reduced collision locus in type A
For a Bernstein component as in (16), put
T s = ∏ a ( C × ) m a , W s = ∏ a S m a , X s = T s / W s . T_{\mathfrak{s}}=\prod_a(\mathbb{C}^{\times})^{m_a},\qquad W_{\mathfrak{s}}=\prod_aS_{m_a},\qquad X_{\mathfrak{s}}=T_{\mathfrak{s}}/W_{\mathfrak{s}}. T s = a ∏ ( C × ) m a , W s = a ∏ S m a , X s = T s / W s .
Let D s ⊂ X s D_{\mathfrak{s}}\subset X_{\mathfrak{s}} D s ⊂ X s be the reduced image of the union of the hyperplanes z a , i = z a , j z_{a,i}=z_{a,j} z a , i = z a , j , i ≠ j i\ne j i = j .
Proposition 6.1. The ideal of D s D_{\mathfrak{s}} D s is the principal ideal ( Δ s ) (\Delta_{\mathfrak{s}}) ( Δ s ) . In particular, the common denominator of Theorem 5.2 defines the reduced collision divisor on the quotient, although its pullback has order two on every collision hyperplane of the ordered torus.
Proof. Work first in one symmetric group factor. If an invariant Laurent polynomial f f f vanishes on z i = z j z_i=z_j z i = z j , it is divisible by z i − z j z_i-z_j z i − z j . Write f = ( z i − z j ) g f=(z_i-z_j)g f = ( z i − z j ) g . Invariance under the transposition $(ij) makes g g g anti-invariant. Its restriction to z i = z j z_i=z_j z i = z j is therefore zero, so $f is divisible by ( z i − z j ) 2 (z_i-z_j)^2 ( z i − z j ) 2 . Applying this to every pair and using unique factorization in the Laurent polynomial ring gives
∏ i < j ( z i − z j ) 2 ∣ f . \prod_{i<j}(z_i-z_j)^2 \mid f. i < j ∏ ( z i − z j ) 2 ∣ f .
Up to an invertible symmetric Laurent monomial and a nonzero constant, this product is ∏ i ≠ j ( 1 − z i / z j ) \prod_{i\ne j}(1-z_i/z_j) ∏ i = j ( 1 − z i / z j ) . The quotient by this product is again invariant. Repeating the argument for each a a a proves the assertion. □
Under local Langlands for G L ( N ) \mathrm{GL}(N) GL ( N ) , the points on this divisor have repeated isomorphic irreducible Weil constituents in their semisimple parameters. If two such constituents are placed in different blocks of an endoscopic Levi, their off-diagonal intertwining spaces increase the centralizer dimension. Thus the divisor has the centralizer-jump interpretation motivating § 4 .
The Fitting ideals for endoscopic Levis
For G L ( N ) \mathrm{GL}(N) GL ( N ) , ordinary untwisted endoscopic transfers are the standard block-sum transfers from products of general linear groups, up to equivalence and changes of source coordinates. Indeed, in the untwisted case the Weil action fixes the semisimple endoscopic element; its centralizer is a product of general linear groups, and the action on that product is inner. This assertion would not apply to endoscopy with a nontrivial twisting character.
Let H = ∏ b G L ( n b ) H=\prod_b\mathrm{GL}(n_b) H = ∏ b GL ( n b ) and fix a source component mapping to X s X_s X s . Assign each irreducible Weil constituent indexed by ( a , i ) (a,i) ( a , i ) to its H H H -block b ( a , i ) b(a,i) b ( a , i ) . This describes the source quotient as T s / W t T_s/W_t T s / W t , where W t = ∏ a , b S m a , b W_t=\prod_{a,b}S_{m_{a,b}} W t = ∏ a , b S m a , b . Its map to X s X_s X s is finite. For fixed s s s only finitely many block partitions and assignments occur. The same description applies on each bounded parameter piece. Under the local Langlands correspondence, these quotients are the corresponding Bernstein components. The map Ψ G \Psi_G Ψ G is this identification, by [3 Theorem IX.7.4] and the characteristic-zero description of the invariant parameter rings in [8 Theorem 1.7 and Corollary 1.8] . The latter is an isomorphism of schemes in characteristic zero, so this identification is not limited to closed geometric points.
Define on the source component
C s / t = ∏ a ∏ i , j b ( a , i ) ≠ b ( a , j ) ( 1 − z a , i z a , j ) . (32) C_{s/t}=\prod_a\prod_{\substack{i,j\\b(a,i)\ne b(a,j)}}\left(1-\frac{z_{a,i}}{z_{a,j}}\right). \tag*{(32)} C s / t = a ∏ i , j b ( a , i ) = b ( a , j ) ∏ ( 1 − z a , j z a , i ) . (32)
Both orientations of every pair occur. With the compatible coordinates of Section 5.7, this is the restriction of the analytic cross factor C M C_M C M in (25) to the source component. The internal-block discriminant Δ t \Delta_t Δ t satisfies
f ∗ Δ s = Δ t C s / t . (33) f^*\Delta_s=\Delta_t C_{s/t}. \tag*{(33)} f ∗ Δ s = Δ t C s / t . (33)
Lemma 6.2. On the parameter scheme over this source component, the pullback of C s / t C_{s/t} C s / t belongs to the zeroth Fitting ideal of the universal coinvariant module for V = g l N / ⨁ b g l n b V=\mathfrak{gl}_N/\bigoplus_b\mathfrak{gl}_{n_b} V = gl N / ⨁ b gl n b . Consequently f ∗ Δ s f^*\Delta_s f ∗ Δ s belongs to the contracted ideal J e , U \mathcal{J}_{e,U} J e , U on this component.
Proof. We work in characteristic zero. Locally on fixed-inertial-type charts, write the universal parameter in Weil–Deligne form ( r , N ) (r,N) ( r , N ) , with r r r having finite inertia image and r ( F r ) = φ ( F r ) r(\mathrm{Fr})=\varphi(\mathrm{Fr}) r ( Fr ) = φ ( Fr ) , where φ \varphi φ is the universal Weil parameter. These are scheme-theoretic charts: see [8 Propositions 2.6–2.7, Section 3, and (4.2)–(4.5)] . In particular, the argument below applies to the universal family, not only to geometric points. After a faithfully flat change of charts, we may fix the finite inertia action. Let R R R be the chart ring and put
W = ( R ⊗ V ∨ ) r ( I F ) . W=(R\otimes V^{\vee})_{r(I_F)}. W = ( R ⊗ V ∨ ) r ( I F ) .
Taking coinvariants for a finite group in characteristic zero is a direct-summand operation, so W W W is locally free. Frobenius acts on W W W . Let N V \mathcal{N}_V N V be the nilpotent operator induced by monodromy on R ⊗ V ∨ R \otimes V^\vee R ⊗ V ∨ . A suitable power of a tame inertia generator acts as exp ( c N V ) \exp(c\mathcal{N}_V) exp ( c N V ) for some nonzero scalar c c c . Since
exp ( c N V ) − 1 = N V u ( N V ) , u ( 0 ) = c ≠ 0 , \exp(c\mathcal{N}_V)-1=\mathcal{N}_V u(\mathcal{N}_V), \qquad u(0)=c\ne0, exp ( c N V ) − 1 = N V u ( N V ) , u ( 0 ) = c = 0 ,
the operator u ( N V ) u(\mathcal{N}_V) u ( N V ) is invertible. Thus its coinvariant relation imposes the N V \mathcal{N}_V N V -relation. The original inertia relations then impose the finite-inertia relations. Consequently the full coinvariant module M M M is obtained by imposing the monodromy relations on
coker ( 1 − r ( Fr ) : W ⟶ W ) . \operatorname{coker}(1-r(\operatorname{Fr}):W\longrightarrow W). coker ( 1 − r ( Fr ) : W ⟶ W ) .
In particular, M M M is a quotient of this cokernel. Zeroth Fitting ideals increase under taking quotients. Therefore
det ( 1 − r ( Fr ) ∣ W ) ∈ Fitt 0 R ( M ) . (34) \det(1-r(\operatorname{Fr})\mid W)\in\operatorname{Fitt}^{R}_{0}(M). \tag*{(34)} det ( 1 − r ( Fr ) ∣ W ) ∈ Fitt 0 R ( M ) . (34)
We compute this determinant using the irreducible constituents of the semisimplification of r r r . Finite-inertia coinvariants are exact, and determinants are multiplicative along invariant filtrations, so this calculation also applies when Frobenius is not semisimple. Two irreducible Weil representations have a nonzero inertia intertwiner precisely when they differ by an unramified twist. For the class indexed by a a a , the restriction of its representative to inertia has t a t_a t a pairwise nonisomorphic constituents, cyclically permuted by Frobenius. For two twists in this class, Frobenius acts cyclically on the corresponding t a t_a t a -dimensional inertia-intertwiner space, with cycle product z a , i / z a , j z_{a,i}/z_{a,j} z a , i / z a , j . Its contribution to det ( 1 − r ( Fr ) ) \det(1-r(\operatorname{Fr})) det ( 1 − r ( Fr )) is therefore 1 − z a , i / z a , j 1-z_{a,i}/z_{a,j} 1 − z a , i / z a , j , with the orientation reversed on the dual. The module V V V consists exactly of the cross-block Hom \operatorname{Hom} Hom spaces, and both orientations occur. Thus the determinant in (34) is C s / t C_s/t C s / t , up to a unit. This calculation is an identity in the chart ring: it is the characteristic-polynomial identity for the finite inertia types, not merely an equality on a generic open subset.
Fitting ideals commute with base change, and ideal membership descends under faithfully flat maps. This proves the first assertion on the parameter scheme. Multiplying by Δ t \Delta_t Δ t and using (33) proves the second assertion after contraction to the invariant ring. In particular, monodromy directions do not invalidate the containment. □
Computation of the intersection
Theorem 6.3. For G = GL ( N ) G=\operatorname{GL}(N) G = GL ( N ) and every Bernstein component s s s , the endoscopic intersection, under the spectral-to-Bernstein identification, is
I s ∩ = ( Δ s ) . \mathcal{I}^{\cap}_s=(\Delta_s). I s ∩ = ( Δ s ) . Globally,
Z G ⋅ Ψ G ( I G ∩ ) = Δ G Z G = J G . Z_G\cdot\Psi_G(\mathcal{I}^{\cap}_G)=\Delta_G Z_G=J_G. Z G ⋅ Ψ G ( I G ∩ ) = Δ G Z G = J G . Thus Conjecture 1.5 holds for GL ( N ) \operatorname{GL}(N) GL ( N ) .
Proof. Lemma 6.2 shows that Δ s \Delta_s Δ s belongs to every endoscopic ideal on this component. Hence
( Δ s ) ⊂ I s ∩ . (\Delta_s)\subset\mathcal{I}^{\cap}_s. ( Δ s ) ⊂ I s ∩ .
For the reverse inclusion, use the finest support Levi
H s = ∏ a GL ( d a ) m a H_s=\prod_a \operatorname{GL}(d_a)^{m_a} H s = a ∏ GL ( d a ) m a
and the source component in which each factor has one irreducible Weil constituent. Its coarse parameter space is the ordered torus T s T_s T s . One may take an additional finite étale cover
T ~ s ⟶ T s \widetilde{T}_s\longrightarrow T_s T s ⟶ T s
to choose the individual unramified twists before quotienting by their self-twists. On this cover there is an actual family of block-diagonal Weil parameters with these irreducible constituents and with N = 0 \mathcal{N}=0 N = 0 .
For this family there are no monodromy relations and no repeated constituents inside any one block. The universal coinvariant module is exactly the cokernel of 1 − r ( Fr ) 1-r(\operatorname{Fr}) 1 − r ( Fr ) on the finite-inertia coinvariants. Its zeroth Fitting ideal on T ~ s \widetilde{T}_s T s is consequently
( Δ s ) O ( T ~ s ) . (\Delta_s)\mathcal{O}(\widetilde{T}_s). ( Δ s ) O ( T s ) .
This is a calculation of the full ideal, including at multiple collisions. If a ∈ I s Π a \in\mathcal{I}^{\Pi}_{s} a ∈ I s Π , its pullback to this family must lie in that ideal, since Fitting ideals commute with base change. The map T ~ s → X s \widetilde{T}_{s} \to X_{s} T s → X s is faithfully flat: the self-twist power maps are finite étale, and the symmetric Laurent-polynomial extension for each S m a S_{m_a} S m a is finite free. Contraction of a principal extended ideal therefore gives
( Δ s ) O ( T ~ s ) ∩ O ( X s ) = ( Δ s ) . (\Delta_s)\mathcal{O}(\widetilde{T}_s) \cap\mathcal{O}(X_s) = (\Delta_s). ( Δ s ) O ( T s ) ∩ O ( X s ) = ( Δ s ) .
It follows that a ∈ ( Δ s ) a \in(\Delta_s) a ∈ ( Δ s ) . If the support Levi is G G G itself, then Δ s = 1 \Delta_s = 1 Δ s = 1 , and the first inclusion already gives the assertion without using a proper transfer.
The formulas use only the multiset of irreducible constituents, are independent of coordinate representatives, and agree upon changing ramification bounds. They therefore compute the global ideal defined by (10) as the product of these component ideals. The spectral-to-Bernstein identification and Theorem 1.4 give the global equality. Finally Δ G \Delta_G Δ G is a non-zero-divisor, as in Theorem 5.2. □
The conjectures for SL(2)
Put G = S L ( 2 ) G = \mathrm{SL}(2) G = SL ( 2 ) and G ~ = G L ( 2 ) \widetilde{G} = \mathrm{GL}(2) G = GL ( 2 ) . We prove both Conjectures 1.1 and 1.5 for G G G . The additional phenomenon, compared with G L ( 2 ) \mathrm{GL}(2) GL ( 2 ) , is the splitting of a principal series at a nontrivial quadratic character. Differences of the two constituent characters force additional zeros in the denominator. The nonsplit endoscopic tori account for these zeros.
The statement and its coordinates
Identify the diagonal torus T ( F ) T(F) T ( F ) with F × F^\times F × by a ↦ diag ( a , a − 1 ) a \mapsto\operatorname{diag}(a,a^{-1}) a ↦ diag ( a , a − 1 ) , and write I ( χ ) I(\chi) I ( χ ) for normalized induction of a smooth character χ \chi χ of T ( F ) T(F) T ( F ) to G ( F ) G(F) G ( F ) . A principal series Bernstein component is specified by ϑ = χ ∣ O F × \vartheta= \chi|_{\mathcal{O}_F^\times} ϑ = χ ∣ O F × , up to inversion. Choose an extension χ 0 \chi_0 χ 0 of ϑ \vartheta ϑ and write χ = χ 0 ν z \chi= \chi_0\nu_z χ = χ 0 ν z , where ν z \nu_z ν z is unramified and ν z ( ϖ ) = z \nu_z(\varpi) = z ν z ( ϖ ) = z . If ϑ 2 = 1 \vartheta^2 = 1 ϑ 2 = 1 , choose χ 0 ( ϖ ) = 1 \chi_0(\varpi) = 1 χ 0 ( ϖ ) = 1 , so that χ 0 2 = 1 \chi_0^2 = 1 χ 0 2 = 1 and inversion acts by z ↦ z − 1 z \mapsto z^{-1} z ↦ z − 1 . The component rings are
Z s = { C [ z ± 1 ] , ϑ 2 ≠ 1 , C [ z ± 1 ] z ↦ z − 1 = C [ t ] , t = z + z − 1 , ϑ 2 = 1. Z_s =
\begin{cases}
\mathbb{C}[z^{\pm1}], & \vartheta^2 \ne1,\\
\mathbb{C}[z^{\pm1}]_{z\mapsto z^{-1}} = \mathbb{C}[t],\quad t = z + z^{-1}, & \vartheta^2 = 1.
\end{cases} Z s = { C [ z ± 1 ] , C [ z ± 1 ] z ↦ z − 1 = C [ t ] , t = z + z − 1 , ϑ 2 = 1 , ϑ 2 = 1.
A cuspidal Bernstein component is a point, since G ( F ) G(F) G ( F ) has no nontrivial unramified characters. Define d ∈ Z G d \in Z_G d ∈ Z G componentwise by
d s = { 1 , s is cuspidal , 1 , s is principal series and ϑ 2 ≠ 1 , t 2 − 4 , s is principal series and ϑ 2 = 1. (35) d_s =
\begin{cases}
1, & s\text{ is cuspidal},\\
1, & s\text{ is principal series and }\vartheta^2 \ne1,\\
t^2 - 4, & s\text{ is principal series and }\vartheta^2 = 1.
\end{cases} \tag*{(35)} d s = ⎩ ⎨ ⎧ 1 , 1 , t 2 − 4 , s is cuspidal , s is principal series and ϑ 2 = 1 , s is principal series and ϑ 2 = 1. (35)
Changing the quadratic extension χ 0 \chi_0 χ 0 replaces t t t by − t -t − t , so the last formula is independent of that choice.
Theorem 7.1. For G = S L ( 2 ) G = \mathrm{SL}(2) G = SL ( 2 ) ,
J G = d Z G = Z G ⋅ Ψ G ( G ~ ) . J_G = dZ_G = Z_G \cdot\Psi_G(\widetilde{G}). J G = d Z G = Z G ⋅ Ψ G ( G ) .
The element d d d is a non-zero-divisor. In particular, d ∗ D G c ⊂ Z G d * D^c_G \subset Z_G d ∗ D G c ⊂ Z G and D G c ⊂ R G D^c_G \subset R_G D G c ⊂ R G .
We use the standard restriction theory from G L ( 2 , F ) \mathrm{GL}(2,F) GL ( 2 , F ) to S L ( 2 , F ) \mathrm{SL}(2,F) SL ( 2 , F ) : restriction of an irreducible representation is a finite sum of irreducibles, its constituents form a single G L ( 2 , F ) \mathrm{GL}(2,F) GL ( 2 , F ) -orbit, and every irreducible representation of S L ( 2 , F ) \mathrm{SL}(2,F) SL ( 2 , F ) occurs this way. Cuspidality is preserved under restriction. See [1 Section 2 and Theorem 3.1] . In the principal series, I ( χ ) I(\chi) I ( χ ) is irreducible except when χ = ∣ ⋅ ∣ ± 1 \chi= \lvert\cdot\rvert^{\pm1} χ = ∣ ⋅ ∣ ± 1 or χ \chi χ is nontrivial quadratic. In the latter case
I ( χ ) = π χ + ⊕ π χ − , I(\chi) = \pi_\chi^+ \oplus\pi_\chi^-, I ( χ ) = π χ + ⊕ π χ − ,
with the two inequivalent constituents interchanged by G L ( 2 , F ) \mathrm{GL}(2,F) GL ( 2 , F ) -conjugation. At ∣ ⋅ ∣ ± 1 \lvert\cdot\rvert^{\pm1} ∣ ⋅ ∣ ± 1 the constituents are the trivial and Steinberg representations, each fixed by this conjugation action. These assertions also follow directly from the restriction theory: an irreducible principal series of GL(2, F F F ) has a nontrivial determinant self-twist precisely when its character ratio is nontrivial quadratic; see [1 Sections 2–3] .
Descent of the GL(2) denominator
Let j : G ( F ) ↪ G ~ ( F ) j:G(F)\hookrightarrow\widetilde{G}(F) j : G ( F ) ↪ G ( F ) be the inclusion. Define d 0 ∈ Z G d_0\in Z_G d 0 ∈ Z G to be 2 − t = ( 1 − z ) ( 1 − z − 1 ) 2-t=(1-z)(1-z^{-1}) 2 − t = ( 1 − z ) ( 1 − z − 1 ) on the unramified principal series component and 1 1 1 on every other component. It is invariant under G ~ ( F ) \widetilde{G}(F) G ( F ) -conjugation.
Lemma 7.2. If α ∈ D G c \alpha\in D_G^c α ∈ D G c is invariant under G ~ ( F ) \widetilde{G}(F) G ( F ) , then d 0 ∗ α ∈ Z G d_0\ast\alpha\in Z_G d 0 ∗ α ∈ Z G .
Proof. For a distribution γ \gamma γ on G ( F ) G(F) G ( F ) let j ∗ γ ( f ) = γ ( f ∣ G ( F ) ) j_\ast\gamma(f)=\gamma(f|_{G(F)}) j ∗ γ ( f ) = γ ( f ∣ G ( F ) ) . An outer-invariant central distribution b ∈ Z G b\in Z_G b ∈ Z G gives a central distribution j ∗ b ∈ Z G ~ j_\ast b\in Z_{\widetilde{G}} j ∗ b ∈ Z G . To check essential compactness, use the homeomorphism
F × × G ( F ) ⟶ G ~ ( F ) , ( a , h ) ⟼ s ( a ) h , s ( a ) = diag ( a , 1 ) . F^\times\times G(F)\longrightarrow\widetilde{G}(F),\qquad(a,h)\longmapsto s(a)h,\qquad s(a)=\operatorname{diag}(a,1). F × × G ( F ) ⟶ G ( F ) , ( a , h ) ⟼ s ( a ) h , s ( a ) = diag ( a , 1 ) . Every compactly supported locally constant function in these coordinates is a finite sum of functions ϕ ( a ) f 0 ( h ) \phi(a)f_0(h) ϕ ( a ) f 0 ( h ) . Convolution by j ∗ b j_\ast b j ∗ b sends such a function to ϕ ( a ) ( b ∗ f 0 ) ( h ) \phi(a)(b\ast f_0)(h) ϕ ( a ) ( b ∗ f 0 ) ( h ) , because b b b is invariant under conjugation by s ( a ) s(a) s ( a ) . This is compactly supported. The same observation shows conversely that, for an outer-invariant distribution γ \gamma γ , essential compactness of j ∗ γ j_\ast\gamma j ∗ γ implies essential compactness of γ \gamma γ : take ϕ ( 1 ) ≠ 0 \phi(1)\ne0 ϕ ( 1 ) = 0 and restrict the convolution to the fiber a = 1 a=1 a = 1 .
The central element j ∗ d 0 j_\ast d_0 j ∗ d 0 is precisely Δ G ~ \Delta_{\widetilde{G}} Δ G of (17) . Indeed, on a principal series of G ~ ( F ) \widetilde{G}(F) G ( F ) with inducing characters χ 1 , χ 2 \chi_1,\chi_2 χ 1 , χ 2 , restriction to G ( F ) G(F) G ( F ) has character ratio χ 1 χ 2 − 1 \chi_1\chi_2^{-1} χ 1 χ 2 − 1 . The GL(2) discriminant is ( 1 − z ) ( 1 − z − 1 ) (1-z)(1-z^{-1}) ( 1 − z ) ( 1 − z − 1 ) when this ratio is unramified, and is 1 1 1 otherwise. On cuspidal representations both elements act by 1 1 1 . The same scalar comparison holds on every constituent at reducibility points, and hence identifies the central elements.
Now j ∗ α j_\ast\alpha j ∗ α is invariant under G ~ ( F ) \widetilde{G}(F) G ( F ) and has bounded support modulo conjugation there. Theorem 5.2 gives
j ∗ ( d 0 ∗ α ) = Δ G ~ ∗ j ∗ α ∈ Z G ~ . j_\ast(d_0\ast\alpha)=\Delta_{\widetilde{G}}\ast j_\ast\alpha\in Z_{\widetilde{G}}. j ∗ ( d 0 ∗ α ) = Δ G ∗ j ∗ α ∈ Z G . The converse compactness observation proves the lemma. □
The part not invariant under GL(2, F)
The conjugation action on invariant distributions and on the cocenter of G ( F ) G(F) G ( F ) factors through
C = G ~ ( F ) / ( Z G ~ ( F ) G ( F ) ) ≃ F × / ( F × ) 2 . C=\widetilde{G}(F)/(Z_{\widetilde{G}}(F)G(F))\simeq F^\times/(F^\times)^2. C = G ( F ) / ( Z G ( F ) G ( F )) ≃ F × / ( F × ) 2 .
This is a compact group. It is finite if char F ≠ 2 \operatorname{char}F\ne2 char F = 2 ; in characteristic 2 we use normalized Haar averaging on C C C . The action on every test function modulo commutators is smooth, so averaging is well-defined. Write Av for this projector, on either cocenters or invariant distributions. It preserves bounded support modulo conjugation. Indeed, C C C has representatives s ( a ) s(a) s ( a ) with a ∈ O F × ∪ ϖ O F × a\in O_F^\times\cup\varpi O_F^\times a ∈ O F × ∪ ϖ O F × , a compact set, and conjugating a fixed compact support set by these representatives again gives a compact set.
For nontrivial quadratic χ \chi χ , put
η χ ( h ) = tr π χ + ( h ∨ ) − tr π χ − ( h ∨ ) . \eta_\chi(h)=\operatorname{tr}\pi_\chi^+(h^\vee)-\operatorname{tr}\pi_\chi^-(h^\vee). η χ ( h ) = tr π χ + ( h ∨ ) − tr π χ − ( h ∨ ) .
Let x χ x_\chi x χ be its common central character in the principal series Bernstein component. Then u ∗ η χ = u ( x χ ) η χ u\ast\eta_\chi=u(x_\chi)\eta_\chi u ∗ η χ = u ( x χ ) η χ for u ∈ Z G u\in Z_G u ∈ Z G .
Lemma 7.3. For any invariant distribution β \beta β with Av ( β ) = 0 \operatorname{Av}(\beta)=0 Av ( β ) = 0 , its projection to a principal series component is a linear combination of the distributions η χ \eta_\chi η χ in that component. Its projections to the cuspidal components together form an element of Z G Z_G Z G .
Proof. Apply the trace description to ℓ β ( h ) = β ( h ∨ ) \ell_\beta(h)=\beta(h^\vee) ℓ β ( h ) = β ( h ∨ ) . On a principal series component consider the image of 1 − Av 1-\operatorname{Av} 1 − Av in the cocenter. The character of any G ~ ( F ) \widetilde{G}(F) G ( F ) -fixed irreducible representation vanishes on this image. By the preceding classification, only the two constituents at each nontrivial quadratic character remain, and their traces on this image are opposite. There are at most two such characters in any one component. More explicitly, let
H ‾ G , s \overline{\mathcal{H}}_{G,s} H G , s denote the image of the Hecke algebra of the component s \mathfrak{s} s in H ‾ G \overline{\mathcal{H}}_G H G . Spectral density [10 Theorem 2.8] gives an injective map
( 1 − Av ) H ‾ G , s ⟶ ⨁ χ 2 = 1 , χ ≠ 1 χ ∈ s C , (1-\operatorname{Av})\overline{\mathcal{H}}_{G,\mathfrak{s}} \longrightarrow\bigoplus_{\substack{\chi^2=1,\ \chi\ne1\\ \chi\in\mathfrak{s}}}\mathbb{C}, ( 1 − Av ) H G , s ⟶ χ 2 = 1 , χ = 1 χ ∈ s ⨁ C ,
[ h ] ⟼ ( tr π χ + ( h ) − tr π χ − ( h ) ) χ . [h]\longmapsto\left(\operatorname{tr}\pi_\chi^+(h)-\operatorname{tr}\pi_\chi^-(h)\right)_\chi. [ h ] ⟼ ( tr π χ + ( h ) − tr π χ − ( h ) ) χ .
Here χ ∈ s \chi\in\mathfrak{s} χ ∈ s means that the cuspidal support of I ( χ ) I(\chi) I ( χ ) belongs to s \mathfrak{s} s . Indeed, if these differences vanish, then all irreducible traces vanish on the given class, so density makes that class zero. Every functional on the image of this map extends to its finite-dimensional target. Dualizing therefore gives the assertion for ℓ β \ell_\beta ℓ β , and inversion gives the assertion for β \beta β .
Each cuspidal component contains a single irreducible representation π \pi π . Density makes its cocenter one-dimensional. By Plancherel, the Bernstein projector e π e_\pi e π , regarded as a distribution, satisfies
e π ( h ) = d π tr π ( h ∨ ) , e_\pi(h)=d_\pi\operatorname{tr}\pi(h^\vee), e π ( h ) = d π tr π ( h ∨ ) ,
where d π d_\pi d π is the nonzero formal degree of π \pi π . Thus every invariant distribution on this component is a scalar multiple of e π e_\pi e π and belongs to Z G \mathcal{Z}_G Z G . Arbitrary scalars on the cuspidal components define an element of the product Bernstein center; only finitely many components are visible on any given compact-open level. This proves the last assertion. □
Define r ∈ Z G r\in\mathcal{Z}_G r ∈ Z G by
r s = { t + 2 , ϑ = 1 , t 2 − 4 , ϑ ≠ 1 , ϑ 2 = 1 , 1 , on all other components . r_{\mathfrak{s}}=
\begin{cases}
t+2, & \vartheta=1,\\
t^2-4, & \vartheta\ne1,\ \vartheta^2=1,\\
1, & \text{on all other components}.
\end{cases} r s = ⎩ ⎨ ⎧ t + 2 , t 2 − 4 , 1 , ϑ = 1 , ϑ = 1 , ϑ 2 = 1 , on all other components .
It vanishes at every x χ x_\chi x χ with χ \chi χ nontrivial quadratic. Consequently Lemma 7.3 gives r ∗ ( α − Av ( α ) ) ∈ Z G r*(\alpha-\operatorname{Av}(\alpha))\in\mathcal{Z}_G r ∗ ( α − Av ( α )) ∈ Z G for every invariant distribution α \alpha α . If α ∈ D G c \alpha\in\mathcal{D}^c_G α ∈ D G c , Lemma 7.2 also gives d 0 ∗ Av ( α ) ∈ Z G d_0*\operatorname{Av}(\alpha)\in\mathcal{Z}_G d 0 ∗ Av ( α ) ∈ Z G . Combining these,
d 0 r ∗ α ∈ Z G . d_0r*\alpha\in\mathcal{Z}_G. d 0 r ∗ α ∈ Z G .
The element d 0 r d_0r d 0 r differs from d d d only by the unit − 1 -1 − 1 on the unramified principal series component. Thus
d Z G ⊂ J G . (36) d\mathcal{Z}_G\subset\mathcal{J}_G. \tag*{(36)} d Z G ⊂ J G . (36)
Every component of d d d is nonzero, so this already proves Conjecture 1.1 for SL ( 2 ) \operatorname{SL}(2) SL ( 2 ) .
Necessity of the factors
First, η χ ∈ D G c \eta_\chi\in\mathcal{D}^c_G η χ ∈ D G c . To see this, on the split regular semisimple set every GL ( 2 , F ) \operatorname{GL}(2,F) GL ( 2 , F ) -conjugacy class is one SL ( 2 , F ) \operatorname{SL}(2,F) SL ( 2 , F ) -conjugacy class: the determinant map on the split centralizer in GL ( 2 , F ) \operatorname{GL}(2,F) GL ( 2 , F ) is surjective. The two constituent characters therefore agree there. In particular, their difference vanishes outside the set
B 0 ( G ( F ) ) = { g ∈ G ( F ) : both eigenvalues have absolute value 1 } . B_0(G(F))=\{g\in G(F):\text{ both eigenvalues have absolute value }1\}. B 0 ( G ( F )) = { g ∈ G ( F ) : both eigenvalues have absolute value 1 } .
Indeed, every element outside this set is split regular semisimple. This set is bounded modulo G ( F ) G(F) G ( F ) -conjugation. Rational canonical form first proves boundedness modulo GL ( 2 , F ) \operatorname{GL}(2,F) GL ( 2 , F ) -conjugation, and the compact representatives of C C C used above convert this into boundedness modulo G ( F ) G(F) G ( F ) -conjugation.
The nonzero distribution η χ \eta_\chi η χ is not central. If it were represented by b ∈ Z s b\in\mathcal{Z}_{\mathfrak{s}} b ∈ Z s , then ( a − a ( x χ ) ) b = 0 (a-a(x_\chi))b=0 ( a − a ( x χ )) b = 0 for every a ∈ Z s a\in\mathcal{Z}_{\mathfrak{s}} a ∈ Z s . Since Z s = C [ t ] \mathcal{Z}_{\mathfrak{s}}=\mathbb{C}[t] Z s = C [ t ] is a domain of positive dimension, this would force b = 0 b=0 b = 0 . Hence, for u ∈ J G u\in\mathcal{J}_G u ∈ J G ,
u ( x χ ) = 0 ( χ ≠ 1 , χ 2 = 1 ) . u(x_\chi)=0\qquad(\chi\ne1,\ \chi^2=1). u ( x χ ) = 0 ( χ = 1 , χ 2 = 1 ) .
This forces t + 2 t+2 t + 2 on the unramified component and t 2 − 4 t^2-4 t 2 − 4 on every ramified quadratic component. It remains to force t − 2 t-2 t − 2 on the unramified component. Let e T e_T e T be the Bernstein projector to the unramified characters of T ( F ) T(F) T ( F ) and put
α T ( h ) = e T ( r ‾ T G h ) . \alpha_T(h)=e_T(\overline{r}_T^G h). α T ( h ) = e T ( r T G h ) .
The projector e T e_T e T is supported on T ( O F ) T(\mathcal{O}_F) T ( O F ) . The constant-term formula therefore shows that α T ∈ D G c \alpha_T\in\mathcal{D}^c_G α T ∈ D G c , supported in B 0 ( G ( F ) ) B_0(G(F)) B 0 ( G ( F )) . The rank-one Plancherel formula gives
d 0 ∗ α T = c P , P ( t ) = ( 1 − q − 1 z ) ( 1 − q − 1 z − 1 ) = 1 + q − 2 − q − 1 t , c ≠ 0. (37) d_0\ast\alpha_T=cP,\qquad P(t)=(1-q^{-1}z)(1-q^{-1}z^{-1})=1+q^{-2}-q^{-1}t,\qquad c\ne0. \tag*{(37)} d 0 ∗ α T = c P , P ( t ) = ( 1 − q − 1 z ) ( 1 − q − 1 z − 1 ) = 1 + q − 2 − q − 1 t , c = 0. (37)
where P P P is extended by zero off the unramified component. For completeness, the continuous Plancherel density on its unitary torus is a nonzero constant times ( 1 − z ) ( 1 − z − 1 ) / P ( t ) (1-z)(1-z^{-1})/P(t) ( 1 − z ) ( 1 − z − 1 ) / P ( t ) ; see [14 Proposition 3.12, Theorems 3.14 and 4.5(2)(iii), and Sections 4.7–4.8] , specialized to the unramified case q 1 = q 2 = q q_1=q_2=q q 1 = q 2 = q . Multiplication by P P P leaves the torus trace weighted by d 0 d_0 d 0 . The same formulas show that the only discrete-series contribution in this component is Steinberg, with t = q + q − 1 t=q+q^{-1} t = q + q − 1 , where P = 0 P=0 P = 0 . This proves (37) ; the isolated reducible unitary principal series contributes no Plancherel atom.
For u ∈ J G u\in\mathcal{J}_G u ∈ J G write b = u ∗ α T ∈ Z G b=u\ast\alpha_T\in\mathcal{Z}_G b = u ∗ α T ∈ Z G . On the unramified component, (37) gives
( 2 − t ) b = c u P ( t ) . (2-t)b=cuP(t). ( 2 − t ) b = c u P ( t ) .
Since P ( 2 ) = ( 1 − q − 1 ) 2 ≠ 0 P(2)=(1-q^{-1})^2\ne0 P ( 2 ) = ( 1 − q − 1 ) 2 = 0 , one has ( t − 2 ) ∣ u (t-2)\mid u ( t − 2 ) ∣ u . Together with the quadratic-character factors, this proves J G ⊂ d Z G \mathcal{J}_G\subset d\mathcal{Z}_G J G ⊂ d Z G . Combining with (36) proves the first equality in Theorem 7.1.
The endoscopic ideal
The dual group is G ^ = PGL ( 2 ) \widehat{G}=\operatorname{PGL}(2) G = PGL ( 2 ) . Besides the datum with dual group G ^ \widehat{G} G , which imposes no condition, the connected endoscopic dual groups are tori. Their Weil action is either trivial or inversion. Thus the proper endoscopic groups are the split torus and the norm-one tori of quadratic extensions E / F E/F E / F .
For the split torus, the quotient Lie ( G ^ ) / Lie ( T ^ ) \operatorname{Lie}(\widehat{G})/\operatorname{Lie}(\widehat{T}) Lie ( G ) / Lie ( T ) is the sum of the two root lines. On the family χ = χ 0 ν z \chi=\chi_0\nu_z χ = χ 0 ν z their Weil characters are χ \chi χ and χ − 1 \chi^{-1} χ − 1 . If ϑ ≠ 1 \vartheta\ne1 ϑ = 1 , finite-inertia coinvariants vanish, so this datum imposes no condition. If ϑ = 1 \vartheta=1 ϑ = 1 , the zeroth Fitting ideal on the source torus is
( ( 1 − z ) ( 1 − z − 1 ) ) , ((1-z)(1-z^{-1})), (( 1 − z ) ( 1 − z − 1 )) ,
whose contraction to C [ t ] \mathbb{C}[t] C [ t ] is ( t − 2 ) (t-2) ( t − 2 ) .
Now let H H H be the norm-one torus of E / F E/F E / F , and let κ E \kappa_E κ E be the corresponding quadratic character of W F W_F W F , or of F × F^\times F × by local class field theory. Its L L L -embedding has image in the normalizer of T ^ \widehat{T} T , with nontrivial Weyl action specified by κ E \kappa_E κ E . Its source coarse parameter components are reduced points. Indeed, H ( F ) H(F) H ( F ) is compact, and [3 Proposition IX.6.4] identifies its spectral center with
l i m → K Λ [ H ( F ) / K ] . \varinjlim_K\Lambda[H(F)/K]. K lim Λ [ H ( F ) / K ] .
For compact open K ⊂ H ( F ) K\subset H(F) K ⊂ H ( F ) , the quotient H ( F ) / K H(F)/K H ( F ) / K is finite abelian, so its group algebra over Λ \Lambda Λ is a product of copies of Λ \Lambda Λ . Together with the characteristic-zero coarse-quotient description in [8 Theorem 1.7] , this shows that the bounded coarse parameter pieces are finite and reduced.
The Lie algebra of the centralizer inside H ^ \widehat{H} H is zero, since Weil acts on it by inversion. The centralizer inside G ^ \widehat{G} G has positive-dimensional Lie algebra precisely when the transferred parameter has image the order-two group generated by a Weyl reflection. Indeed, the centralizer of a reflection has connected component a torus; requiring the whole parameter to centralize that torus leaves only this order-two image. Up to conjugacy this is the parameter
w ⟼ diag ( κ E ( w ) , 1 ) in PGL ( 2 ) . w\longmapsto\operatorname{diag}(\kappa_E(w),1)\quad\text{in }\operatorname{PGL}(2). w ⟼ diag ( κ E ( w ) , 1 ) in PGL ( 2 ) .
We apply the construction of § 4 on one such source component. Its coordinate ring is S = Λ S=\Lambda S = Λ . Write R R R for the coordinate ring of the parameter scheme above it and M M M for the universal coinvariant module. The ideal used on the source quotient is
J = Fitt 0 R ( M ) ∩ S . J=\operatorname{Fitt}_0^R(M)\cap S. J = Fitt 0 R ( M ) ∩ S .
At the exceptional source point, the fiber of M M M is nonzero, so Fitt 0 R ( M ) \operatorname{Fitt}^{R}_{0}(M) Fitt 0 R ( M ) is proper. Its contraction to the field S S S is therefore zero. Hence the annihilator of the pushforward of S / J = Λ S/J=\Lambda S / J = Λ on the target is exactly the maximal ideal of x κ E x_{\kappa_E} x κ E .
At every other source point, the closed-orbit representative has zero coinvariant space. The invariant-support argument of § 3 shows that the support has empty image on this component. Thus the contracted ideal is the unit ideal. In particular, the exceptional contribution is a reduced point, with multiplicity one.
It follows that the endoscopic intersection on the unramified principal series component is
( t − 2 ) ∩ ( t + 2 ) = ( t 2 − 4 ) . (t-2)\cap(t+2)=(t^2-4). ( t − 2 ) ∩ ( t + 2 ) = ( t 2 − 4 ) .
On a ramified quadratic component the two quadratic extensions of ϑ \vartheta ϑ give the ideals ( t − 2 ) (t-2) ( t − 2 ) and ( t + 2 ) (t+2) ( t + 2 ) , again with intersection ( t 2 − 4 ) (t^2-4) ( t 2 − 4 ) . On all other principal series components the ideal is the unit ideal. At a cuspidal parameter the centralizer in PGL ( 2 ) \operatorname{PGL}(2) PGL ( 2 ) has zero Lie algebra, and hence there is no exceptional locus; the ideal is again the unit ideal. These are calculations of the Fitting and annihilator ideals, not only of their supports.
To compare with Z G Z_G Z G , the Fargues–Scholze parameter of a constituent of a restricted GL ( 2 , F ) \operatorname{GL}(2,F) GL ( 2 , F ) -representation is the projectivization of its semisimple parameter, by [3 Theorems IX.6.1 and IX.7.4] . On principal series components this gives exactly the coordinates used above. Several cuspidal Bernstein points can have the same spectral parameter; this causes no change to the extended ideal, since the spectral ideal is the unit ideal at every such parameter. The componentwise functions just calculated define a spectral element whose image is d d d , and every element of the spectral ideal pulls back to a multiple of d d d . Thus
Z G ⋅ Ψ G ( I G Γ ) = d Z G . Z_G\cdot\Psi_G(I_G^\Gamma)=dZ_G. Z G ⋅ Ψ G ( I G Γ ) = d Z G .
The calculations are independent of ramification bounds and of the representatives of the endoscopic data. They prove the second equality of Theorem 7.1 and complete the proof of Conjecture 1.5 for SL ( 2 ) \operatorname{SL}(2) SL ( 2 ) .