Bessel periods, seesaw and separation of functionals
The timeline says that the period of an iterated theta lift is a Bessel period, and that interlacing weights control it. The Lean code has a theorem of this shape, called FunctionalsSeparate. This page shows what it states, how the proof goes up a tower of Bessel and Fourier-Jacobi periods, and where the seesaw is.
draft
Result
The unitary hypotheses node needs a statement called FunctionalsSeparate. It gets it from one theorem:
theorem functionalsSeparate_of_howeUnram_of_siegelWeil (P : ThetaLocalInputs)
(hH : GlobalTheory.Hk3.HoweUnram) (hSW : GlobalTheory.Published.SiegelWeilUnitaryConvergent) :
FunctionalsSeparate :=
real_functionals_separate_of_hypothesis P hH hSW realInputE_holds
This is at FunctionalsSeparateOfHoweUnramOfSiegelWeil.lean#L17. Its dominator subtree has 4,530 declarations.
In words, FunctionalsSeparate says this. Take an iterated theta lift of a form along a row of steps, and take its Bessel period against a vector valued form on a smaller unitary group. For families with large parameters, there is one real test datum , chosen from the input vector only, such that this Bessel period is not zero for every nonzero Hecke eigen-direction . So the Bessel periods separate the inputs.
The proof has two parts. The first part goes up a tower of Bessel and Fourier-Jacobi periods. It is proved in Lean with no literature hypothesis. The second part is a local multiplicity one statement at the real place. It turns "for each there is a test datum" into "one test datum for all ".
The Lean code is not reviewed by a person. It compiles, but this page does not claim that the statements are the intended ones.
Idea
The timeline of the plan has two steps that this page compares with the code:
- Step 7: "Determinants for that contain the GGP periods. The values and mean interlacing weights."
- Step 12: "Seesaw: the period of the iterated lift is a Bessel period, and it does not depend on the Hecke characters (Rallis)."
The code has the first half of Step 12 in a precise form. The quantity in FunctionalsSeparate is, by definition, a Bessel period (a GGP period) of an iterated theta lift. The proof that it is not zero uses a seesaw identity at the bottom of a tower, and then goes up one rung at a time. Each rung changes a Fourier-Jacobi period into a Bessel period, or back. This is the method of Ginzburg, Rallis and Soudry for the descent.
The code has interlacing weights only as a sign condition on Harish-Chandra parameters at the real place. The signs take values in Bool, so they are the " and " of the timeline. This page did not find a determinant whose entries are GGP periods. It also did not find a statement that the period does not depend on the Hecke characters. See the Notes.
Before you read
- The theta correspondence for a pair of unitary groups, and the Weil representation.
- Hermitian forms over a CM field , and the adele ring .
- Hecke characters. In the code, the splitting characters of a Weil datum are maps .
- A fundamental domain for in . The code writes each automorphic integral as an integral over such a set.
- Automorphic forms and cusp forms, and the Bessel and Fourier-Jacobi periods of Gan, Gross and Prasad.
- Harish-Chandra modules and discrete series of , for the real place.
Plan
- Define the iterated lift
thetaChiand the Bessel periodBplus. - State
FunctionalsSeparate. - Follow the proof of
real_functionals_separate_of_hypothesis. - Look at the "seesaw pair"
PairEx. - Define the bottom period
per0and say why it is not zero. - Go up the tower:
Ascent, proved with no literature input. - Get one nonzero period for each :
WeakSep. - Get one test datum for all : real multiplicity one,
BurGapPos. - Say where interlacing appears.
The steps
Step 1. The iterated lift and the Bessel period
The code works with a grid of hermitian spaces over , with nodes such as .grid i 0, .core i and .top 0. For a node , Fn G X is the type of functions on . The type M : Vol G gives each node a set M.fd X and a measure M.ν X, and an integral over M.fd X plays the role of an integral over .
One lift step along an arrow of the grid is liftA. It is the integral of a theta kernel against the input form, over the fundamental set of the source. The iterated lift liftP composes these steps along a path. For a form on , a real datum and a Schwartz datum , the code sets (BplusOfEquivariant.lean#L263)
the lift along the row path from .grid 0 0 to .grid m 0. Each uses the Weil datum on the edge and the Schwartz function given by .
The Bessel period is Bplus (BplusOfEquivariant.lean#L286). Let be the space of the node .top 0, and let be the stabilizer embedding embTop for an orthogonal decomposition of . Let be the unipotent radical of a flag in (the type NFlag), and let be the generic character besChar. Then for on and on ,
The lemma Bplus_eq says that this is the general Bessel period Per.besPer of the pair, with in place of . So Bplus is the global GGP period of Bessel type.
The last definition is fPsiPhi (BplusOfEquivariant.lean#L326):
Here is a vector valued function and is a linear functional on its target space. The scalar function pairVec c ξ Ψ is .
So the code does contain the first claim of timeline Step 12, as a definition: the number that is studied is the Bessel period of the iterated lift.
Step 2. The statement FunctionalsSeparate
FunctionalsSeparate is at LevelPlusSubgroup.lean#L83. Leaving out the bookkeeping of the families, it says:
For all and all there is such that the following holds for each family whose size is at least and which satisfies the size conditions
SizeRows. For each grid with Weil data , each choice of measures of row type, and each nonzero vector in the spaceSmT (R.MY 0), there are and withSeparates R W M w ρ ξ.
And Separates (line 65 of the same file) says: for each nonzero in the eigenspace R.EEat R.triv there are in the column space Cols G W.θ and a datum such that the integrals converge (FConv) and
where = rowForm R w t is the form built from and .
The order of the quantifiers is the content. The pair depends on , but not on . The forms and may depend on .
Step 3. The proof of real_functionals_separate_of_hypothesis
The theorem real_functionals_separate_of_hypothesis (RealFunctionalsSeparateOfHypothesis.lean#L15, 3,524 declarations) takes four inputs: the literature record ThetaLocalInputs, the hypotheses HoweUnram and SiegelWeilUnitaryConvergent, and the real statement RealInputE. The fourth input is proved: it is realInputE_holds (942 declarations). The first three are hypotheses that come from the parent node.
The proof is a sequence of have steps:
rowModels,rowStep,rowTower: one theta step maps a space of forms of a given type to the next one, so the whole row is a tower of such spaces (RowTower). The steprowStepuses the "seesaw pair" of Step 4.conv: the integrals of the row converge.per: the bottom period is not zero (PeriodNe, Step 5).onto: the core lifts are onto (CoreOnto). This istwist_core_surjective, with 332 declarations. It usesHoweUnramandSiegelWeilUnitaryConvergent.weak: for each some Bessel period is not zero (WeakSep, Step 7). This step calls the ascentbessel_period_nonvanishing_ascent_of_depth_data(Step 6).- The last line,
bessel_separate_of_step_gap, makes the test datum independent of (Step 8).
Step 4. The "seesaw pair" PairEx
The anchor exists_seesaw_pair_of_theta_data (ExistsSeesawPairOfThetaData.lean#L40) has the type ThetaLocalInputs → PairEx. Its name says "seesaw". Its content is different.
PairEx (ParRowZero.lean#L17) says: for each pair of real parameters and each pair of discrete series models , , there is a linear map
where is the space PolF of the code, such that, for each Weil representation and each vacuum vector , the operator has the property IsItW (LinksChar.lean#L58). This property says that is nonzero, bounded, equivariant for one group of the dual pair and invariant for the other, at the real places.
So PairEx is the existence of a real intertwiner for one theta step. It is not a seesaw identity between two periods. Its proof uses the fields discreteSeriesCover and harishChandraModules of ThetaLocalInputs, so it rests on literature hypotheses. Its own subtree has only 3 declarations, because the lemmas it calls are shared with other branches.
Step 5. The bottom period
At the bottom of the tower, the code uses the period per0 (OffQOfSizeRows.lean#L81):
where is a form on the core node .core 0 and is a stabilizer embedding. This is a period of GGP type for the smallest pair of the tower.
PeriodNe (line 119) says that, for some core vectors, . It is proved by Branch.periodNe_of_eigen (651 declarations) from HarishChandraModules, CasselmanWallach, CuspRapidDecay and RealInputE. The statement RealInputE is a real matrix coefficient identity of the form
with bounds on decay. So the real place gives a nonzero local period, and a cusp form with a nonzero global period exists.
Step 6. The ascent
This is the main step. Ascent (OffQOfSizeRows.lean#L127) says:
Let be a chain of rapidly decreasing forms on the core nodes, with the down lift of (
IsColChain). Let be rapidly decreasing on the row. If , then there are and with .
In words: if the bottom period is not zero, then the Bessel period at the top of the iterated lift is not zero.
The theorem bessel_period_nonvanishing_ascent_of_depth_data : DepFixed → Ascent (BesselPeriodNonvanishingAscentOfDepthDataStatement.lean#L11, 2,123 declarations) proves it. The hypothesis DepFixed says that the depth vectors are fixed by the level groups. It is proved by dependence_fixed (DependenceFixedStatement.lean#L11).
The proof ascent_of_depFixed (BesselPeriodNonvanishingAscentOfDepthData.lean#L21) puts together four facts about a tower of periods:
Rung0: if , then the first Fourier-Jacobi period of the tower is not zero.RungB: if the Fourier-Jacobi period at level is not zero, then the Bessel period at level is not zero.RungJ: if the Bessel period at level is not zero, then the Fourier-Jacobi period at level is not zero.- The frames (
existsGoodGridFrames_of_depthFixed, 1,356 declarations): the grid has the orthogonal decompositions that each rung needs, and the top Bessel period of the tower isBplus.
Each rung is a statement about the lift of one form, UpBF and UpFB in TowerEx.lean. The rungs share one fact about the Heisenberg group: a tame functional on theta functions that kills the translates of one nonzero vacuum vector kills all vacuum vectors (tame_functional_vacuum_irreducibility, 45 declarations).
The seesaw is in Rung0. Its key lemma is UpLast (TowerEx.lean#L155). For an orthogonal sum with , it says: if
then the Fourier-Jacobi period of the lift against is not zero for some vacuum vectors , . The proof upLast writes both sides as one double integral over of , and changes the order of integration. This exchange of the two integrals is the seesaw identity.
The closure of bessel_period_nonvanishing_ascent_of_depth_data has 14,131 declarations. It contains no literature hypothesis and no axiom. So, if the Lean statements are the intended ones, the ascent is a theorem of the code and not an input.
Step 7. A nonzero period for each
WeakSep (OffQOfSizeRows.lean#L130) says that for each there are with . Here may depend on .
The proof weakSep_of_steps (RowSeparationByRealFunctionalsOfChainSteps.lean#L71) is short:
PeriodNegives core vectors with .CoreOntoputs these vectors in the span of the vectors that a chain can reach.- So there is a chain with and .
Ascentgives .
Step 8. One test datum for all
The last step is bessel_separate_of_step_gap (288 declarations). Its main part is separate_of_burGapPos (SepGap.lean#L222). It uses a factorization, the lemma factor_gap in the same file. It gives one function , the same for all , and , such that for each , and there is a number with
To build , the code uses that the line of real Bessel functionals at the top has dimension at most one. Now take any . WeakSep gives with , so . For a second , WeakSep gives some with , so the number for is not zero. Then . So the one datum works for every . This is FunctionalsSeparate.
The dimension bound is BurGapPos (BurGapOfHyp.lean#L33):
Let and be hermitian matrices with a frame in which and is positive definite. Let be a generic character, an irreducible unitary representation of with a gap at least in its parameter (
IsGapDS), and an irreducible finite dimensional representation of . Then the spacebesselFormsof Bessel functionals for on the smooth vectors of has rank at most .
This is real local multiplicity one for Bessel models, in a case with a large gap. It is proved in the code by burGapPos (145 declarations), not assumed.
Step 9. Interlacing
The interlacing condition of the timeline appears at the real place as Interlaced (ProjExists.lean#L34). Let be the Harish-Chandra parameter of a discrete series of a unitary group in variables, with signs , and let with signs be one for variables. Merge the two lists in increasing order. Then Interlaced ε λ ε' μ says: for two neighbours , of the merged list, the signs are equal if and only if and come from different lists.
This is the sign form of the branching law for discrete series. The signs are values in Bool, so they are the " and " of timeline Step 7. The structure Par (line 52) adds this condition and size bounds to a pair of parameters. The quantifier ForPair (line 81), which is used in the statement RealInputE of Step 5, ranges over such pairs.
There is a second form in Formalism. The statement BesselPeriodGGP (BesselPeriodGGP.lean#L105) says: for even and Harish-Chandra parameters , , with the interlacing InterlacingCondition (, and ) and half-integral differences, a discrete series and a finite dimensional of the right type have a nonzero local Bessel integral
This is the local GGP statement at the real place in the interlacing case. It is not reachable from the main theorem, so it is not part of the proof.
Where it is in the code
Size is the number of declarations that the node dominates in the spine of the main theorem. A size of 1 is a definition that is used in more than one place.
| Declaration | Module | Size |
|---|---|---|
UpperTheory.Lit.functionalsSeparate_of_howeUnram_of_siegelWeil |
FunctionalsSeparateOfHoweUnramOfSiegelWeil.lean#L17 | 4,530 |
UpperTheory.Lit.real_functionals_separate_of_hypothesis |
RealFunctionalsSeparateOfHypothesis.lean#L15 | 3,524 |
GlobalTheory.T.bessel_period_nonvanishing_ascent_of_depth_data |
BesselPeriodNonvanishingAscentOfDepthDataStatement.lean#L11 | 2,123 |
GlobalTheory.Asc.ascent_of_depFixed |
BesselPeriodNonvanishingAscentOfDepthData.lean#L21 | 2,122 |
GlobalTheory.T.existsGoodGridFrames_of_depthFixed |
ExistsGoodGridFramesOfDepthFixedStatement.lean#L12 | 1,356 |
GlobalTheory.T.fockConjugates_span_depth_data |
FockConjugatesSpanDepthDataStatement.lean#L11 | 226 |
GlobalTheory.T.bessel_to_fourierJacobi_nonvanishing_tower |
BesselToFourierJacobiNonvanishingTowerStatement.lean#L11 | 119 |
GlobalTheory.T.fourierJacobi_to_bessel_nonvanishing_tower |
FourierJacobiToBesselNonvanishingTowerStatement.lean#L11 | 71 |
GlobalTheory.T.tame_functional_vacuum_irreducibility |
TameFunctionalVacuumIrreducibilityStatement.lean#L11 | 45 |
GlobalTheory.T.nonvanishing_thetaTower_bottom_rung |
NonvanishingThetaTowerBottomRungStatement.lean#L11 | 15 |
GlobalTheory.T.lastSeesaw_periodNonvanishing |
LastSeesawPeriodNonvanishingStatement.lean#L11 | 12 |
GlobalTheory.T.dependence_fixed |
DependenceFixedStatement.lean#L11 | 30 |
UpperTheory.Lit.realInputE_holds |
RealBranch.lean#L231 | 942 |
UpperTheory.Lit.Branch.periodNe_of_eigen |
BranchPeriod.lean#L14 | 651 |
GlobalTheory.T.twist_core_surjective |
TwistCoreSurjectiveOfLocalComponents.lean#L32 | 332 |
UpperTheory.Lit.bessel_separate_of_step_gap |
BesselSeparate.lean#L9 | 288 |
UpperTheory.Lit.burGapPos |
Separate.lean#L32 | 145 |
UpperTheory.Lit.Bur.separate_of_burGapPos |
SepGap.lean#L222 | 56 |
GlobalTheory.T.rowSeparation_by_realFunctionals_of_chain_steps |
RowSeparationByRealFunctionalsOfChainStepsStatement.lean#L11 | 8 |
GlobalTheory.T.exists_seesaw_pair_of_theta_data |
ExistsSeesawPairOfThetaData.lean#L40 | 3 |
UpperTheory.Lit.Rallis.RallisIPProof.rallis_of_dm |
RallisDM.lean#L24 | 93 |
PI.BesselPeriodGGP.BesselPeriodGGP |
BesselPeriodGGP.lean#L105 | not in the spine |
The node functionalsSeparate_of_howeUnram_of_siegelWeil is a child of nonempty_hyps_definite_side_of_closed_bound (SiegelEisensteinSecondTermHypotheses.lean#L44). The node rallis_of_dm is in a different branch: its parent is rallis_inner_product_convergent_of_literature, under the literature branch.
Notes
- What is hypothesis and what is proof.
ThetaLocalInputs,HoweUnramandSiegelWeilUnitaryConvergentare hypotheses of this branch. They are used in the "seesaw pair" (Step 4), in the bottom period (Step 5) and inCoreOnto(Step 3). The ascent (Step 6), the dependence statementdependence_fixed,RealInputEand the multiplicity one boundBurGapPosare proved in the code. The page did not check whether each hypothesis is a correct form of the published theorem it is named after. - The name "seesaw". The theorem
exists_seesaw_pair_of_theta_databuilds a real intertwiner for one theta step. It does not state a seesaw identity. The real seesaw identity of the branch is the change of the order of integration inupLast(Step 6). - Timeline Step 7, determinants. The page found no determinant for whose entries are GGP periods. In this branch the periods enter one at a time, and the step from many to one test datum is a dimension bound (
BurGapPos), not a determinant. The namespaceFormalism.QuantumDeterminantsis used for the hermitian forms and the matrix groups of the real place, not for a determinant of periods. The quantity of the timeline has no counterpart here. - Timeline Step 7, " and ". The only counterpart is the
Boolvalued sign vectors ofInterlacedand ofeps48inBesselPeriodGGP. The timeline does not say more, so the page does not claim more. - Timeline Step 12, independence of the Hecke characters. The page found no statement that the period does not depend on the Hecke characters. The rung statements
UpBF,UpFBandUpLasthold for all splitting characters that satisfy a product relation (for example ). This is a uniform statement in the characters, not an identity between periods for different characters. - Rallis. The name Rallis appears in
rallis_of_dm, which provesRallisInnerProductConvergentfromResidueKilled.DM,SiegelWeilUnitaryConvergentandWeilThetaConvergence. That statement is a Rallis inner product formula: the Petersson inner product of two theta lifts is a zeta integral at the pointsZero. It is not used byFunctionalsSeparate. The other proof that a theta lift is not zero, through a Rallis identity, is on the page Block determinants and theta Parseval. - The GGP statement in Formalism.
BesselPeriodGGPis a clear statement of local nonvanishing in the interlacing case, but it is not reachable from the main theorem. The proof usesRealInputEandBurGapPosin its place. - Not unfolded. This page does not unfold
SizeRows,Cols,rowForm,RowTower,CoreOntobeyond its role,Frames,Ends,Loc,StBF,StFB,IsGapDSorbesselForms. The plain words for these objects stay at the level of their names and shapes.
The source of this page. Markdown. Something is wrong? Report it.