riemann

The Lean code claims a proof of the Riemann Hypothesis. This site has not checked the claim yet. See what is checked.

Block determinants and theta Parseval

The timeline has a Stepanov determinant and a positivity proof through a determinant. The Lean code does not contain them. This page shows what the code has in their place, in the proof that a theta lift is not zero.

draft

Result

The theorem UpperTheory.TP.K.tLiftK proves the statement TLiftK (see Step 7 of the Galois row). In short, TLiftK says this. Let EE be a CM field, let XX be a definite hermitian space of dimension nn and let UU be an automorphic representation of U(X)U(X) with weights λw\lambda_w at the infinite places. Then there is a d0d_0 such that for each skew-hermitian space WW of dimension d≥d0d \geq d_0 (with a weight condition) there is an automorphic representation U′U' of U(W)U(W) with IsLift: matched weights at infinity and matched local components at split places.

The proof is one line:

theorem UpperTheory.TP.K.tLiftK : UpperTheory.TP.TLiftK :=
  UpperTheory.TP.tLift_holds_k

theorem tLift_holds_k : TLift_k := tLift_of_tools_k tools_holds_k (hgtEx_of_k stab2Cpt_k)

This page is about tLift_of_tools_k (TPGTop.lean#L46). It has 503 declarations in its dominator subtree. Its heart is a proof that the theta lift of UU is not zero. The files are in the folders ThetaParseval/BlockDeterminants and ThetaParseval/ThetaKernels.

The Lean code is not reviewed by a person. This page does not claim that the statements are the intended ones.

Idea

The timeline of the plan has three steps that this page compares with the code:

  • Step 1: "The Stepanov determinant."
  • Step 4: "W(1S)≥0W(1_S) \geq 0 through a determinant. Read it as Stepanov with multiplicity one."
  • Step 8: iterated theta lifts U(X0)→⋯→U(Xm)U(X_0) \to \dots \to U(X_m) that gain multiplicity mm, with mm near X\sqrt{X}.

The word "Stepanov" does not occur in the Lean code. There is no auxiliary polynomial with a zero of high order, no Wronskian in this branch, and no quantity W(1S)W(1_S). The code has a different, standard way to show that something is not zero:

  1. A Rallis inner product identity (the "theta Parseval" of the folder name). If the lift of φ\varphi is zero, then the integral of the Weil matrix coefficient against the matrix coefficient of φ\varphi is zero.
  2. A choice of test data that makes this integral a main term plus a tail. The main term is not zero, and the tail is smaller. So the integral is not zero, and so the lift is not zero.
  3. In this branch, this page found determinants of matrices in one place only: the test vector at infinity is a product of powers of leading principal minors of a matrix of variables. This is the classical highest weight vector.

So the role of "a determinant that is not zero" is played by an inequality "main block is larger than the tail". The determinant itself enters only through the highest weight vector.

Before you read

Plan

  1. Show where the branch sits and how large it is.
  2. Show the existence skeleton tLift_of_k and its fourteen tools.
  3. Show the test vector made of leading minors (joint, realPair_k).
  4. Show the Rallis identity (rallis_k).
  5. Show the main term and the tail (nonVan_k and AbsNe).
  6. Show how a nonzero lift gives one irreducible U′U'.
  7. Say where the other determinants of the code are.

The steps

Step 1. Where the branch sits

The spine of tLiftK is short:

  • UpperTheory.TP.K.tLiftK (2,709 declarations) is dominated by galoisLocalGlobalStrong_recParameter (51,591).
  • tLift_holds_k (2,708) is the whole proof.
  • tools_holds_k (2,169) builds the structure Tools_k (GalTPDefs.lean#L27). It has fourteen fields: swFull (a full Siegel-Weil statement), boxSec, dbl (doubling), stab2, cpt (compactness of the quotients), basic, polSch, unfold0, pairEx (existence of a model of the dual pair), split, vacFix, acts, lie1 and lie2. This page does not describe their proofs.
  • tLift_of_tools_k (503) uses the tools to prove the lift. This page describes it.

The page the Galois row gives older sizes (5,201 for tLiftK and 2,998 for tLift_of_tools_k). The numbers on this page come from the current full spine.

Step 2. The existence skeleton

tLift_of_tools_k is one term:

theorem tLift_of_tools_k (T : Tools_k) (hH : HgtEx_k) : TLift_k :=
  tLift_of_k T hH brEx (hgtFst_k T) (realPair_k T)
    (nonVan_k T (rallis_k T) (decomp_k T) (bumpAny_k T) (datum_k T) (tail_k T) absNe (schurStep_k T))
    (liftAut_k T) projEx (readLift_k T (liftAut_k T) absIso)

So the proof is split into named statements. Each one is a def ... : Prop in GalTPDefs.lean#L45, and each one has its own proof ..._k. The theorem tLift_of_k (GalTPDefs.lean#L141) puts them together in this order:

  1. From HgtEx_k, get a height function aa on the doubled group. From HgtFst_k, get t0t_0 such that g↦a(g,1)tg \mapsto a(g,1)^t is integrable for t≥t0t \geq t_0.
  2. Take d0=max⁡(N,2n)d_0 = \max(N, 2n) with N>2t0N > 2 t_0. So each d≥d0d \geq d_0 has 2n≤d2n \leq d and t0<d/2t_0 < d/2.
  3. From pairEx, get a model of the dual pair (U(X),U(W))(U(X), U(W)) and its Weil representation. From brEx, get the identifications of the adelic groups.
  4. From RealPair_k, get a polynomial f0≠0f_0 \neq 0 and a form φ∈U\varphi \in U, φ≠0\varphi \neq 0, of matched types (Step 3).
  5. From NonVan_k, get finite test data φf\varphi_f with θ(f0⊗φf,φ)≠0\theta(f_0 \otimes \varphi_f, \varphi) \neq 0 (Steps 4 and 5).
  6. From LiftAut_k, ProjEx and ReadLift_k, get U′U' and its properties (Step 6).

The theta lift is the function on U(W)(𝔸)U(W)(\mathbb{A})

θ(Φ,φ)(h)=∫U(X)(F)\U(X)(𝔸)θ(g,h;Φ) φ(g)‾ dg,\theta(\Phi, \varphi)(h) = \int_{U(X)(F) \backslash U(X)(\mathbb{A})} \theta(g, h; \Phi) \, \overline{\varphi(g)} \, dg,

which is the Lean lift M om Φ φ h (SchurStep.lean#L22). The matrix coefficient of φ\varphi is mc M φ g, the Petersson product of the translate y↦φ(yg)y \mapsto \varphi(yg) with φ\varphi (SchurStep.lean#L25). The matrix coefficient of the Weil representation is coef om ν Φ g.

Step 3. Test vectors made of leading minors

This is the only place in the branch where determinants of matrices appear.

Let x=(xij)x = (x_{ij}) be a d×nd \times n matrix of variables with n≤dn \leq d. For s≤ns \leq n, let Δs\Delta_s be the leading s×ss \times s minor, lpMinor. For a weight ν1≥ν2≥⋯≥νn≥0\nu_1 \geq \nu_2 \geq \dots \geq \nu_n \geq 0, the Lean hw (StatementA.lean#L63) is

hwν=∏s=1nΔs νs−νs+1,νn+1=0.\mathrm{hw}_\nu = \prod_{s=1}^{n} \Delta_s^{\,\nu_s - \nu_{s+1}}, \qquad \nu_{n+1} = 0.

This is the classical highest weight vector of weight ν\nu for GLn\mathrm{GL}_n in the polynomial ring. The definition joint h μ (RightSub.lean#L32) moves hwμ\mathrm{hw}_\mu into the Fock space PP n d of polynomials in the n×dn \times d variables. The theorem jointNe (JointNe.lean#L13) says that joint h μ ≠ 0 when μ\mu is antitone. Its proof uses Formalism.GradedWeylDivisibility.hwNe.

Two theorems, exists_act1_equivariant_nablaSq_map_hitting_joint and exists_act2_equivariant_nablaSq_map_hitting_joint, give maps that commute with the GLn\mathrm{GL}_n and the GLd\mathrm{GL}_d actions and that hit joint. The proof for GLd\mathrm{GL}_d uses the rule that the top minor is multiplied by det⁡g\det g under the right action.

The theorem realPair_k (TPGRealPair.lean#L29) uses these to build f0f_0 as an image of ⨂wjoint(μw)\bigotimes_w \mathrm{joint}(\mu_w) over the infinite places ww. Its conclusion is that f0≠0f_0 \neq 0, that some φ∈U\varphi \in U with φ≠0\varphi \neq 0 has the same type as f0f_0 (SameType), and that at each ww the weight λw′\lambda'_w seen by f0f_0 on the U(W)U(W) side satisfies IsLiftWt. So the minors pick a test vector whose K∞K_\infty-types match on both sides.

Step 4. The theta Parseval identity

The statement Rallis_k (GalTPDefs.lean#L61) ends with:

Integrable (fun g => coef om ν Φ g * star (mc M φ g)) M.μ ∧
  (lift M om Φ φ = 0 → ∫ g, coef om ν Φ g * star (mc M φ g) ∂M.μ = 0)

In standard notation: if t0<d/2t_0 < d/2, then g↦⟨ω(g)Φ,Φ⟩ ⟨φ(⋅ g),φ⟩‾g \mapsto \langle \omega(g)\Phi, \Phi \rangle \, \overline{\langle \varphi(\cdot\, g), \varphi \rangle} is integrable on U(X)(𝔸)U(X)(\mathbb{A}), and

θ(Φ,φ)=0  ⟹  ∫U(X)(𝔸)⟨ω(g)Φ,Φ⟩ ⟨φ(⋅ g),φ⟩‾ dg=0.\theta(\Phi, \varphi) = 0 \implies \int_{U(X)(\mathbb{A})} \langle \omega(g)\Phi, \Phi\rangle \, \overline{\langle \varphi(\cdot\, g), \varphi\rangle} \, dg = 0.

This is one direction of the Rallis inner product formula. The proof rallis_k (TPGRallis.lean#L99) works like this:

  1. It bounds the Weil coefficient by a constant times a(g,1)d/2a(g,1)^{d/2}. With t0<d/2t_0 < d/2, this gives the integrability.
  2. It takes the Siegel Eisenstein series E^(z)\hat E(z) on the doubled group from swFull and its pairing with φ\varphi (kpair). For Re z\mathrm{Re}\, z large, the doubling tool dbl gives kpair(E^(z),φ)=c c′∫a(g,1) z−s0 ⟨ω(g)Φ,Φ⟩ ⟨φ(⋅ g),φ⟩‾ dg,s0=d−n2.\mathrm{kpair}(\hat E(z), \varphi) = c \, c' \int a(g,1)^{\,z - s_0} \, \langle \omega(g)\Phi, \Phi\rangle \, \overline{\langle \varphi(\cdot\, g), \varphi\rangle} \, dg, \qquad s_0 = \tfrac{d-n}{2}.
  3. A private lemma zero_of_half_pk (analytic continuation on a connected open set) extends this identity to z=s0z = s_0.
  4. At z=s0z = s_0, the Siegel-Weil statement writes the left side as a double integral of ⟨θ(g1),θ(g2)⟩\langle \theta(g_1), \theta(g_2) \rangle against φ\varphi. If the lift is zero, this is zero, and so the right side is zero.

Step 5. A main term that is larger than the tail

NonVan_k (GalTPDefs.lean#L114) says: if 2n≤d2n \leq d, t0<d/2t_0 < d/2 and f0f_0, φ\varphi have the same type, then there are finite test data φf\varphi_f with lift M om (pure f₀ φf) (φ ∘ e.symm) ≠ 0. By Step 4, it is sufficient to show that the integral of Step 4 is not zero. The proof nonVan_k (TPGNonVan.lean#L26) uses five inputs.

  • Decomp_k: for a finite set SS of places, U(X)(𝔸)=G∞×GS×GSU(X)(\mathbb{A}) = G_\infty \times G_S \times G^S as a product of three closed subgroups (gInf, gAt, gOff). G∞G_\infty is compact, because XX is definite.
  • BumpAny_k: at each place v∈Sv \in S there is a "bump" test function whose Weil coefficient is supported in a given small subgroup HvH_v.
  • Datum_k: with these bumps, the Weil coefficient vanishes for b∉∏Hvb \notin \prod H_v, and on x b cx \, b \, c with b∈∏Hvb \in \prod H_v it equals s(c)s(c) times its value at xx, where ∣s(c)∣=a(c,1)d/2|s(c)| = a(c,1)^{d/2} and s(c)=1s(c) = 1 when cc is integral at all places.
  • SchurStep_k: the integral over the compact G∞G_\infty of coef(x) mc(xc)‾\mathrm{coef}(x) \, \overline{\mathrm{mc}(xc)} is κ coef(1) mc(c)‾\kappa \, \mathrm{coef}(1) \, \overline{\mathrm{mc}(c)} with κ≠0\kappa \neq 0. By its name and its form, this is a Schur orthogonality step. It needs the matched types of Step 3.
  • Tail_k: for SS large, the integral of a(c,1)ta(c,1)^t over the part of GSG^S that is not integral is smaller than the measure of the integral part.

The abstract statement AbsNe (SchurStep.lean#L183) then closes the proof. In words: let ff on A×B×CA \times B \times C vanish for b∉Hb \notin H and equal s(c) m(x,c)s(c) \, m(x,c) for b∈Hb \in H, with 0<μ(H)<∞0 < \mu(H) < \infty. Let ∫Am(x,c) dx=κ c0(c)\int_A m(x,c)\,dx = \kappa \, c_0(c) with κ≠0\kappa \neq 0. Let s=1s = 1 and c0=r>0c_0 = r > 0 on a set KK, with ∣c0∣≤r|c_0| \leq r and ∣s∣≤w|s| \leq w everywhere. If ∫Kcw<μ(K)\int_{K^c} w < \mu(K), then ∫f≠0\int f \neq 0. The reason is that ∫f=μ(H) κ∫Cs c0\int f = \mu(H) \, \kappa \int_C s \, c_0, the part over KK has absolute value μ(K) r\mu(K) \, r, and the part over KcK^c has absolute value at most r∫Kcwr \int_{K^c} w, which is smaller. The bound ∣c0∣≤r|c_0| \leq r with r=c0r = c_0 on KK comes from UpperTheory.matrix_coefficient_bounded_by_value_at_one: a matrix coefficient is largest at 11.

This is the counterpart, in the code, of the nonvanishing of an auxiliary function. The main block is the integral part, where everything is explicit. The tail is controlled by the height. No determinant is involved.

Step 6. From a nonzero lift to one representation U′U'

  • LiftAut_k gives a map Θ\Theta that sends φ\varphi to the lift, shows that the lift is an automorphic form on U(W)U(W), and shows that Θ\Theta commutes with the group actions.
  • ProjEx (proof projEx, ProjEx.lean#L98) takes a nonzero automorphic form and gives an irreducible automorphic representation U′U' and an equivariant projection onto U′U' that does not kill the form. Its proof uses the Petersson inner product and Zorn's lemma.
  • ReadLift_k uses the abstract lemma AbsIso (an isotypic projection lemma) to read the weights of U′U' from the types of f0f_0, and the local pairs LocPair at split places.

So the code takes one irreducible constituent of the lift. It does not count the multiplicity of the constituent.

Step 7. The other determinants of the code

The folder Development/WeilBound/ThetaLifts/LeviDeterminants has 116 files. None of them is below tLiftK on the spine. Most of its declarations on the spine are below theta_integrals_agree_and_local_matching_of_literature (8,520), on the side of the theta matching. There, the determinants are determinants of Levi blocks and characters of the Siegel parabolic subgroup. An example is PositivityTheory.ThKF0.parabolic_character_trivial_on_diagonal (53 declarations), which needs det⁡J≠0\det J \neq 0. These are not Stepanov determinants either.

Where it is in the code

Declaration Module Size
UpperTheory.TP.K.tLiftK …ThetaKernels.TLiftK 2,709
tLift_holds_k …ThetaKernels.TPGHolds 2,708
tools_holds_k same 2,169
tLift_of_tools_k …ThetaKernels.TPGTop 503
realPair_k …ThetaKernels.TPGRealPair 140
RP.fockOf_k (below realPair_k) 93
RP.formEx (below realPair_k) 15
RP.exists_act2_equivariant_nablaSq_map_hitting_joint …ExponentialEntries.ExistsAct2EquivariantNablaSqMapHittingJoint 10
RP.exists_act1_equivariant_nablaSq_map_hitting_joint …ExponentialEntries.ExistsAct1EquivariantNablaSqMapHittingJoint 7
RP.jointNe …ExponentialEntries.JointNe 2
bumpAny_k …ThetaKernels.TPGBumpAny 128
liftAut_k …ThetaKernels.TPGLiftAut 38
brEx …BlockDeterminants.MemClosureOf 34
decomp_k …ThetaKernels.TPGDecomp 34
datum_k …ThetaKernels.TPGDatum 28
nonVan_k …ThetaKernels.TPGNonVan 18
readLift_k …ThetaKernels.TPGReadLift 14
schurStep_k …ThetaKernels.TPGSchurStep 11
tail_k …ThetaKernels.TPGTail 7
rallis_k …ThetaKernels.TPGRallis 6
hgtFst_k …ThetaKernels.TPGHgtFst 5
projEx …BlockDeterminants.ProjEx 2
absIso …BlockDeterminants.AbsIso 1
absNe …BlockDeterminants.AbsNe 1

The prefix … is Development.AutomorphicArithmetic.Foundations.GaloisRepresentations.ThetaParseval, and ExponentialEntries is inside BlockDeterminants. Names without a namespace are in UpperTheory.TP. "Size" is the number of declarations in the dominator subtree: the declarations that the trunk reaches only through this one. The small sizes of rallis_k, absNe and absIso mean that their proofs use shared lemmas that are also used elsewhere. The definitions hw, lpMinor and hwNe are shared in the same way: their dominator is the top theorem everyCharacter, not this branch.

Notes

  • The timeline step 1 (the Stepanov determinant) has no counterpart in the code. There is no auxiliary polynomial, no vanishing to high order and no Wronskian in this branch. The closest piece is the argument of Step 5, where a main term is shown to be larger than a tail.
  • The timeline step 4 (W(1S)≥0W(1_S) \geq 0 through a determinant) has no counterpart in the code. There is no quantity W(1S)W(1_S) in this branch and no positivity proved by a determinant. The code proves that one integral is not zero. It does not prove that some quantity is at least zero.
  • The idea "Stepanov with multiplicity one" has no counterpart. The code projects onto one irreducible constituent (projEx, absIso) and does not use or count multiplicity.
  • The timeline step 8 (iterated theta lifts U(X0)→⋯→U(Xm)U(X_0) \to \dots \to U(X_m), multiplicity mm, mm near X\sqrt{X}) has only a partial counterpart. The code has one lift U(n)→U(d)U(n) \to U(d) for all d≥d0=max⁡(N,2n)d \geq d_0 = \max(N, 2n). This page found no iteration of lifts in this branch, no gain of multiplicity and no parameter like mm near X\sqrt{X}.
  • The word "Parseval" in the folder name ThetaParseval is not used in any declaration of this branch. This page reads it as the Rallis inner product identity of Step 4, because that is the only identity of Parseval type in the branch.
  • The name BlockDeterminants covers files with no determinant in them (AbsNe, ProjEx, AbsIso, MemClosureOf). Only the subfolder ExponentialEntries and the definition hw use determinants.
  • The sizes on this page come from the full spine. They are smaller than the sizes on the Galois row page, which come from an older spine.
  • This page does not unfold the fourteen tools of Tools_k, the height IsHgt, IsBump, SameType, HasType, IsLiftWt and LocPair. The Rallis step depends on swFull and dbl, which are large statements on their own.
  • The Lean code is not reviewed by a person. The words above describe what the statements say, as read from the code.

The source of this page. Markdown. Something is wrong? Report it.