Analytic Theory of Modular Forms Spring 2012, taught by Jacob Tsimerman. Contents 1 Elliptic Functions 2 2 Hecke Operators 4 3 L-Functions Associated to Modular Forms 8 4 The Functional Equation for the Zeta Function 10 5 Smooth Sums and Sharp Sums 12 6 The Approximate Functional Equation 15 7 Poincaré Series 17 7.1 The Fourier Expansion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 7.2 Properties of Jk−1 and K . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20 8 Congruence Subgroups 22 9 The Theta Function 25 10 Quadratic Forms 27 11 Harmonic Polynomials 34 12 Bounds for Fourier Coefficients 37 13 Metaplectic Groups and Half-Integral Weight Hecke Operators 38 14 Bounds in the Half-Integral Weight Case 41 14.1 Trivial Bounds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44 14.2 A Toy Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44 14.3 A Return to the Original Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48 14.3.1 Estimate I . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49 14.3.2 Estimate II . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50 14.3.3 The Final Result . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51 1 15 Heegner Points and Geodesics 52 16 Maass Forms 53 17 Subconvexity of L-Functions 64 18 The Burgess Estimate 69 19 Amplification 70 20 Applications of Subconvexity 78 1 Elliptic Functions w1 Given w1 , w2 ∈ C with w ∈ / R, let Λ be Zw1 ⊕ Zw2 . f is elliptic if f is meromorphic on C and 2 Λ-periodic. Define EΛ = C/Λ; then elliptic functions are meromorphic on EΛ . An example of an elliptic function is 1 ℘(z; Λ) = 2 + z X w∈Λ\{0} 1 1 − 2 2 (z − w) w . (1.1) The poles of ℘ consist precisely of a double pole at 0 (mod Λ). P Define Gk (Λ) = w∈Λ w1k . Then for |z| < minw∈Λ\{0} |w|, we have ℘(z) = z −2 + X (k + 1)Gk+2 (Λ)z k . (1.2) k≥1 After differentiating and comparing principal parts, let g2 = 60G4 (Λ) and g3 = 140G6 (Λ). Then = 4℘(z)3 − g2 ℘(z) − g3 . We also find that for k > 6, Gk is a polynomial in G4 and G6 with integer coefficients, by equating higher terms. Observe that Gk (λw1 , λw2 ) = λ−k Gk (w1 , w2 ), and if ac db ∈ SL2 (Z), then Gk (aw1 + bw2 , cw1 + dw2 ) = Gk (w1 , w2 ). Let Gk (z) = Gk (1, z); then ℘0 (z)2 Gk ( az + b ) = (cz + d)k Gk (z). cz + d (1.3) For H the upper half-plane, Gk can be viewed as a function H → C. f : H → C is a modular function of weight k if f satisfies (1.3) for every ac db ∈ SL2 (Z). If f is also holomorphic and “bounded at ∞”, we say f is a modular form. Define F = {z ∈ H : |z| > 1, <(z) < 21 }. Then F is a fundamental domain for SL2 (Z) on H. That is, for every z ∈ H, there exists γ ∈ SL2 (Z) such that γ · z ∈ F ∪ ∂F, and if z1 , z2 ∈ F such that there exists γ ∈ SL2 (Z) with γ · z1 = z2 , then z1 = z2 . If f is a modular function, then f (z + 1) = f (z), so f is a function of q = e2πiz defined for q ∈ D \ {0}. “Bounded at ∞” means this function can be extended to q ∈ D. 2 k zeros modulo SL2 (Z). What this means Lemma 1.1. f (z) modular of weight k implies f (z) has 12 1 is that for p ∈ C, define m(p) = 2 |stabSL2 (Z) (p)|, so m(i) = 2 and m(ρ) = 3. Then define mf (p) = ordp f and mf (∞) = ordq=0 f (q). Then we have X p (mod SL2 (Z)) mf (p) k = . m(p) 12 (1.4) f0 f dz over ∂F, with appropriate tweaking. k −2 a b Alternatively, d( az+b ∈ SL2 (Z), so f (z)(dz) 2 is invariant under c d cz+d ) = (cz + d) dz for SL2 (Z), so descends to SL2 (Z)\H. Sketch of proof. Integrate Denote the space SL2 (Z)\H by Y0 (1), and let X0 (1) = Y0 (1) ∪ {∞}, so X0 (1) is a compact Riemann surface of genus 0. Let Ω be the line bundle of complex differential 1-forms on X0 (1). k k Then f (z)(dz) 2 induces a global meromorphic section of Ω 2 . dq ; near i, w = (z − i)2 so dw = 2(z − i)dz; In local coordinates: near ∞, q = e2πiz so dz = 2πiq near ρ, w = (z − ρ)3 so dw = 3(z − ρ)2 dz. We obtain deg k Ω2 k 2 (f (z)(dz) ) = X p6=i,ρ,∞ = X mf (p) p m(p) As X0 (1) has genus 0, deg Corollary 1.2. k mf (p) + mf (∞) − + 2 k Ω2 − mf (i) − 2 k 2 ! 13k . 12 + mf (ρ) − k 3 (1.5) (1.6) = −k, implying the desired result. • G4 (z) is modular of weight 4, so G4 has ρ as its only zero modulo SL2 (Z). • G6 has i as its only zero modulo SL2 (Z). Define ∆(z) = g2 (z)3 − 27g3 (z)2 , so that ∆ is modular of weight 12 and not identically zero (as ∆(i) 6= 0). Furthermore, ∆(∞) = (60 · 2ζ(4))3 − 27(140 · 2ζ(6))2 = 0. (1.7) So ∆ is nonvanishing in C. 3 2 (z)) so j is a modular function of weight 0, having a single pole at ∞, Now define j(z) = (12g ∆(z) so j descends to a map X0 (1) → P1 of degree 1, therefore an isomorphism. Theorem 1.3. For k even, k 12 dim Mk (SL2 (Z)) = k +1 12 k≡2 (mod 12) (1.8) otherwise Moreover, Mk (SL2 (Z)) is spanned by the Ga4 Gb6 , where 4a + 6b = k. 3 To do There’s a gap here. Fill in? (1) 2 Hecke Operators Recall that for L the space of lattices in C, we have an equivalence between modular functions of weight k under SL2 (Z) and G : L → C such that for λ ∈ C× , G(λΛ) = λ−k G(Λ). Now for n ∈ N, define X (Tn G)(Λ) = nk−1 G(Λ0 ). (2.1) Λ0 ⊆Λ [Λ:Λ0 ]=n Lemma 2.1. • If gcd(n, m) = 1, then Tn ◦ Tm = Tnm . • For p prime, Tp ◦ Tpa = Tpa+1 + pk−1 Tpa−1 . Proof. The first part is easy to show. For the second, consider lattices up to scale, L/R× , and draw an edge between Λ1 and Λ2 if there exists λ ∈ R× with λΛ2 ⊆ Λ1 and [Λ1 : λΛ2 ] = p. (Then [Λ2 : λp Λ1 ] = p, so the graph is symmetric.) Then Tpa consists of averaging over points a away. Alternatively, we could split up the sum for Tpa+1 , keeping track of homogeneity. Remark. The graph above can be thought of as P GL2 (Qp )/P GL2 (Zp ). Corollary 2.2. Tpa ∈ Z[Tp ]. Corollary 2.3. Tn ◦ Tm = Tm ◦ Tn for every m, n. In fact, Tn ◦ Tm = X dk−1 T mn . 2 (2.2) d d| gcd(m,n) T = Z[Tn ]n∈N is called the Hecke algebra. Let M2 (n) ⊆ M2 (Z) be the matrices of determinant n. Then the space of index n sublattices is parameterized by [ a b SL2 (Z)\M2 (n) = . 0 d (2.3) ad=n b mod d Now if f is a modular function of weight k, then (Tn f )(z) := (Tn F )h1, zi = n k−1 X F hd, az + bi = n k−1 X ad=n ad=n b mod d b mod d −k d f az + b d . (2.4) Corollary 2.4. T preserves the spaces Mk of modular forms and Sk of cusp forms of weight k. 4 Now if f (z) = P∞ m=1 af (m)e(mz), then ∞ X nk−1 X maz + mb (Tn f )(z) = , af (m)e d dk (2.5) m=1 ad=n b mod d which, after summing over b first, gives ∞ X af (m) m=0 X ak−1 e maz d ad=n . (2.6) d|m So for fn = Tn f , we have ∞ X X af (m) = m=0 dk−1 af mn d|(m,n) d2 . (2.7) In particular, afn (1) = af (n). Remark. These operators exist since for α ∈ M (n), αSL2 (Z)α−1 ∩ SL2 (Z) has finite index in SL2 (Z). The set of such α is called COM (SL2 (Z)) and equals SL2 (Q). But for “most” discrete Γ ∈ SL2 (R), COM (Γ) = Γ. Recall that ∆ generates S12 , so is an eigenfunction of T. We get (Tn ∆)(z) = τ (n)∆(z). It follows 11 that τ is multiplicative, and if αp satisfies τ (p) = p 2 (αp + αp−1 ), then τ (pa ) = p 11a 2 a X αpa−2b . (2.8) b=0 As a result, the Ramanujan conjecture for ∆ is equivalent to |αp | = 1. Claim. For each δ > 0, if |τ (n)| = O (nδ+ ), then |τ (n)| ≤ σ0 (n)nδ . This follows from the above structure: assume |αp | ≤ 1. Then |τ (pa )| ≤ (a + 1)|αp |a p 11a 2 . Now −(a+2) τ (pa ) = p If |αp | > 1, then this is Ωp (p 11a 2 11a 2 · αpa+2 − αp αp2 − αp−2 . (2.9) |αp |a ). Now we want a basis of Sk or Mk consisting of eigenfunctions of T. To get existence, it is enough to show that each Tn is semisimple. 2 2 Consider H with the metric dx y+dy , invariant under SL2 (R), and the invariant measure 2 For f (z), g(z) ∈ Mk , with f or g in Sk , define the Petersson inner product Z f (z)g(z)y k hf, gik = SL2 (Z)\H Lemma 2.5. If g ∈ Sk , then hg, Ek i = 0. 5 dx dy . y2 dx dy . y2 (2.10) Proof. Z hg, Ek i = g(z) SL2 (Z)\H Z = ζ(k) g(z) F yk dx dy k (cz + d) y 2 (c,d)6=(0,0) X (2.11) yk dx dy . k y2 (cz + d) (c,d)=1 X (2.12) Define h(z) = g(z)y k . Then h az + b cz + d = (cz + d)k g(z) yk yk = g(z) . |cz + d|2k (cz + d)k ∼ Define Γ∞ = {( 01 n1 ) , n ∈ Z}; then Γ∞ \SL2 (Z) − → {(c, d) = 1} via a b c d (2.13) 7→ (c, d). Now our integral is Z X ζ(k) h(γz) F γ∈Γ \SL (Z) ∞ 2 dx dy = ζ(k) y2 Z X h(z) γ∈Γ∞ \SL2 (Z) γF Z = 2ζ(k) Z h(z) ∞Z = 0 0 Γ∞ \H 1 g(z)y k dx dy y2 dx dy y2 dx dy y2 =0 (2.14) (2.15) (2.16) (2.17) since g is a cusp form. Lemma 2.6. For f, g ∈ Sk , hTn f, gi = hf, Tn gi. Proof. Define Lp = {(Λ, Λ0 ) : Λ0 ⊆ Λ, [Λ : Λ0 ] = p}, and let wp : Lp /R× → Lp /R× be the involution (Λ, Λ0 ) 7→ (Λ0 , pΛ). We also have a map H → Lp /C× by z 7→ (h1, zi, h1, pzi). We have Lp ,→L × L via the natural embedding. We obtain a map H → L/C× × L/C× . Defining Γ0 (p) = a b ∈ SL (Z), p|c , then Γ (p) preserves h1, zi and h1, pzi. Now Γ (p)\H (L → L/C× × 2 0 0 p c d L/C× ) with wp = have 0 −p−1/2 p1/2 0 hf, Tp gi = p k−1 1 . (Note that wp (z) = − pz , and that wp Γ0 (p)wp−1 = Γ0 (p).) Now we Z (2.18) f (z)y k (2.19) Γ0 (p)\H = pk−1 Z Γ0 (1)\H dx dy y2 ! p−1 X z+i dx dy g + pk−1 g(pz) . p y2 f (z)g(pz)y k i=0 1 Now setting w = − pz , we get 6 pk−1 Z f Γ0 (p)\H =p k−1 1 1 v k du dv − g − pw w pk |w|2k v 2 Z f (pw)(p|w|k )g(w)(|w|k ) Γ0 (p)\H = pk−1 Z f (pw)g(w)v k Γ0 (p)\H v k du dv pk |w|2k v 2 du dv v2 (2.20) (2.21) (2.22) = hTp f, gi. (2.23) As the Tp generate T, we get the desired result. The lemma implies that there exist {f1 , . . . , fk } an orthogonal basis of Sk such that fi is an eigenfunction of every Tn , called a Hecke cusp form. Theorem 2.7 (Multiplicity One). If f, g ∈ Sk are Hecke cusp forms with the same eigenvalues under all of the Tn , then f and g differ by a constant. This is true since Tn determines the Fourier coefficients. For f a Hecke cusp form, let λn (f ) = Tn f f . Define Mk (Z) ⊆ Mk to be the set of modular forms having integral Fourier coefficients. Since E4a E6b are in Mk (Z) after an appropriate scalar multiplication, Mk (Z) is a lattice in Mk . And T preserves Mk (Z), so the λn (f ) are algebraic integers. Corollary 2.8. For (λn (f ))n∈N the Hecke eigenvalues for f ∈ Sk , and γ ∈ Gal(Q/Q), then there exists f γ ∈ Sk such that λn (f γ ) = (λn (f ))γ . The square-root cancellation heuristic: for a1 , . . . , aN complex numbers which are “random” and |ai | ∼ 1, then N X √ ai ∼ N . (2.24) i=1 Theorem 2.9. Fix f ∈ Sk . Then N X |af (n)| f n k+1 2 . (2.25) f (x + iy)e(−nx) dx, (2.26) n=1 That is, the sum is Of (n k+1 2 ). Proof. |f (z)|2 y k = O(1). Then bound af (n) by e −2πiny Z af (n) = 1 0 trivially bounding, and optimizing with respect to y. 7 A better bound can be obtained using Parseval’s identity: 1 Z |f (θ)|2 dθ = 0 X |an |2 . (2.27) n∈Z Applying to the above, we get ∞ X |af (n)|2 e−4πy = 1 N, |f (x + iy)|2 dx y −k . (2.28) 0 n=1 For y = 1 Z we get N X |af (n)|2 e− 4πn N N k =⇒ n=1 N X |af (n)|2 N k . (2.29) n=1 Then Cauchy implies the theorem. 3 L-Functions Associated to Modular Forms Suppose f is a Hecke cusp form of weight k. Define the (analytically normalized) L-function L(f, s) = ∞ X λf (n) n=1 ns+ k−1 2 . (3.1) This converges absolutelt for <(s) > 1 by the above. Because λf (n) is multiplicative, ∞ Y X λf (pn ) L(f, s) = p Letting λf (p) = p k−1 2 n=0 pn(s+ ! k−1 ) 2 . (3.2) −1 (αp,f + αp,f ), ! n k−1 2 n X λf (p ) = p −b a αp,f αp,f . (3.3) a+b=n We finally end up with L(f, s) = 1 Y p (1 − αp,f p−s )(1 −1 −s − αp,f p ) . Here is how to obtain a meromorphic continuation and functional equation: holomorphic for every s. It equals 8 (3.4) R∞ 0 f (iy)y s dy y is Z ∞ ∞X 0 af (n)e −2πny s dy y y n=1 = ∞ X n=1 ∞ X Z ∞ af (n) e−2πny y s 0 dy y Z ∞ af (n) dy −s e−y y s = (2π) s n y 0 n=1 k−1 (2π)−s Γ(s). = L f, s − 2 (3.5) (3.6) (3.7) But f ( yi ) = f (iy)(iy)k , so ∞ Z 0 ∞ i dy f (iy)y f y s−k =i y y y Z0 ∞ dy = ik f (iy)y k−s y 0 k+1 = L f, − s (2π)s−k Γ(k − s). 2 s dy k Z (3.8) (3.9) (3.10) We conclude that L(f, s)(2π) k−1 −s 2 k−1 k+1 k s− k+1 2 Γ = i L(f, 1 − s)(2π) −s . Γ s+ 2 2 (3.11) L is also holomorphic everywhere since the integral and Γ1 are. We also find that L(f, s) has k+1 “trivial zeros” at − k−1 2 , − 2 , . . .. The generalized Riemann Hypothesis asks if all other zeros of 1 L(f, s) lie on <(s) = 2 . Theorem 3.1 (Strong Multiplicity One). If f, g are Hecke cusp forms of weight k succh that for all but finitely many primes p (∀0 p) λf (p) = λg (p), then f and g differ by a constant factor. Proof. Let R(s) = L(f,s) L(g,s) = Qr i=1 Ri (s), Ri (s) = for −1 −s (1 − αpi ,f p−s i )(1 − αpi ,f pi ) −1 −s (1 − αpi ,g p−s i )(1 − αpi ,g pi ) . (3.12) We have R(s) = R(1 − s). As a result, it turns out that for every i, Ri (s) = Ri (1 − s). (We can expand the Dirichlet series, and need a finite product to expand for s < 12 .) Letting X = p−s i , the rational function (1 − αpi ,f X)(1 − αp−1 X) i ,f (1 − αpi ,g X)(1 − αp−1 i ,g X) (3.13) pi is invariant under X 7→ X . Either we get {αpi ,f , αp−1 } = {αpi ,g , αp−1 } (we’re done if this is true i ,g i ,f pi for every i), or the poles of Ri (s) are invariant under X 7→ X , which is impossible. 9 The subconvexity problem: for F a family of L-functions L(ϕi , s), we want to understand |L(ϕi , 12 )|. This is typically done for ϕi the Hecke cusp forms of weight k, or consider |L(f, 12 + it)| as t → ∞ for f a Hecke cusp form. For σ > 1, |L(f, σ + it)| is bounded as a function of t. By the functional equation, for σ < 0, |L(f, σ + it)| ∼ |t|ψ(σ) for some explicit linear function ψ. We can then interpolate to get |L(f, σ + it)| ∼ |t|ψ0 (σ) . 4 The Functional Equation for the Zeta Function For f ∈ L2 (R), define fb by fb(y) = Fact. If for every j ≤ A, dj f dxj R∞ Then fb ∈ L2 (R), and f 7→ fb is an isometry. −∞ e(xy)f (x) dx. ∈ L1 ∩ L∞ , then fb(y) = O((1 + |y|)−A ). For integration by parts gives fb0 (y) = −2πifb(y). P P Let g(x) = n∈Z f (x + n), so g ∈ C 0 (R/Z). We have g(x) = n∈Z gb(n)e(nx) for Z gb(n) = Z g(x)e(−nx) dx = f (x)e(−nx) dx = fb(−n). R/Z (4.1) R By taking x = 0, we recover Theorem 4.1 (Poisson Summation Formula). X X f (n) = fb(n). n∈Z (4.2) n∈Z 2 Apply this to ht (x) = e−πtx . We have Z b ht (y) = e R = e− πy 2 t = e− πy 2 t = e− πy 2 t ix2 t + xy dx 2 Z 2 ix t iy 2 e + xy − dx 2 2t R ! Z it iy 2 x− dx e 2 4t R Z 2 e−πtx dx =(x)==( iy ) 4t πy 2 e− t = √ t By a contour argument, the integral is Z 2 =(x R Re √ t)==( iy ) 4t −πx2 e−πx dx. (4.3) (4.4) (4.5) (4.6) (4.7) dx = 1, so we have 2 πy h 1 (y) e− t t b . ht (y) = √ = √ t t 10 (4.8) Now for =(z) > 0, define θ0 (z) = X 2z eπin n∈Z = X h−iz (n). (4.9) n∈Z By Poisson, we get 1 X 1 1 i 2 i 2 h i (n) = θ0 − . z 2 2 z (4.10) n∈Z θ0 is a modular form of half-integral weight. R∞ s Now for <(s) > 1, 0 θ0 (iy)−1 y 2 dy 2 y converges. It equals ∞ Z X ∞ 2 s e−πn y y 2 n=1 0 s s dy = π 2 ζ(s)Γ . y 2 But we can also express the left hand integral by splitting into (4.11) R∞ 1 and 1 Z ∞ θ0 (iy) − 1 s dy θ0 (iy)y 2 − 1 − s dy 2 y + y 2 2 y 2 y 1 1 Z ∞ dy 1−s 1 1 θ0 (iy) − 1 s y2 + y 2 + − . = 2 y s−1 s 1 Z R1 0 . We obtain ∞ (4.12) (4.13) This is meromorphic everywhere with poles only at 0 and 1 and symmetric under s 7→ 1 − s. We get s 2 π ζ(s)Γ s An exercise is to show that ζ(s) = √ 2 1 s−1 =π 1−s 2 ζ(1 − s)Γ 1−s 2 . (4.14) + γ + · · · as s → 1. For σ ∈ R, define ψ(σ) = inf{r : |ζ(σi t)| |t|r }. For σ > 1, ψ(σ) = 0. Also |Γ(σ + it)| ≈ 1 π|t| 2π|t|σ− 2 e− 2 for each σ and t → ∞. This and the functional equation gives us, for σ < 0, ψ(σ) = 12 − σ. In fact, for every σ, ψ(σ) = ψ(1 − σ) + 12 − σ. Theorem 4.2. ψ(σ) is convex in σ. Lemma 4.3. Suppose f (s) is holomorphic for a ≤ σ ≤ b, |f (s)| ≤ M on σ = a and σ = b, and α |f (σ + it)| = O(e|t| ). Then |f (s)| ≤ M for a ≤ σ ≤ b. m Proof of lemma. Pick m ≡ 2 mod 4 with m > |α| and look at f (s)e−s . This decays at ∞, so we m can apply the maximum modulus principle. We get |f (s)| ≤ maxσ=a,b M |es | for every , and then take → 0. Proof of theorem. Let L(s) = lemma. b−s b−a ψ(a) + a−s a−b ψ(b) and consider ζ(s)(−is)−L(s) , and then apply the 11 The theorem implies that for 0 ≤ σ ≤ 1, ψ(σ) ≤ Hypothesis is that ψ( 21 ) = 0. 1 2 − σ2 . In particular, ψ( 12 ) ≤ 14 . The Lindelof Theorem 4.4. The Riemann Hypothesis implies the Lindelof Hypothesis. Proof. Assuming Riemann, log(ζ(s)(s − 1)) is holomorphic for <(s) > 12 . And for any σ > <(log(ζ(s)(s − 1))) ≤ C log |t|. By complex analysis, we obtain | log(ζ(s)(s − 1))| log |t|. 1 2, (Borel-Catheodory: if f (z) is holomorphic on |z| < R with |f (z)| ≤ M on |z| < r, and sup <(f ) ≤ N , then |f (z)| ≤ ϕ(M, N, r, R) for some ϕ.) Then convexity gives | log(ζ(s)(s−1))| (log |t|)γ for some γ < 1. Exponentiating gives Lindelof. 5 Smooth Sums and Sharp Sums 1 For <(s) > 1, ζ(s) = Riemann Hypothesis. P∞ µ(n) n=1 ns . We can then understand partial sums of µ(n), assuming the Mellin Transform: For ϕ(x) ∈ C 0 (R+ ), ∞ Z ϕ(s) e = ϕ(x)xs 0 dx x (5.1) converges for a < σ < b. For every c with a < c < b, 1 2πi Z σ=c −s ϕ(s)x e dx = ϕ(x). (5.2) (After substituting x = ey , this becomes Fourier inversion.) j Suppose that for j ≤ A, ddxϕj ∈ L1 ∩ L∞ (R+ ). Then |ϕ(σ e + it)| = O((1 + |t|)−A ). In particular, if ϕ ∈ Cc∞ (R+ ), then |ϕ(σ e + it)| = O((1 + |t|)−A ) for every A. Theorem 5.1. The Riemann Hypothesis is true if and only if for every ϕ ∈ Cc∞ (R+ ), ∞ X µ(n)ϕ n x n=1 1 = O (x 2 + ). (5.3) Proof. For the “only if” implication, consider 1 2πi Z σ=a ∞ X 1 1 s ϕ(s)x e dx = µ(n) · ζ(s) 2πi = n=1 ∞ X n=1 for a where and |t| large. 1 ζ(s) µ(n)ϕ Z σ=a ϕ(s) e n 12 (5.4) (5.5) x converges. Assuming Riemann, we can take a = xs ds ns 1 2 + and bound for t near 0 For the “if” implication, given the bound, then ∞X Z µ(n)ϕ n 0 x x−s dx x (5.6) converges for σ > 12 , so we get ζ −1 (s)ϕ(s) e holomorphic for σ > 21 . But for every s, there exists ϕ ∈ Cc∞ (R+ ) with ϕ(s) e 6= 0. So ζ −1 (s) is holomorphic for σ > 12 . P Givenn a sequence (an )n∈N , we can either consider n an ϕ( x ). Here are some examples: • For an = 1, P n≤x 1 P n≤x an , or choose ϕ ∈ Cc∞ and consider = bxc = x + O(1), while X X n = xϕ(0) b +x ϕ(nx) b = xϕ(0) b + OA (x−A ). ϕ x n (5.7) n6=0 P 1 • Let r2 (n) = #{a, b ∈ Z : a2 + b2 = n}. Then n r2 (n) = πx + O(x 2 ). The Gauss circle problem is to lower the exponent as much as possible (it is believed that 14 + is possible). But X r2 (n)ϕ n x n = X ϕ a,b∈Z a2 + b2 x . (5.8) For ψ(a, b) = ϕ(a2 + b2 ), we get b 21 , dx 12 ) = xψ(0, b 0) + OA (x−A ). xψ(cx X b 0) + xψ(0, (5.9) (c,d)6=(0,0) • Suppose f ∈ Sk . Then Z X an (f ) n 1 ϕ = L(f, s)x−s ϕ(s) e ds k−1 x 2πi σ=2 n n 2 Z 1 = L(f, s)x−s ϕ(s) e ds 2πi σ=−A = OA (x−A ). (5.10) (5.11) (5.12) Computing sharp sums is in fact more difficult than computing smooth sums, in that sharp sum estimates imply smooth sum estimates. For if ϕ ∈ Cc∞ (R+ ), then (5.13) N X n X N +1 N δ an ϕ N ϕ −ϕ xδ . n x x x (5.14) n So if PN n=1 an N X X ! N +1 N an ϕ −ϕ . x x X an ϕ n x = N n=1 = O(xδ ), then n=1 13 Theorem 5.2. The Riemann Hypothesis holds if and only if X 1 µ(n) = O (x 2 + ). (5.15) n≤x We only need to prove the “only if” implication. Perron’s Formula: for δ > 0 and y > 0, 1 2πi 1 ys ds = 0 s 1 Z σ=1+δ y>1 0<y<1 y=1 (5.16) ds X = µ(n). s (5.17) 2 Given this, 1 2πi Z ζ −1 (s)xs σ=1+δ n≤x Proposition 5.3 (Quantitative Perron). For δ > 0 and y > 0, Z 1+δ+iT 1+δ−iT 1+δ y 1 + O s T log y y ds = 1+δ s y O T log y y>1 (5.18) 0<y<1 Proof. For y > 1, consider To do Rectangular contour. (2) Integrating over the rectangle is 1. The top is bounded by 1+δ Z −A yσ 1+δ dσ ≤ . T T log y (5.19) The bottom is similar. For the left, we get at most Z T −T y −A dt → 0 |A + it| (5.20) as A → ∞. Proof of theorem. Pick y to be a half-integer. Then Z 1+δ+iT ζ 1+δ−iT We now estimate −1 (s)y s ds s = X n≤y ∞ X (y/n)1+δ µ(n) + O T | log(y/n)| . (5.21) n=1 (y/n)1+δ y 1+δ n n=1 T | log(y/n)| . If 1 ≤ y ≤ 2 , then | log(y/n)| ≥ log 2, so we get O( T ζ(1 + 1+δ get O( y T ζ(1 + δ)). Finally, for y2 ≤ n ≤ 2y, | log ny | = | log(1 + n−y y )| ∼ P∞ δ)). If n ≥ 2y, we also 1 | n−y y | ≥ 2y , so ! 14 2y X x= y2 We could get y 1+δ log y T Z 2y X y 2+δ (y/n)1+δ y2 ≤2 ≤ . 1+δ T T | log(y/n)| T y n (5.22) n= 2 by a more careful approximation. So 1+δ+iT ζ −1 (s)y s ds s 1+δ−iT = X µ(n) + Oδ n≤y y 2 y 1+δ + T T . (5.23) Now consider the contour To do Rectangle (3) Assuming Riemann, the integral over the rectangle is 0. Now Z top, bottom Z 1+δ |ζ −1 (σ + it)|y σ σ= 12 +δ dσ , T (5.24) 1+δ which by Riemann again is O ( Ty 1− ). And Z left Z 1 +δ+iT 2 1 1 ds |ζ −1 (s)|y 2 +δ = O (y 2 +δ T ). 1 |s| 2 (5.25) +δ−iT We obtain X 1 µ(n) = O (y 2 +δ T ) + Oδ n≤y y 2 y 1+δ + T T . (5.26) 1 For T = y 2 , we get O (y 2 + ). (Had we instead only known that |ζ −1 (σ + iT )| O (T a+ ), we 2 1+δ 1 would get O (y 2 +δ T a ) + Oδ ( yT + Ty 1−a ) and would optimize accordingly.) 6 The Approximate Functional Equation For 0 < <(s0 ) < 1, choose ϕ(x) ∈ Cc∞ (R+ ). The idea is to consider 1 2πi Z σ=2 ϕ(s0 + s)xs ϕ e ds X −s0 n = n ϕ . s x n (6.1) This also equals 1 2πi ds x−s0 ϕ(−s0 ) ζ(s0 + s)xs ϕ(s) e − + ϕ(0)ζ(s e 0 ). s s0 σ=−2 | {z } Z (?) Now (?) equals 15 (6.2) Z ζ(s0 − s)x − −s σ=−2 ds (?) ϕ(−s) e = s Z π 1+s−s0 ) 2 ζ(1 s0 −s Γ( 2 ) −1−2s+2s0 Γ( σ=2 + s − s0 )x−s ϕ(−s) e ds . s (6.3) Fix Me (4) We need to estimate 1 2πi Z y −s σ=2 0 Γ( 1+s−s ) ds 2 ϕ(−s) e . s0 −s s Γ( 2 ) (6.4) s Let Λ(s) = ζ(s)π 2 Γ( 2s ), and choose ϕ such that ϕ(x) = ϕ( x1 ), so that ϕ(−s) e = ϕ(s). e Then 1 2πi Z Λ(s + s0 ) s ds x−s0 ϕ(s e 0) Λ(s0 ) 1 x ϕ(s) e = − + ϕ(0) e s0 s0 s0 + Γ( 2 ) s s0 Γ( 2 ) Γ( 2 ) 2πi Z σ=−2 | Λ(s + s0 ) s ds x ϕ(s) e . s0 Γ( 2 ) s {z } (6.5) (?) (?) equals Z − σ=2 ds Λ(s0 − s) −s x ϕ(s) e =− s0 Γ( 2 ) s Z σ=2 Λ(1 + s − s0 ) −s ds x ϕ(s) e . s0 Γ( 2 ) s (6.6) We need to understand 1 Vs0 (y) = 2πi Lemma 6.1. Z σ=2 Γ( s02+s ) −s ds y ϕ(s) e . s0 Γ( 2 ) s !σ0 −δ y 1 + Oδ p 1 + |s0 | Vs0 (y) = !−A p y OA 1 + |s | (6.7) 0≤y≤1 (6.8) y≥1 0 and Vs00 (y) = Oδ √ y 1+|s0 | σ0 −δ ! for 0 ≤ y ≤ 1. Proof. Γ( s0 +s ) σ0 +σ− 12 s +s s − π2 (|=( 02 )|−|=( 20 )|) |t0 + t| 2 . ≈e 1 Γ( s20 ) |t0 |σ− 2 (6.9) We have | t02+t | − | t20 | ≤ |t|. Now 1 Vs0 (y) = 1 + 2πi Z σ=−σ0 +δ 16 Γ( s02+s ) −s ds y ϕ(s) e s0 Γ( 2 ) s (6.10) 1 and for t small, the second integral is Vs0 (y) = |t0 +t|δ− 2 |t0 | 1 2πi σ0 −δ −σ0 +δ− 1 2 Z σ=A Oδ ( yσ0 −δ ). Also t0 Γ( s02+s ) −A ds y ϕ(s) e . Γ( s20 ) s (6.11) For Vs00 (y), differentiating eliminates the pole at s = 0. For x > 0, we have ζ(s0 )π s0 2 X = n −s0 V s0 n x n 1−s0 + X n s0 −1 n x1−s0 π 2 V1−s0 (nx) + . Γ( s20 ) √ Pick s0 = 21 + iT . Then the two sums have length x T and (roughly) that ζ 1 + iT 2 π = Γ( iT 2 1 +iT 2 2 + 1 X ) n = X n 1 +iT 2 Vs0 (n) + √ T x , 1 X n n so both 1 −iT 2 √ (6.12) T if x = 1. We get Vs0 (n). (6.13) Morally, ζ 1 + iT 2 1 n− 2 −iT + √ n≤ T X 1 n− 2 +iT , (6.14) √ n≤ T so estimating, X 1 1 ζ 1 + iT ≤ 2 n− 2 ∼ T 4 . 2 √ (6.15) n≤ T In order to improve this, we need to exploit some cancellation in the above sums. 7 Poincaré Series Fix a positive even integer k and write jk ac db , z = (cz +d)−k . For γ ∈ SL2 (Z), write (f |k γ)(z) = f (γz)jk (γ, z) so that f ∈ Mk ⇐⇒ f |k γ = γ for every γ ∈ SL2 (Z). Also, for γ1 , γ2 ∈ SL2 (Z), we have the cocycle relation jk (γ1 γ2 , z) = jk (γ1 , γ2 z)jk (γ2 , z). Corollary 7.1. SL2 (Z) acts on C ∞ (H) by f 7→ f |k γ. Recall that Ek (z) = 1 2 X jk (γ, z) = γ∈Γ∞ \Γ 1 2 X (1|k γ)(z), (7.1) γ∈Γ∞ \Γ so we see easily that Ek is a modular form. More generally, for m ≥ 0, let em (z) = e(mz), and define the Poincaré series 17 pm,k (z) = 1 2 X (em |k γ)(z). (7.2) γ∈Γ∞ \Γ For m ≥ 0 and k > 2, pm,k ∈ Mk . Claim. If m > 0, then pm,k ∈ Sk . Proof. pm,k (z) = e(mz) + X −k (cz + d) c>0 az + b e m cz + d . (7.3) d∈Z As =(z) → ∞, e(mz) → 0 and P c>0 d∈Z Lemma 7.2. For f ∈ Sk , hf, pm,k ik = Proof. For dµ(z) = dx dy , y2 |cz + d|−k → 0, so pm,k (z) → 0. af (m) Γ(k (4πm)k−1 we have 1 hf, pm,k ik = 2 Z 1 = 2 Z = Z Γ\H 0 apm,k (n) nk−1 = (7.4) f (γz)em (γz)=(γz)k dµ(z) (7.5) Γ\H γ∈Γ \Γ ∞ f (z)e(mz)y k dx dy y2 f (z)e(mz)y k 0 dx dy y2 af (m) Γ(k − 1). (4πm)k−1 apn,k (m) mk−1 em (γz) jk (γ, z)y k dµ(z) γ∈Γ∞ \Γ X = = X f (z) Γ∞ \H Z ∞Z 1 Corollary 7.3. − 1). (7.6) (7.7) (7.8) . This follows from hpm,k , pn,k ik = hpn,k , pm,k ik . Corollary 7.4. The pm,k over m > 0 span Sk . As an exercise, it can be shown that for d = dim Sk , (p1,k , . . . , pd,k ) is a basis for Sk . Let hf1 , . . . , fd i be the orthonormal Hecke basis for Sk . Then we have pm,k (z) = d X hpm,k , fi ifi (z), i=1 so 18 (7.9) d X afi (m)afi (n) apm,k (n) = Γ(k − 1) . (4πm)k−1 (7.10) i=1 In particular, the apm,k (n) are real. Also, d Γ(k − 1) X apm,k (m) = afi (m)2 . (4πm)k−1 (7.11) i=1 7.1 The Fourier Expansion We have 1+iy Z apm,k (n) = pm,k (z)e(−nz) dz iy (7.12) 1+iy X az + b e(mz) + = (cz + d)−k e m e(−nz) dz cz + d iy c>0 Z (7.13) (c,d)=1 X Z = δm,n + 1+iy az + b e m − nz (cz + d)−k dz. cz + d iy c>0 (7.14) (c,d)=1 Now az+b cz+d = a c − ad−bc c(cz+d) δm,n + = X Z a c 1 c(cz+d) , − 1+iy e iy c>0 so we get ma m − − nz (cz + d)−k dz c c(cz + d) (7.15) (c,d)=1 = δm,n + Z X e =(z)=y c>0 ma m − − nz (cz + d)−k dz c c(cz + d) (7.16) d∈(Z/cZ)× z7→z− dc = δm,n + X c>0 e ma nd + c c Z m (cz)−k e − 2 − nz dz. c z =(z)=y (7.17) d∈(Z/cZ)× Define the Kloosterman sums X K(m, n; c) = x∈(Z/cZ)× e mx + nx c . (7.18) xx=1 Also define the Bessel functions k Jk−1 (4πx) = (2πi) Z x z −k e −xz − dz. z =(z)=y 19 (7.19) So we have the formula √ k−1 X 4π mn K(m, n; c) 2 k n (2πi) apm,k (n) = δm,n + Jk−1 . m c c (7.20) c>0 7.2 Properties of Jk−1 and K • Jk−1 maps R to R. • Jk−1 (x) k min{xk−1 , √1x }. • |K(m, n; c)| ≤ c. k Corollary 7.5. apm,k (n) = O (n 2 + ). This is almost Hecke’s bound. More properties of K: 1. K(m, n; c) = K(n, m; c). 2. If (a, c) = 1, then K(am, n; c) = K(m, an; c). 3. If d|(m, n, c), then K(m, n; c) = ϕ(c) m n c ϕ( dc ) K( d , d ; d ). 4. If c = c1 c2 where (c1 , c2 ) = 1, then K(m, n; c) = K(mc1 , nc1 ; c2 )K(mc2 , nc2 ; c1 ) where, as before, the bar denotes multiplicative inverse. 5. An exercise (due to Selberg): K(m, n; c) = X dK mn d|(m,n,c) d2 , 1; c . d (7.21) As an example, consider k = 12. For m > 0, pm,12 (z) = cm ∆(z) where cm = hpm,12 , ∆i12 Γ(11) τ (m) = , h∆, ∆i12 (2πm)11 h∆, ∆i12 (7.22) (n) so apm,12 (n) = v τ (m)τ for some constant v. Now (5) implies m11 apm,k (n) = m1−k X dk−1 ap1,k d|(m,n) mn d2 , (7.23) so we find that τ (m)τ (n) = v −1 m11 am,12 (n) = X d|(m,n) recovering the Hecke identity. Now consider K(1, m; pa ) for p a prime. 20 d11 τ mn d2 , (7.24) Lemma 7.6. For p a prime, |K(1, m; p2a )| ≤ 2pa . Proof. x0 + pa y ≡ x0 − x0 2 pa y (mod p2a ), so X X x0 mod pa y mod pa 2a K(1, m; p ) = e x0 + mx0 p2a 1 − mx0 2 e y p2a (7.25) p-x0 =p X a e p-x0 x0 + mx0 p2a (7.26) x20 ≡m mod pa with the sum having at most two terms. 1 Lemma 7.7. |K(1, m; p2a+1 )| ≤ 2pa+ 2 for a > 0. Proof. x0 + pa+1 y = x0 − x0 2 pa+1 y, so by the same method as before, K(1, m, p 2a+1 )=p X a e x0 mod pa+1 , p-x0 x20 ≡m mod pa x0 + mx0 p2a+1 . (7.27) Now x0 + pa z = x0 − x0 2 pa z + x0 3 p2a z 2 , so we get p a X x0 mod pa e x0 + mx0 p2a+1 a X ! p p z(1 − mx0 2 ) + x0 3 p2a z 2 e . p2a+1 (7.28) z=1 x20 ≡m mod pa To get the desired bound, we use: P √ p nx2 × Lemma 7.8. For n ∈ (Z/pZ) , x=1 e( p ) = p. Proof. We have p 2 2 p X X nx2 nx − ny 2 e e . = p p x=1 (7.29) x,y=1 For p > 2, take a = x − y and b = x + y to get p X a,b=1 e nab p 21 = p. (7.30) It remains to deal with K(1, m; p). The Weil Conjectures imply K(1, m; p) = αp + βp where √ √ |αp | = |βp | = p, so |K(1, m; p)| ≤ 2 p. We’ll prove this result soon. k 1 The result is that apm,k (n) = O (n 2 − 4 + ). 1 3 Theorem 7.9. |K(1, a; p)| < 2 4 p 4 . Proof. Let Vp = Pp−1 a=1 |K(1, a; p)| 4. Note that ( −1 a 6≡ 0 (mod p) K(0, a; p) = p − 1 a ≡ 0 (mod p) (7.31) Now by property (2), (p − 1)Vp = p−1 X |K(a, b; p)|4 (7.32) |K(b, a; p)|4 − 2(p − 1) − (p − 1)4 . (7.33) a,b=1 = p−1 X a,b=0 The above sum equals p−1 X p−1 X e a,b=0 x,y,z,w=1 2 b(x + y − z − w) + a(x + y − z − w) p = p · #{x, y, z, w ∈ (Z/pZ)× : x + y = z + w, x + y = z + w}. (7.34) (7.35) This occurs if and only if either x = −y and z = −w, or x + y = z + w and xy = zw (implying {x, y} = {z, w}). We end up with #{ } = (p − 1)(2p − 5), so Vp = p2 (2p − 5) − 2 − (p − 1)3 < 2p3 . 8 (7.36) Congruence Subgroups ϕ ∗∗ Fix N ∈ N, and let Γ(N ) = ker(SL2 (Z) −−N → SL2 (Z/N Z)), Γ0 (n) = ϕ−1 N ({( 0 ∗ )}), and Γ1 (N ) = −1 ϕN ({( 10 ∗1 )}). Define Y (N ) = Γ(N )\H and define Y0 (N ) and Y1 (N ) similarly. 22 Fact. Y 1 1− 2 p p|N Y 1 1+ |Γ : Γ0 (N )| = N p p|N Y 1 2 1− 2 |Γ : Γ1 (N )| = N p |Γ : Γ(N )| = N 3 (8.1) (8.2) (8.3) p|N Lemma 8.1. Let F be a fundamental domainS for Γ, and γ1 , . . . , γd left representatives of Γ0 \Γ, for Γ0 a subgroup of Γ of finite index. Then F 0 = di=1 γi F is a fundamental domain for Γ0 . As an example, for Γ0 = Γ0 (N ), SL2 (Z/N Z) acts on P1 (Z/N Z) = {[x, y] : x, y ∈ Z/N Z, (x, y) = a b 1}/(scaling by (Z/N Z)× ). Then Γ0 (N ) = stab([0 : 1]). So Γ → Γ1 00(N )\Γ is c d = [c : d]. For the 0 −1 case N = p, we can take γi = 1 i for 1 ≤ i ≤ p and γp+1 = ( 0 1 ). To do Fundamental domain picture (5) Now look at the cusps. For H ⊆ P1 (C), ∂(H) = P1 (R), and SL2 (Z) acts transitively on P1 (Q). For Γ0 ⊆ SL2 (Z), Γ0 \P1 (Q) are the cusps of Γ0 \H. Suppose that c ∈ Γ0 \P1 (Q). Then there exists σc ∈ SL2 (Z) such that σc (∞) = c, so σc−1 F is a fundamental domain taking c to ∞ (and two arcs meeting at c to vertical lines). For Γc = Γ0 ∩ σc Γ∞ σc−1 , Γc is precisely the set of matrices in Γ0 preserving c. Define X0 (N ) = Y0 (N ) ∪ Γ0 (N )\P1 (Q). S Fact. • P1 (Q) = Γ0 (N ) · uv where the union is disjoint, taken over v|N , (u, v) = 1, and N u mod (v, v ). • Determine the genus of X0 (N ). To do (6) Now for k ∈ N, possibly odd, and Γ0 ⊆ Γ, define Mk (Γ0 ) to be the set of holomorphic f : H → C such that for every γ ∈ Γ0 , f |k γ 0 = f , and f is holomorphic at the cusps. That is, for any c ∈ Γ0 \P1 (Q), (f |k σc )(z) = X af,c n≥0 n nz e m m (8.4) for m = [Γ∞ : Γ∞ ∩ σc−1 Γ∞ σc ]. We say that f ∈ Sk (Γ0 ) if for every c ∈ Γ0 \P1 (Q), af,c (0) = 0. We know that Γ1 (N ) Γ0 (N ) with Γ0 (N )/Γ1 (N ) ∼ = (Z/N Z)× . So for χ : (Z/N Z)× → C× , let Mk (Γ0 (N ), χ) = f ∈ Mk (Γ1 (N )) : f |k L Then Mk (Γ1 (N )) = χ Mk (Γ0 (N ), χ). a b c d = χ(a)f ∀ a b c d ∈ Γ0 (N ) . (8.5) We can extend the construction of th Poincaré series as follows: for k > 2, define pm,N,k = 1 2 X (em |k γ)(z), pcm,N,k = γ∈Γ∞ \Γ0 (N ) 23 1 2 X (em |k σc−1 γ)(z). γ∈Γc \Γ0 (N ) (8.6) c . Denote pc0,N,k by EN,k Theorem 8.2. • c (c0 ) EN,k • If m > 0, then pcm,N,k ∈ Sk (Γ0 (N )). ( 1 = 0 c = c0 . otherwise Proof. First suppose c 6= c0 . Then lim=(z)→∞ (pcm,N,k |k σc0 )(z) = pm,N,k (c0 ). But we have (pcm,N,k |k σc0 )(z) = 1 2 X (em |k σc−1 γσc0 )(z). (8.7) γ∈Γc \Γ0 (N ) For c 6= c0 , 1 6∈ σc−1 Γ0 (N )σc0 , so 1 c (pm,N,k |k σc0 )(z) ≤ 2 X |jk (γ, z)| (8.8) γ∈Γ∞ \SL2 (Z) γ6=1 which tends to 0 as =(z) → ∞. Now if c = c0 , then (pcm,N,k |k σc )(z) = e(mz) + O X γ∈Γ∞ \SL2 (Z) |jk (γ, z)| . (8.9) γ6=1 The second sum tends to 0 as =(z) → ∞, while e(mz) tends to 1 if m = 0 and 0 if m > 0. L c · C. Corollary 8.3. For k > 2, Mk (Γ0 (N )) = Sk (Γ0 (N )) ⊕ c EN,k If f ∈ Sk (Γ0 (N )) and g ∈ Mk (Γ0 (N )), define Z N hf, gik f (z)g(z)y k = Γ0 (N )\H We have L c c EN,k dx dy . y2 (8.10) · C = Sk (Γ0 (N ))⊥ by unfolding. And we have the identity N hf, pm,N,k ik = Γ(k − 1) af (m) (4πm)k−1 (8.11) for every f ∈ Sk (Γ0 (N )). The Fourier expansion of pm,N,k is given by apm,N,k (n) = δm,n + P n>0 apm,N,k (n)e(nz), n k−1 X 2 m c>0 N |c 24 Jk−1 where √ 4π mn K(m, n; c) . c c (8.12) 1 Iwaniec’s trick: Sk ,→ Sk (Γ0 (N )) by inclusion. We have N kf kk = [Γ : Γ0 (N )] 2 kf kk . Let f 1 . Take {f1 , . . . , fd } an orthonormal basis for Sk . Then for some gi , Nf = [Γ:Γ0 (N )] 2 {N f1 , . . . ,N fd , g1 , . . . , gr } (8.13) is an orthonormal basis for Sk . Then pm,N,k Γ(k − 1) = (4πm)k−1 d X afi (m) i=1 [Γ : Γ0 (N )] 2 1 fi + r X ! . agi (m)gi (8.14) i=1 Taking mth Fourier coefficients, 1+ X Jk−1 c>0 4πm c d r X X |afi (m)|2 + |agi (m)|2 [Γ : Γ0 (N )] K(m, m; c) Γ(k − 1) = c (4πm)k−1 i=1 N |c (8.15) m δ−1 Nm |afi (m)|2 , N 1+o(1) or 1 + mδ N 1 2 so (N + mδ ) |afi (m)|. For the left hand side, the length of the sum is about −1 2 m m would expect the sum to behave as √ mN = N 9 . i=1 Theorem 8.4. If |K(m, m; c)| ≤ cδ for some δ < 1, then 1 + |afi (m)|2 , N 1+o(1) ! 1 and each term is about m− 2 , so we m N, √1 . N The Theta Function √ P πim2 z , we had θ (z + 2) = θ (z) and θ (− 1 ) = Recall that for θ0 (z) = −izθ0 (z). Fix 0 0 0 m∈Z e z γ = ac db ∈ SL2 (Z) with a ≡ b ≡ 0 (mod 2). Then θ0 ( az+b ) is given by cz+d θ0 a 1 − c c(cz + d) = 2 X e 2 iπm iπ mc a − c(cz+d) = X eπi m2 a c m 2 c cz+d X m 2 e−iπ( c +t) c cz+d . (9.2) t∈Z m mod c Now let f (t) = e−iπ( c +t) (9.1) m∈Z 2 . Then by scaling and shifting by e−iπt , 1 1 ms d 2 fb(s) = (ic)− 2 (cz + d) 2 e2πi c +iπs (z+ c ) . (9.3) Apply Poisson summation to get θ0 az + b cz + d 1 1 = (ic)− 2 (cz + d) 2 X m mod c Now for a = 2α, 25 eiπ m2 a c X s∈Z e2πi ms +iπs2 (z+ dc ) c . (9.4) X eiπ m2 a +2πi ms c c m mod c αm2 + ms c m mod c 2 X m + ms m=αn e α = c m mod c X α(m + s · 2)2 4s2 α = e e − c c m mod c X αm2 −iπ s2 d ad≡1 mod c c = e e c m mod c X s2 d αm2 = e e−iπ c . c X = e (9.5) (9.6) (9.7) (9.8) (9.9) m mod c Fix Me Maybe deal with the overset alignments? (7) Let β(α, c) = z 7→ − z1 , we get αm2 m mod c e( c ). P θ0 1 1 −2 Then θ0 ( az+b (cz + d) 2 β(α, c)θ0 (z). After mapping cz+d ) = (−ic) bz − a dz − c = β(α, c) c 1 2 1 (dz − c) 2 θ0 (z). (9.10) We will now compute the Gauss sum β(α, c). We have the following properties: • If c = c1 c2 with gcd(c1 , c2 ) = 1, then β(α, c) = β(αc1 , c2 )β(αc2 , c1 ). • If c = pn with p prime and n > 1, then β(α, pn ) = pβ(α, pn−2 ). • If gcd(x, p) = 1, then β(αx2 , c) = β(α, c). • If x is such that xp = −1, then β(1, p) + β(x, p) = 0. So it suffices to determine β(1, p). Theorem 9.1. Sketch. ( 1 p2 β(1, p) = 1 ip 2 p ≡ 1 (mod 4) p ≡ 3 (mod 4) (9.11) Fix Me We may have confused notation. (8) 2 This proof works even for composite p. Let f (x) = e( xp ) for 0 ≤ x ≤ p and 0 otherwise. Then we have p X i=0 # p−1 X X Z p 1 1 = f (0) + f (i) + f (p) = lim f (x)e(−mx) dx N →∞ 2 2 0 !" 0 f (i) i=1 |m|≤N so that 26 (9.12) p X Z β(1, p) = lim N →∞ = lim N →∞ 0 |m|≤N X e x2 − mx p dx (9.13) 1 Z e(p(m2 − mx)) dx p (9.14) 0 |m|≤N Z 1 m2 p m 2 pe − = lim e p x− dx N →∞ 4 2 0 |m|≤N Z 1− m X 2 m2 p pe − = lim e(px2 ) dx m N →∞ 4 − 2 |m|≤N Z ∞ e(px2 ) dx = p(1 + i−p ) −∞ Z ∞ √ = p(1 + i−p ) e(x2 ) dx. X (9.15) (9.16) (9.17) (9.18) −∞ By plugging in p = 1, we get R∞ 2 −∞ e(x ) dx i 1+i , = −p √ so β(1, p) = i p 1+i 1+i . √ So forodd c, let c be 1 for c ≡ 1 (mod 4) and i for c ≡ 3 (mod 4). Then β(α, c) = c c where αc is the Jacobi symbol. Therefore if b ≡ c ≡ 0 (mod 2), θ0 az + b cz + d = 2c d −1 1 d (cz + d) 2 θ0 (z). Now let θ(z) = θ0 (2z). Then ∀, γ ∈ Γ0 (4), if γ = 1 2 where j 1 (γ, z) = dc −1 d (cz + d) . a b c d α c (9.19) , then we have θ(γz) = j 1 (γ, z)θ(z), 2 2 For k ∈ Z, 4|N , we define M k (Γ0 (N ), χ) to be the set of f : H → C such that f (γz) = 2 χ(γ)j 1 (γ, z)k f (z), and f is holomorphic on H and at the cusps. 2 10 Quadratic Forms Now suppose Q is a positive definite quadratic form on Zg such that ∀ v ∈ Zg , Q[v] ≡ 0 (mod 2). Theorem 10.1. For N ∈ N, A a positive definite symmetric matrix with even diagonal entries, such that N A−1 has integral entries (so (Zg )∗ ⊆ N −1 Zg , with the dual being with respect to A), then define θQ (z) = X e m∈Zg 1 Q[m]z 2 where Q[v] = v T QV . Then θQ ∈ M g (Γ0 (2N ), χQ ) for χQ 2 Let rQ (n) = #{m ∈ Zg : Q[m] = 2n}. Then 27 (10.1) a b c d = 2g ·|A| d . θQ (z) = X rQ (n)e(nz). (10.2) n≥0 We have that g 1 X 1 rQ (n) = #{m ∈ ZG : Q[m] ≤ 2M } ∼ cQ M 2 −1 . M M (10.3) n≤M But if f is a cusp form of weight g2 , then g 1 1 X |af (n)| ∼ M 4 − 4 . M (10.4) n≤M We’ll later find Siegel’s mass formula, which determines the Eisenstein series part of θQ (z), so it would remain to deal with the cusp form part. P 2 As an exercise, n≡a (b) q n is a modular form of weight 21 for some congruence group Γ. The P 2 same holds for n≡a (b) q rn for r ∈ Q, r > 0. So a product of k such sums will be a modular form of weight k2 . We can treat quadratic forms either as a positive definite matrix Q in Zk , or a lattice L in Rk . (Here L = Q[Zk ].) If L is a lattice, then θL = X 1 q 2 hv,vi = v∈L X exp(−πthv, vi) (10.5) v∈L for q = e2πiz and z = it. To be invariant under some integer translation, we want our quadratic form to have rational values. θL also satisfies a Poisson summation formula: in general, X 1 Fb(v ∗ ) vol(Rk /L) ∗ ∗ v ∈L v∈L Z Fb(x∗ ) = e(hx∗ , xi)F (x) dx. X F (v) = (10.6) (10.7) Rk We obtain θL = X 1 − k2 t exp(−(πt−1 hv ∗ , v ∗ i)). vol(Rk /L) ∗ ∗ (10.8) v ∈L We can also consider the restricted sum over a “shifted lattice” v ≡ v0 (mod L0 ) where [L : L0 ] < ∞. Also, we have θL1 ⊕L2 = θL1 θL2 . The exercise will show that shifted rectangular lattices are modular forms. But then any lattice L has a rectangular sublattice L0 of finite index. Then choose appropriate L0 . 28 The same will hold for lattices). As an example, P L ω(v)q 1 hv,vi 2 1 η = q 24 for ω(v) L0 -periodic (this is a linear combination of shifted ∞ Y (1 − q n ) = n=1 1 2 1X χ(n)q 24 n 2 (10.9) n∈Z for some χ by Euler’s pentagonal number formula. We have X v ω(v)F (v) = X 1 ω b (v ∗ )Fb(v ∗ ) |Rk /L0 | ∗ ∗ (10.10) v ∈L0 for ω b the discrete Fourier transform. Also, 1 η 3 = q 8 (1 − 3q + 5q 3 − 7q 6 + 9q 10 − · · · ) = for some character χ. Also, this shows that in weight expected weight. 3 2 1 2 1X χ(n)nq 8 n 2 n (10.11) or 12 , coefficients can have larger than Fix Me This part appears in the notes after equidistribution. May want to organize differently? (9) Let Q be a positive definite quadratic form over Z in d variables. For each m ∈ N, let Sm = {v ∈ d−1 given by v 7→ √ v . This can either be Zd : Q(v) = m}. We have a measure ϕ : Rd \ {0} → SQ Q(v) d−1 induced from the metric given on Q, or the µ-unique invariant thought of as the µ-measure on SQ measure under OQ (R) = {A ∈ GLn (R) : ∀ v, Q(Av) = Q(v)}. (10.12) Then: d Theorem 10.2. Given {mi } ∈ N, assume rQ (mi ) mi2 equidistributed with respect to µ. −1− , and d ≥ 4. Then ϕ(Smi ) become For R any ring, a quadratic form over R is a symmetric matrix Q ∈ Md (R), with Q(v) = v T Qv. We say that Q1 ∼R Q2 if there exists γ ∈ GLn (R) such that γQ1 γ T = Q2 . Theorem 10.3 (Hasse-Minkowski). If Q1 , Q2 are quadratic forms over Q in d variables, then Q1 ∼Q Q2 if and only if Q1 ∼R Q2 and Q1 ∼Qp Q2 for every prime p. For Q1 , Q2 quadratic forms over Z in d variables, we say that Q1 ∈ genus(Q2 ) if Q1 ∼R Q2 and Q1 ∼Zp Q2 for every prime p. Remark. This is weaker than equivalence. For example, consider Q1 = ( 1 14 ) and Q2 = ( 2 7 ). Theorem 10.4. For any quadratic form Q over Z, genus(Q) (as a set of equivalence classes) is finite. 29 e ∈ genus(Q), then det Q e = det Q. Then use Minkowski’s Proof in case Q is positive definite. If Q d e e induction theory: Pick v1 ∈ Z such that Q(v1 ) = minv∈Zd \{0} Q(v), and inductively choose vi+1 = e minv∈Zd \{0} Q(v) such that hv1 , . . . , vi+1 i is a primitive sublattice of Zd (can be extended to a basis). Fact. e d det Q d Y e i ). Q(v (10.13) i=1 e j | are all bounded, and determine Q. e As they are integers, there are finitely As a result, |viT Qv many options. Now let L be a lattice over Z and Q a quadratic form on L. For each p, let Lp be LT ⊗Z Z(p) , a lattice over Z(p) , and all contained in the Q-space LQ = L ⊗Z Q. We can recover L as p Lp . For L0 ⊆ LQ another lattice, we have L0p = Lp ∀0 p (for all but finitely many p). Now (L, Q) determines some integral quadratic form, and so does (L0 , Q) if Q(L0 ) ⊆ Z. We have (L0p , Q) ∼Zp (Lp , Q) if and only if there exists γp ∈ OQ (Qp ) such that γp Lp = L0p . Likewise, (L0 , Q) ∼Z (L, Q) if and only if there exists γ ∈ OQ (Q) such that γL = L0 . Q 0 Let OQ (Af ) = 0p OQ (Qp ) (consisting T of (γ2 , γ3 , γ5 , . . .) such that γp ∈ OQ (Qp ), and ∀ p, γp ∈ OQ (Zp )). For γ ∈ OQ (Af ), let γ · L = p (γp · Lp ). We have that - , genus(L) = OQ (Q) OQ (Af ) ! Y OQ (Zp ) (10.14) p via γ · L →7 γ. Given (L, Q), define 1 #Aut(L, Q) (L) w(L) = P o(L) = L0 ∈genus(L) o(L Let rgenus(L,Q) (n) = (10.15) 0) . (10.16) 0 L0 ∈genus(Q) w(L )rL0 (n). P Theorem 10.5. Let (L, Q) be a Z-lattice with a positive definite quadratic form, and n > 0 such that for every p, there exists vp ∈ Lp such that Q(vp ) = n. Then rgenus(L,Q) 6= 0. Proof. By Hasse-Minkowski, there exists v ∈ LQ such that Q(v) = n. ∀0 p, v ∈ Lp . Consider wp = v − vp , and let γp (w) = w − 2 wT Qwp wp . Q(wp ) Then γp ∈ OQ (Zp ) with γp (vp ) = v. So v ∈ γp · Lp , implying v ∈ We have γp = 1 if v ∈ Lp , so L0 ∈ genus(L). Define 30 (10.17) T p γp · Lp ; define this to be L0 . θgen(L) (z) = X w(L0 )θL0 (z) = L0 ∈gen(L) X rgen(L) (n)e(nz). (10.18) n≥0 1. If L0 ∈ gen(L), then θL (z) − θL0 (z) is a cusp form. Theorem 10.6 (Siegel-Weil). 2. θgen(L) (z) lies in the Eisenstein spectrum. That is, for all cusp forms f , hθgen(L) (z), f (z)i = 0. Q 3. rgen(L) (m) = r (m) · r∞ (m), where: p p • For prime p and d = dim L, #{v ∈ L/pr L : Q(v) ≡ m r→∞ pr(d−1) rp (m) = lim (mod pr )} = volZp (Q−1 (m)). (10.19) (The limit actually stabilizes.) d • r∞ (m) = volR (Q−1 (m)) = m 2 −1 volR (Q−1 (1)). The product converges absolutely for d ≥ 4, and conditionally for d = 2, 3. P Remark. (2) =⇒ (3), theoretically speaking. For we then have θgen(L) (z) = c ac Ec (z) for some ac , and we know the Fourier coefficients of Ec . We can also show that (3) =⇒ (2), by the theory of Hecke operators. Pick p - 2 det Q, and m such that 2 det(Q) - m. Then rgen(L) (p2 m) #{Fp : Q(v) = m} = pd−2 · . rgen(L) (m) #{Fp : Q(v) = 0} (10.20) This shows that θgen(L) is, ∀0 p, an eigenfunction for the p2 Hecke operator with a very large eigenvalue. P As an example, take Q = Id , so Q(x1 , . . . , xd ) = di=1 x2i . Pick p ≡ 3 (mod 4). The number of Fp solutions to Q(v) = m is given by: m d\ 1 2 3 4 5 2 (F× p) 0 1 2 1 p+1 p2 p2 − p p3 + p2 − p p3 − p ··· p4 − p2 × 2 F× p \ (Fp ) 0 p−1 p2 + p p3 − p p4 + p2 (10.21) e = Q + x2 + y 2 , we have Observe that for Q rQe (m) = X rx2 +y2 (n) · rQ (m − n) (10.22) m∈Fp = (p + 1) X n d rQ (n) − prQ (m) = pd+1 + p − prQ (m). 31 (10.23) (10.24) It then follows that ∀0 p, p d−1 rp (m) = d d d−1 − (−1) 2 p 2 −1 p d−1 d−1 pd−1 + (−1) 2 p 2 2|d m p (10.25) 2 - d. d For d ≥ 4, rp (m) = 1 + O(p−2 ). To show that rgen(Q) (m) m 2 −1− , we need to consider p| det Q. Given x0 , y0 , z0 with Q(x0 , y0 , z0 ) ≡ m (mod pr ), we want to lift Q(x0 +pr x, y0 +pr y, z0 +pr z) ≡ Q(x0 , y0 , z0 )+pr Qx (x0 , y0 , z0 )x+pr Qy (x0 , y0 , z0 )y+pr Qz (x0 , y0 , z0 )z. (10.26) If (x0 , y0 , z0 ) 6= 0, we have exactly p2 lifts. For p| det(Q), m, rp (m) p−m` for ` such that p` km. For n ≥ 5, rp (m) p−n . n Corollary 10.7. For n ≥ 5, rgen(Q) (m) m 2 −1− if m is locally represented. The same holds for n = 4 and squarefree m. As a consequence, |rgen(Q) (m)| → ∞ implies equidistribution for n ≥ 4. Let A ∈ Mn (Z) such that AT = A, and v T Av > 0 for v 6= 0. Fix N ∈ N such that N A−1 ∈ Mn (Z), and assume diag(A) ≡ 0 (mod 2). Then A determines a quadratic form. Let A[v] = v T Av, and define θA (z) = X e v∈Zn 1 A[v]z , 2 (10.27) and H ⊆ Zn /N Zn consisting of h such that Ah ≡ 0 (mod N ). (So then (Zn )∗ = define the theta function X θA (z; h) = v≡h (N ) e A[v]z 2N 2 HZn N .) For h ∈ H, . (10.28) Then θA (z; 0) = θA (z). Theorem 10.8. For h ∈ H, θA T n X 1 h A` − 21 e − ; h = det(A) (−iz) 2 θA (z; `). z 2N 2 (10.29) `∈H Observe that h`, hi 7→ e Proof. Let f (x) = e 1 2 A[x]z `T Ah 2N 2 is well-defined and nondegenerate on H. . Then Z fb(y) = e(−xT y)f (x) dx Rn − 12 = det(A) n T i 2 h y e . z N 32 (10.30) (10.31) Poisson summation for f h yields N X v∈Zn e n X i 2 1 h A−1 [v] hT v − 21 z = det(A) . A v+ e − + 2 N z 2z N n (10.32) v∈Z So − 12 θA (z; h) = det(A) n X X A−1 [v] hT Av i 2 e − + . z 2zN 2 N2 (10.33) `∈H v≡` (N ) Theorem 10.9. For B also a nondegenerate symmetric positive definite matrix, B ∈ gen(A) ⇐⇒ ∀ d ∈ Z, ∃ γ ∈ GLn (Z/dZ) such that γAγ T ≡ B (mod d). Proof. We have N B −1 ∈ Mn (Z) by assumption. We can assume that A ≡ B (mod N ). Then: Claim. ∀ c ∈ P1 (Q) and ∀ ` ∈ H, ((θA (; `) − θB (; `))| n2 σc )(∞) = 0, (10.34) where σc ∈ SL2 (Z) such that σc (∞) = c. 1 Proof of claim. For c = ∞, θA (∞; `) = δ` = `). Now T go from c to − c by the transformation θTB (∞; formula and the fact that for all ` ∈ H, e `2NAh = e `2NBh . Finally, we can go from c to c + 1 2 2 T and similarly for B. since θA (z + 1; `) = θA (c; `)e `2NA` 2 In fact, we just needed the case d = N . This is a reflection of the following: Theorem 10.10. If A, B are positive definite quadratic forms with det(A) = det(B) and A ≡ B (mod Z/ det(A) det(B)Z), if diag(A) is even, then A ∈ gen(B). Proof of modularity of θA (z). Take V = hθA (z; `)i`∈H ∼ = L2 (H). We have a right action of B by SL2 (Z), by θA (z; `)|γ = θA (z; `)| n2 γ. We have f |( 1 1 ) (`) = f (`)h`, `i 01 1 X f (h)h`, −hi f| 0 1 (`) = 1 −1 0 |H| 2 h∈H (10.35) = fb(`). (10.37) (Here h`, hi is as before.) Then we have 33 (10.36) (H, h , i) ∼ = n M i=1 for ha, bi = e ab ni (Z/ni Z, h , i) | {z } (10.38) ϕni . a b c d Claim. Γ(2ni ) ⊆ ker ϕni , and if ∈ Γ0 (2ni ), then γ · δ0 = d ni · δ0 . Proof. Consider θ(ni z). Claim. For ac db ∈ Γ0 (4N ), az + b θ N· cz + d = j1 2 d a b ,z · θ(N z). c d N (10.39) Proof. az + b θ N· cz + d =θ a(N z) + bN (c/N )(N z) + d . (10.40) By the modularity of θ(z), we get 1 2 θ(N z) = d (cz + d) · (This trick works since 11 N 1 a b c d c/N d N −1 = θ(N z) · j 1 2 1 = a b/N c/N d d a b ,z . c d N (10.41) .) Harmonic Polynomials Let Σ be the unit sphere in Rk , and let Sn = { √vn : v ∈ L, hv, vi = n, possible congruence conditions}. In general, take Sn finite nonempty subsets of Σ. We say that the Sn are equidistributed as n → ∞ n ∩R) if the Sn are nonempty and for every open ball R ⊆ Σ, #(S → µ(R) as n → ∞. Equivalently, #Sn for every f ∈ C(Σ, R), Z 1 X f (v) → f dµ. #Sn Σ v∈Sn It actually suffices to check for polynomial f . Recall the operators 34 (11.1) X ∂2 ∂x2i i X ∂ E= xi ∂xi i X F = x2i . ∆= (11.2) (11.3) (11.4) i Let Pd denote the homogeneous polynomials of degree d, and Pd0 = ker(∆ : Pd → Pd−2 ) = ker(F ∆ : Pd → Pd ). (11.5) Theorem 11.1. Pd decomposes as Pd = M 0 F k Pd−2k (11.6) k 0 where each F k Pd−2k is an eigenspace of F ∆, with different eigenvalues. Proof. We will show that F ∆|F k P 0 d−2k = λk (d) where the λk (d) are distinct. But also 0 0 ≥ dim Pd−2k − dim Pd−2k−2 , dim F k Pd−2k = dim Pd−2k (11.7) so a dimension count forces the sum to be all of Pd . Also, equality holds everywhere. In particular, ∆ : Pd → Pd−2 is surjective. We can check easily that [∆, F ] = 4E + 2n. So if f is harmonic of degree d, then F ∆(F f ) = (4d + 2n)F f . More generally, if (F ∆)f = λf , then (F ∆)(F f ) = (4d + 2n + λ)F f . We end up with λk (d) = k(4(d − k − 1) + 2n). (We also have [E, ∆] = −2∆ and [E, F ] = 2F .) Because of Theorem 11.1, it suffices to check equidistribution for the Pd0 , because F is the identity map on We now claim that if P is harmonic of degree d and L is a rational lattice, P the unit sphere. 1 then v∈L P (v)q 2 hv,vi is modular (with congruence group depending on L) of weight d + 21 rank(L), and is a cusp form if d > 0. Proposition 11.2. Define Gt (P ) = eπthx,xi P . (Then Gt F = F Gt .) If f = Gt P for P ∈ Pd , then P n 0 0 fb = G 1 Pb where Pb = d0 ≤d Pbd0 with Pbd0 = id t−( 2 +d ) P . t Fix Me Are they both d0 , or does d appear? (10) ∂ b \ \ Proof. We have G t xj P = xj Gt P = c ∂xj (G 1 P ) for appropriate c. Then use induction. t Now suppose P ∈ Pd0 . Then ∆P = 0 and EP = dP . We have 0 = (Gt ∆G−1 t )Gt P . We can show −1 −1 that Gt F G−1 = F , G EG = E + 2πtF , and G ∆G = ∆ + πt(4E + 2n) + (2πt)2 F . We can then t t t t t n −( +d) d 0 check that Pb = i t 2 P ⇐⇒ P ∈ Pd . 35 R Now Rn e−πthx,xi P (x) dx = cPb(0) = 0 if P is harmonic of degree at least 1, so the Sj are equidistributed if and only if for every nonconstant harmonic polynomial P , 1 X P (v) → 0. |Sj | (11.8) j d 1 1 X P (v) = · j − 2 · (coefficient of q 2 ). |Sj | |Sj | (11.9) v∈Sj But we have that v∈Sj d n So we need this coefficient to be o(|Sj | · j 2 ). If we have that |Sj | = j 2 −1− and that the q j n d coefficient of a cusp form is o(j 2 −1−+ 2 ), then we are done. n d The trivial bound o(j 4 + 2 ) suffices if n ≥ 5. For n = 4, any power estimate better than the n d 1 trivial bound will suffice. (The Kloosterman bound is o(j 4 + 2 − 4 + ).) For n = 3, any power estimate better than the Kloosterman bound would suffice. Claim. Let p be an odd prime. Then γ= p−1 X j=0 2πij 2 e p ( √ X a 2πia p p≡1 p = e = √ p i p p≡3 a (p) (mod 4) (mod 4). (11.10) 2πijk Proof. Let V = (e p )p−1 j,k=0 , the discrete Fourier transform modulo p. Our desired expression is √ tr(V ). We can show that the spectrum is p · (1, i, −1, −i, 1, i, . . . , ip−1 ). To get the spectrum, we first know that 1 1 1 1 1 2 V = p (11.11) 1 which is p times a matrix with square the identity and trace 1, so V 2 has spectrum p with √ p−1 multiplicity p+1 2 and −p with multiplicity 2 . This implies V has spectrum {± p} with multiplicity p−1 √ p+1 p−1 2 2 p shows the spectrum is as close to 2 and {±i p} with multiplicity 2 . Then γ = (−1) balanced as possible. Finally, det V = Y k j −e . e p p j<k Now e(α) − e(β) has argument π 2 + 12 (α + β). 36 (11.12) 12 Bounds for Fourier Coefficients Recall that K(m, n; c) = formula, mx+nx c P x (c) e , and that for f ∈ Sk (SL2 (Z)), the Petersson trace X |af (m)|2 Γ(k − 1) positive terms + Jk−1 ≤ 1 + (4πm)k−1 [Γ : Γ0 (N )] · kf k c>0 4πm c K(m, m; c) . c (12.1) N |c Pick large Q and sum over Q ≤ N ≤ 2Q. We have [Γ : Γ0 (N )] = N 1+o(1) , so 2Q X N =Q 1 = Qo(1) . [Γ : Γ0 (N )] (12.2) So we find that 1−k m 2 o(1) af (m) Q 2Q X ∞ X 1 4πm ≤Q+ Jk−1 K(m, m; qr). qr qr (12.3) q=Q r=1 Lemma 12.1. 1 X |K(am, am; r)|2 ≤ 4 gcd(m, r)σ0 (r) · r. r (12.4) a (r) Proof. LHS = 1 r X e a,x,y (r) amx + amx + amy + amy r (12.5) xx=yy=1 = #{x, y (r) : mx + mx + my + my ≡ 0 (r)} (12.6) = #{x, y (r) : m(x + y)(xy + 1) ≡ 0 (r), gcd(xy, r) = 1}. (12.7) Sum over prime powers and count. Now if gcd(q, r) = 1, K(m, m; qr) = K(mq, mq; r)K(mr, mr; q). Let A(Q, R) = X X |K(mq, mq; r)|2 (12.8) q≤Q r≤R (q,r)=1 ≤ X Q r≤R ≤ X r +1 X |K(am, am; r)|2 (12.9) q (r) 4(Q + r) gcd(m, r)rσ0 (r) (12.10) r≤R ≤ (Q + R)m R2+ . 37 (12.11) Next, let B(Q, R) = X X |K(m, m; qr)| (12.12) |K(mq, mq; r)||K(mr, mr; q)| (12.13) q∈Q r≤R (q,r)=1 = X X q≤Q r≤R (q,r)=1 1 1 ≤ A(Q, R) 2 B(R, Q) 2 (12.14) ≤ (Q + R)m Q1+ R1+ (12.15) where the inequality (12.14) follows from Cauchy-Schwarz. Finally, let C(Q, R) = XX |K(m, m; qr)| (12.16) q≤Q r≤R ≤ X X X a≤Q q0 ≤ Q r0 ≤ R a a |K(m, m, q0 r0 a2 )| (12.17) |K(m, m, q0 r0 )|a1+ (12.18) (q0 ,r0 )=1 ≤ ≤ X X X a≤Q q0 ≤ Q a X B a≤Q 1+ r0 ≤ R a (q0 ,r0 )=1 Q R , a a a1+ (12.19) R1+ (Q + R)m . (12.20) X X |K(m, m; qr)| ≤ (Q + R)1+ m . qr (12.21) ≤Q By dyadically summing, we find that q≤Q r≤R 1 We also have Jk−1 (x) min{xk−1 , x− 2 }, and 1−k m ∞ XX 4πm m 1+ 1 Jk−1 |K(m, m; qr)| ≤ Q + Q + |af (m)| Q + m + O(1). qr qr Q 2 q≤Q r=1 (12.22) 1 2 Taking Q = m , m1−k |af (m)|2 m 13 1 + 2 , so |af (m)| m k − 41 + 2 . Metaplectic Groups and Half-Integral Weight Hecke Operators (Reference: Shimura, “On Modular Forms of Half-Integral Weight”.) Define the group 38 e= G cz + d a b + 2 (α, ϕ) : α = ∈ GL2 (Q), ϕ holomorphic on H with ϕ(z) = ± √ . c d det α (13.1) Then (α, ϕ(z))(β, ψ(z)) = (αβ, ϕ(βz)ψ(z)) defines a group multiplication. We have an exact sequence e → GL+ (Q) → 1. 1 → µ4 → G 2 (13.2) e does not arise from an algebraic group. But G e by i(γ) = (γ, j 1 (γ, z)). If 4|N and k ∈ 1 Z, define (f |k ge)(z) = f (g(z))ϕg (z)−2k , Define i : Γ0 (4) ,→ G 2 2 for ge = (g, ϕg ). Then f ∈ Mk (Γ0 (N )) iff f |k i(γ) = f for every γ ∈ Γ0 (N ). 0 e let Γ^ Now if α e ∈ G, e−1 Γ0 (N )e α ∩ Γ^ 0 (N ) = i(Γ0 (N )) and Γ = α 0 (N ). 0 Claim. [Γ^ 0 (N ) : Γ ] < ∞. (We omit the proof of the claim.) Sr Pr 0 e . Let f | [e So we can write Γ^ eγ ei . 0 (N ) = i k α] = i=1 Γ γ i=1 f |k α Claim. f 7→ f |k [e α] sends Mk (Γ0 (N )) → Mk (Γ0 (N )). (Also, this is well-defined on Mk (Γ0 (N )).) For k ∈ Z, Tp = [e α] where α e= p 0 − 12 ,p . 0 1 (13.3) 1 0 Theorem 13.1. Let αm = ( m em = (α, m− 4 ). Then unless k ∈ Z or m is a perfect square, 0 1 ) and α f |k [e αm ] = 0. Proof. −1 Γ (N )α , and let γ ∈ • Take γ ∈ Γ0 (N ) ∩ αm 0 m −1 i(αm γαm ) a b c d . Then we have 1 c/m 2 = +d d m d 1 c m z 2 −1 = αm γαm , c + d d m d d m −1 =α em i(γ)e αm · 1, . d −1 αm γαm , cz (13.4) (13.5) (13.6) −1 Γ (N )α such that i(γ) 6∈ Γ0 . So if m is not a perfect square, then there exists γ ∈ Γ0 (N )∩αm 0 m −1 Γ ^ • i(γ)e αm αm i(γ)−1 is equal to 0 (N )e m m −1 −1 −1 −1 −1 ^ ) ·α em = α em Γ0 (N )e αm , α em i(αm γαm ) 1, · Γ^ · i(αm γαm 0 (N ) · 1, d d −1 Γ ^ so i(γ) stabilizes α em αm , and therefore Γ0 , under conjugation. 0 (N )e 39 (13.7) • We have f |k [e αm ] = r X f |k α em γ ei (13.8) f |k α em · i(γ) · γ ei (13.9) i=1 = r X i=1 = r X −1 f |k i(αm γαm )e αm · (1, −1) · γ ei , (13.10) f |k α em · γi = −f |k [e αm ]. (13.11) i=1 which for k 6∈ Z, is − r X i=1 Now define 2 n 0 − 12 ,n Tn2 f = f |k . 0 1 (13.12) ^ e−1 As an exercise, we have that Γ^ Γ^ (N )e αp2 ) decomposes as 0 (N )/(Γ0 (N ) ∩ α p2 0 2 −1 p[ b=0 If f (z) = P∞ n=0 af (n)e(nz), p−1 [ p h 1 −h 1 b −1 , p2 ∪ , p 0 p2 0 p p h=1 2 1 p 0 ∪ , p− 2 . 0 1 then we have (Tp2 f )(z) = 2 bf (n) = af (np ) + −1 p P∞ n=0 bf (n)e(nz), (13.13) (13.14) where k− 1 2 n n k− 32 k−2 ·p · af (n) + p af . p p2 (13.15) (At least this is true if gcd(p, N ) = 1.) For f, g ∈ Sk (Γ0 (N )), define Z f (z)g(z)y k hf, gik = Γ0 (N )\H dx dy . y2 (13.16) Then the Tp2 are self-adjoint with respect to the h , ik . So the Tp2 are simultaneously diagonalizable. Suppose f ∈ Sk (Γ0 (N )) is such that for every p, Tp2 f = up f . Then for t ∈ N squarefree, 40 ∞ X af (tn2 ) n=1 ns = af (t) · 1− Y 1 − up · p t p p−s 3 pk− 2 −s + p2k−2−2s . (13.17) Theorem 13.2 (Waldspurger,Shimura,Shintani,Konen-Zagier). • Let k ∈ 21 + Z, k ≥ 32 , and f ∈ Sk (Γ0 (N )) such that for all p, Tp2 f = up f . Then there exists g ∈ M2k−1 (Γ0 (N/2)) such that for every p, Tp g = up g. So then ∞ X ag (n) i=1 ns Y = 1 − up p p−s 1 . + p2k−1−2s (13.18) • Let D be squarefree and gcd(D, N ) = 1. Define L(g ⊗ χD , s) = ∞ X ag (n) n=1 n D ns+k−1 . (13.19) Then k − 32 ! |D|k−1 |af (D)|2 1 = · L g ⊗ χD , . hf, f ik π hg, gik 2 (13.20) So Ramanujan for half-integral weight is equvialent to Ramanujan for integral weight along with Lindelof in the D-aspect. We have g ∈ S2k−1 (Γ0 (N/2)) if k ≥ 52 , or if k = functions. 3 2 and f is orthogonal to one-variable theta For a, b ∈ N and χ : (Z/bZ)× → C× , define θa,χ (z) = ∞ X nχ(n)e(an2 z). (13.21) n=1 Then θa,χ (z) ∈ S 3 (Γ0 (4ab2 ), ψ) for ψ(d) = 1 2 2 3/2 − 14 2 −a d χ(d). The coefficients, at their largest, have size roughly n = n . Under the Shimura correspondence, they correspond to Eisenstein series, and Ramanujan fails for these. 14 Bounds in the Half-Integral Weight Case Fix Me Disgtinguish between p and P . (11) Recall that pm k,N (z) = 1 2 X (em |γ )(z) γ∈Γ∞ \Γ0 (N ) for em (z) = e(mz). If f ∈ Sk (Γ0 (N )), hf, pm k,N i is a constant times af (m), specifically 41 (14.1) Γ(k − 1) af (m). (4πm)k−1 hf, pm k,N i = (14.2) So it suffices to bound coefficients for Poincaré series. We have Z i+1 pm k,N (z)e(−nz) dz nth term = i Z = i+1 i (14.3) 2k az + b X −k −2k c e(−nz) dz (14.4) (cz + d) d e m d cz + d X e(mz) + (d,c)=1 c>0 c≡0 (N ) X = δm,n + X c>0 −2k d (d,c)=1 c 2k Z i+∞ d i−∞ X −k (cz + d)−k e c≡0 (N ) d (c) = δm,n + m k−1 2 n −k (2πi) c Jk−1 c>0 c≡0 (N ) ma m − − nz c c(cz + d) dz (14.5) √ 4π mn X d md + nd e · d−2k . c c c d (c) | {z } denoted Sk (m,n;c) (14.6) Lemma 14.1. If c = (q, r) and 4|r, then Sk (m, n; c) = Sk−q+1 (mq, nq; r)S(mr, nr; q) for X d md + nd e . c c S(m, n; c) = (14.7) d (c) (d,c)=1 This lemma is an exercise. S(m, n; c) is called the Salié sum. Lemma 14.2. If gcd(m, q) = gcd(n, q) = 1, then (in addition to S(m, n; q) = S(n, m; q)): 1. S(m, n; q) = m q 2. S(1, m; q) = 0 if S(1, mn; q). m q = −1. √ P 3. S(1, n2 ; q) = q q x2 ≡1 Proof. (q) e 2xn q . 1. Use the change of variables d → dm. 2. S(1, m; q) = S(m, 1; q) = m q S(1, m; q). 3. Define X x x + n2 x h(n) = e . q q x (q) 42 (14.8) Then we have mn h(n)e − q n (q) X x x X x(n2 − xmn) = e e q q q n x (q) X x x xm2 √ x e e − q q = q q 4 q x m2 x √ X = q q 1− . e q 4 b h(m) = X (14.9) (14.10) (14.11) (14.12) (x,q)=1 So √ b h(m) = q q X q µ d d|(q,m2 −4) d. (14.13) Fourier inversion implies 1 X b mn h(n) = h(m)e q q (14.14) m (q) q X =√ n X m (q) d|(q,m2 −4) q mn dµ e . d q (14.15) Switch summations. Then as gcd(n, q) = 1, the sum vanishes if d 6= q. So we get q h(n) = √ q q X e m2 ≡4 (q) mn q . (14.16) Now suppose c is odd and relatively prime to n. We have a bijection between {x : x2 ≡ 1 (mod c)} and {ab = c, (a, b) = 1} by x 7→ ((x − 1, c), (x + 1, c)). Then x a b aa − bb = − = . c b a c (14.17) We then get 2 √ S(1, n ; c) = c c X ab=c (a,b)=1 so that 43 e 2n a b − b a (14.18) S(n, n; c) = √ n X a b e 2n . S(1, n ; c) = c c − c c b a n 2 (14.19) ab=c (a,b)=1 Theorem 14.3. For k ∈ 1 2 + Z with k ≥ 52 , and f ∈ Sk (Γ0 (N )), if n is squarefree, then k 3 |af (n)| n 2 − 7 + . (14.20) For every q, Sk (Γ0 (N )) ,→ Sk (Γ0 (N q)). The Petersson trace formula implies X |af 4πn 1 + 2π c−1 Kk (n, n; c)Jk−1 . [Γ0 (N ) : Γ0 (N q)] c c≡0 (qN ) (n)|2 14.1 (14.21) Trivial Bounds 4πn c If q is prime, then [Γ0 (N ) : Γ0 (N q)] = q + 1, Jk−1 1 min X n c n 1 2 , n c 3 o 2 , and |Kk (n, n; c)| ≤ 1 σ0 (c)(n, c) 2 c 2 . By convention, write L = O (n ). If we fix C ≤ N ≤ D, then X c−1 Kk (n, n; c)Jk−1 c≤C 4πn c 1 1 n− 2 (c, n) 2 σ0 (c) (14.22) c≤C c≡0 (qN ) c≡0 (qN ) L·C 1 , (14.23) qn 2 and X −1 c Kk (n, n; c)Jk−1 c≥D 4πn c c≡0 (qN ) X c≥D 3 1 n 2 (c, n) 2 σ0 (c) c2 (14.24) c≡0 (qN ) 3 n2 . L· qD 14.2 (14.25) A Toy Problem We will try to bound KN (x) = X S(n, n; c) √ c c≤x N |c for x ∼ n. The trivial bound is xL. 44 (14.26) Let X n a b e 2n − Fy (A, B; N ) = ab b a (14.27) where the sum ranges over a, b such that A ≤ a ≤ 2A, B ≤ b ≤ 2B, y ≤ ab ≤ 2y, and N |ab. We have the reciprocity formula 1 a b + ≡ b a ab (mod 1). (14.28) So X n 4na 2n Fy (A, B; N ) = e − . e ab b ab Summation by parts: if f (x) ∈ C 1 (R) and {an }n∈N are given, let Sx = 2A X an f (n) = f (2A)S2A − n=A 2A X P A≤n≤x an . Then Sn (f (n + 1) − f (n)) n=A Z 2A = f (2A)S2A − (14.29) (14.30) Sx f 0 (x) dx. (14.31) A In particular, 2A X an f (n) ≤ A max |f 0 (x)| max |Sx | + |f (2A)||S2A |. [A,2A] [A,2A] (14.32) n=A 0 Take f (x) = e − 2n bx so that f (x) = Hence 2n e bx2 0 − 2n bx , giving A max[A,2A] |f (x)| X X N n 4na e Fy (A, B; N ) 1+ AB a b N |ab (n,b)=1 A≤a≤m B≤b≤2B n A BA2 = n BA . (14.33) b with mb ≤ 2A. Notice that X n 4na =0 e a b (14.34) a (nb) for n squarefree (or just n not a square). Completing the sum: let F : Z → R be m-periodic. Let Fb(r) = X s (m) rs F (s)e − . m 45 (14.35) Then x X F (u) = u=0 x ur X 1 X b F (r)e m m u=0 (14.36) r (m) x ur xFb(0) 1 X X b F (r)e + m m m u=0 r6≡0 (m) X b x|F (0)| 1 1 b F (r) + r m m 1−e m r6≡0 (m) = X xFb(0) 1 + max |Fb(r)| m m r6≡0 (m) r6≡0 (m) (14.37) (14.38) m r (14.39) x|Fb(0)| + max |Fb(r)| log m. m r6≡0 (m) 4nx b We can apply this to m = nb and F (x) = e Fb(r) = (14.40) . Then Fb(0) = 0. If r 6= 0, then X n 4nnx − rx e . x nb (14.41) x (nb) After using partial fractions, we can factor this into a Salié sum and a Kloosterman sum. If we √ assume the Weil bound, we get |Fb(r)| nbL. So X |Fy (A, B; N )| 1+ (n,b)=1 1 n · L · (nb) 2 AB (14.42) B≤b≤2B n 1 3 n2 B 2 . 1+ AB (14.43) 1 If AB = n, the trivial bound is AB, or n, but the above is better if B n 3 . If A, B are both large, then we can “average over the level”. Let X FP,y (A, B; N ) = |Fy (A, B; pN )|. (14.44) P ≤p≤2P p prime,p-n We will choose P to be a small power of n. We have FP,y (A, B; N ) = X P ≤p≤2P X n a b λp e 2n − ab b a (14.45) where λp = sgn(Fy (A, B; pN )) and the inner sum runs over a, b such that A ≤ a ≤ 2A, B ≤ b ≤ 2B, y ≤ ab ≤ 2y, (a, b) = 1, and pN |ab. Now we can write 46 FP,y (A, B; N ) = Fy0 (A/P, B; N ) +Fy0 (A, B/P ; N ). {z } | (14.46) X X n ap b e 2n − |Fy0 (A/P, B; N )| ≤ λp p b ap p1 ≤p≤p2 (14.47) terms with p|a We have where the outer sum runsnover PA ≤ oa ≤ 2A P , B ≤ b ≤ 2B, (a, b) = 1, and N |ab. Here p1 = y A 2y 2A max P, ab , a and p2 = min 2P, ab , a . By Cauchy-Schwarz, X AB |Fy0 (A/P, B; N )|2 P X λp1 λp2 P1 ≤p1 ,p2 ≤P2 a,b ap1 p2 n b − e 2n(p2 − p1 ) p1 p2 b ap1 p2 (14.48) so then X X ap p AB b 1 2 . |Fy (A/P ; B; N )|2 − e(2n(p2 − p1 ) P p ,p b ap1 p2 1 2 a,b (14.49) Summing over a first, the summand is ap1 p2 e 4n(p2 − p1 ) b 2 −p1 ) so |xf 0 (x)| Let f (x) = e 2n(p xbp1 p2 F (x) = e 4n(p2 −pb1 )xp1 p2 . Then Fb(0) = n AB 2n(p2 − p1 ) e . abp1 p2 (14.50) for x ∈ [ PA , 2A P ]. Completing the sum, let m = b and X e x (b) (p2 − p1 )x b (14.51) (x,b)=1 = X dµ n d|(b,p2 −p1 ) d ≤ (p2 − p1 , b) (14.52) (14.53) for p1 6= p2 . For r 6= 0, Fb(r) = X 4n(p2 − p1 )xp1 p2 − rx e b (14.54) x (b) 1 1 b 2 (p2 − p1 , b) 2 L. 47 (14.55) Therefore for p1 6= p2 , X 1 1 ap1 p2 b n A(p2 − p1 , b) L 1+ e 2n(p2 − p1 ) − + b 2 (p2 − p1 , b) 2 . b ap1 p2 AB Pb 2A A ≤a≤ P P (14.56) Now we have X p1 ,p2 ,b X 1 1 3 n n AB A(p1 − p2 , b) A L 1+ L 1+ (p2 − p1 , b) 2 b 2 + B2 + +P · AB Pb AB P P p1 6=p2 (14.57) n 3 2 L 1+ (B 2 P + AP ) + AB. AB (14.58) Hence |F (A/P, B)| 1 5 1 1 n 12 + 1+ · L(A 2 B 4 P 2 + AB 2 ). AB (14.59) 5 1 1 1 1 n 21 + 1+ · L(A 4 B 2 P 4 + A 2 BP − 2 ). AB (14.60) AB 1 P2 Doing the b-sum first, |F (A/P, B)| AB 1 P2 14.3 A Return to the Original Problem Again take k ∈ 1 2 + Z, 4|N , and n squarefree. For 4|c, c md + nd e . d c (14.61) Kk (m, n; qr) = Kk (mq, nq; r)S(mr, nr; q). (14.62) X Kk (m, n; c) = d 2k d (c) If (q, 2r) = 1, then Now for N |Q, define KQ (x) = X 1 c− 2 Kk (n, n; c) (14.63) c≤x Q|c X Kk (nq, nq; r) S(nr, nr; q) √ = . √ q r 48 (14.64) where the sum runs over q, r with qr ≤ x, (q, r) = 1, (q, nN ) = 1, N |r, r|(N n)∞ , and Q|qr. Let w(c) = #{P < p < 2P : p|c, p - n}. Then w(c) < log c. Then we have X KN P (x) = P <p<2P X X X Kk ( ) S( ) √ √ q r p N |r (14.65) q≤x r|(N n)∞ p|q = X X w(q) √ Kk ( )S( ). qr x r (14.66) q≤ r We have Kk (nq, nq; r) = X fr (2s)e s (r) 2nqs r X r fr (2s) = d 2k . d (14.67) (14.68) d (r) d+d=2s Therefore X w(q) X 1 X 2nqs √ fr (2s) . KN P (x) = √ S(nr, nr; q)e q r r r P <p<2P q≤ xr s (r) | {z } X (14.69) call this Tr,s (x) We then have Tr,s (x) = fr (2s) X nr ra rb sab w(ab) e 2n − + ab b a r (14.70) with the sum running over a, b with ab ≤ xr , (a, b) = 1, and N |ab. Let fy (A, B; N ) be the above summand over y ≤ a, b ≤ 2y, A ≤ a ≤ 2A, B ≤ b ≤ 2B, and (a, b) = 1. 14.3.1 Estimate I Modulo 1, ra rb ra rbrb − ≡ − b a b arb ra 1 − aa ≡ − b abr ra a 1 ≡ + − . b br abr So 49 (14.71) (14.72) (14.73) a b 2nra 2na 2n e 2nr =e e − − + b a b br abr 2n(1 + rr)a 2n =e e − . br abr (14.74) (14.75) Partial summation with respect to a then gives To do Can the (?) be referenced? (12) n X X 2n(1 + rr + sbb)q nr Fy (A, B; N ) 1 + e w(ab) abr br a (n,b)=1 | {z } B≤b≤2B (?) (14.76) with the sum in (?) running over A ≤ a ≤ 2A, y ≤ ab ≤ 2y, (a, b) = 1, and N |ab. Then (a, b) = 1 implies w(ab) = w(a) + w(b). (?) becomes w(b) X e( ) a nr a X + P ≤p≤2P X ··· (14.77) p|a A≤a≤2A is nontrivial in arithmetic progressions Now complete the sum over A. It’s important that nr a with modulus P N with respect to a. (This is where n being squarefree is used.) 1 We get (?) L(nbr) 2 P , so 3 1 n Fy (A, B; N ) 1 + LB 2 (nr) 2 P. y (14.78) Summing over b first would replace B with A. 14.3.2 Estimate II Fy (A, B; N ) = Fy0 (A/P, B; N ) + Fy0 (A, B/P ; N ) for Fy0 (A/P, B; N ) X = P ≤p≤2P X rap rb sabp nr λp e 2n − + . b ap r pab (14.79) To do What is the set the sum runs over? (13) By Cauchy-Schwarz, |Fy0 (A/P, B; N )|2 ap1 p2 r br sabp1 p2 AB X X − . e 2n(p2 − p1 ) + P p ,p b ap1 p2 r 1 2 a,b Partial summation and completing the sum (the a sum first), 50 (14.80) Fy0 (A/P, B; N ) 1 1 1 n 2 1 1 5 1 +L 1+ (r 4 A 2 B 4 P 2 + r− 4 AB 2 ). y (14.81) 1 1 1 1 n 2 1 5 1 1 (r 4 A 4 B 2 P 4 + r− 4 A 2 BP − 2 ). +L 1+ y (14.82) AB 1 P2 If the b sum is done first, Fy0 (A/P, B; N ) AB 1 P2 Optimizing, 5 1 1 5 1 1 7 3 min{A 2 B 4 P 2 , A 4 B 2 P 4 } ≤ (AB) 8 P 8 . (14.83) Therefore Fy (A, B; N ) yr 14.3.3 −1 n p +L 1+ y 1 2 1 2 5 7 1 3 1 5 [y 8 r− 8 p 8 + (A− 2 + B − 2 )yr− 4 ]. (14.84) The Final Result Recall that Estimate I gives n Fy (A, B; N ) L min{A, B} (nr) 1+ . y (14.85) 1 n −4 −1 −7 1 −1 n 2r 8y2p 2. min{A, B} 1 + y (14.86) 3 2 1 2 We’ll use Estimate I if We then get |Tr,s (2y) − Tr,s (y)| yr −1 − 21 p 5 7 5 3 1 3 13 1 n 8 + 1+ · [y 8 r− 8 p 8 + n 8 y 4 r− 16 p 4 ]. y (14.87) We get a similar bound for |Tr,s (y)| as a result. So X 1 X √ |Tr,s (x)| r X KN P (x) L p-n N |r P <p<2P r|(nN )∞ L s (r) 5 1 3 1 1 1 x 8 4 8 8 8 4 + (x + p) [x p + n x p ] . √ p Now if f ∈ Sk (Γ0 (N )), we have 51 (14.88) (14.89) |af (n)|2 1 + 2πi−2k nk−1 p X −1 c Kk (n, n; c)Jk−1 c≡0 (pN ) 4πn c . (14.90) Fix Me Should all of 2πi be raised to power? (14) Summing over P < p < 2P , p prime, and p - n, X |af (n)|2 4πn − 1 P − 12 c w(c)Kk (n, n; c)Jk−1 c 2. + log P c nk−1 log P (14.91) c≡0 (N ) 1 5 Partial summing with respect to c, using (x− 2 Jk−1 (x))0 min{x− 2 , 1}, we get |af (n)|2 L nk−1 log P k 1 1 1 3 3 1 1 P + n2 P−2 + n8 P 8 + n8 P 4 log P 2 For P = n 7 , we get |af (n)| n 2 − 7 + . (Note that half-integral weight.) 2 7 1 4, > . (14.92) so we improved our bound for Recall that there exists g ∈ S2k−1 (Γ0 (N )) such that for t squarefree, af (tn2 ) = af (t) X 3 µ(d)dk− 2 ag n d d|n 1 k 1 . (14.93) δ So ag (m) mk− 2 −δ+o(1) ⇐⇒ ∀ t, af (tn2 ) (n2 ) 2 − 4 − 2 . The trivial bound δ = 0 gives k k 1 3 af (tn2 ) (n2 ) 2 − 4 +o(1) , while δ = 14 gives af (tn2 ) (n2 ) 2 − 8 +o(1) . So this technology would have imporoved our bounds even more. 15 Heegner Points and Geodesics For quadratic forms with negative discriminant, let Q(x, y, z) = x2 − 4yz, then #{Q(a, b, c) = n} equals either 0 or ∞, bu this happens because SOQ (Z) is infinite. Let Vn = {(a, b, c) : b2 − 4ac = n}. For n < 0, we have SOQ (Z)\Vn (Z) ,→ SOQ (Z)\V−1 (R) a b c (a, b, c) 7→ √ , √ , √ . −n −n −n (15.1) (15.2) √ We also have V−1 (R) ∼ = H by (x, y, z) 7→ SL2 (Z) action on H. So we get −x± ϕ x2 −4yz . 2y The SOQ (Z) action translates to the SOQ (Z)\Vn (Z) − → SL2 (Z)\H. (15.3) Let Hn = im ϕ, called the Heegner points of discriminant n. These will be shown to be equidistributed with respect to dxy2dy for squarefree n → ∞. 52 √ −b− b2 −4ac 2a For n > 0, let γa,b,c be the geodesic in H connecting b centered at − 2a ). Then we have / SL2 (Z) and √ −b+ b2 −4ac 2a (a semicircle ψ [ γa,b,c − → SL2 (Z)\H. (15.4) b2 −4ac=n Let Sn = im ψ. Fix Me More stuff on this topic appears later. Reorganize? (15) 16 Maass Forms Let ∆ be the Laplacian with respect to the hyperbolic metric; that is, SL2 (R)-invariant. −y 2 ∂2 ∂x2 + ∂2 ∂y 2 . ∆ is We say that ϕ(z) is a maass form of level N if: • ∆ϕ = λϕ · ϕ for some λϕ ∈ R. • ϕ is invariant under Γ0 (N ). • ϕ(x + iy) = O(y A ) as y → ∞, for some A. As an example, for s ∈ R, ∆y s = s(1 − s)y s . If <(s) > 1, take Γ = SL2 (Z) for simplicity and take E(z.s) = 1 X im(γz)s 2 (16.1) ys 1 X . 2 |cz + d|2s (16.2) Γ∞ \Γ = (c,d)=1 Then ∆E(z, s) = s(1 − s)E(z, s) and E(γz, s) = E(z, s). We have that 1 ζ(2s)E(z, s) = y s 2 We can write E(z, s) = y −s P r∈Z ar (y, s)e(rx) X |cz + d|−2s . (16.3) (c,d)6=(0,0) since E is 1-periodic. We have ζ(2s)ar (y, s) = δ0,r ζ(2s) + XXZ c>0 d∈Z = δ0,r ζ(2s) + 1 |cz + d|−2s e(−rx) dx X Z c (c2 x2 + c2 y 2 )−s e(−rx) dx. c|r Now let 53 R (16.4) 0 (16.5) Ks (y) = 1 2 Z ∞ −1 )/2 e−y(t+t 0 ts dt . t (16.6) Ks (y) converges for y > 0, decays exponentially as y → ∞, and Ks (y) = K−s (y). We end up with ( π −s Γ(s)ζ(2s)y s + π s−1 Γ(1 − s)ζ(2 − 2s)y 1−s Γ(s)π −s ζ(2s)ar (y, s) = 1 √ 2|r|s− 2 σ1−2s (|r|) yKs− 1 (2π|r|y) r=0 r 6= 0 (16.7) 2 We learn that ζ(2s)E(z, s) continues meromorphically in s to C, with a unique pole at s = 1, π and Ress=1 E(z, s) = ζ(2) = π3 . Theorem 16.1. vol(Γ\H) = π3 . Idea of proof. Use Z Z dx dy y2 (16.8) =(γz)s δ(=(γz) ≤ T ). (16.9) π dx dy 1 2 = Ress=1 y 3 Γ\H E(z, s) Γ\H and unfold. Proof. Take T > 1 and let ET (z, s) = 1 2 X γ∈Γ∞ \Γ This is still invariant under Γ, and if z ∈ F, then ET (z, s) = E(z, s) − y s δ(y > T ). Therefore Ress=1 ET (z, s) = π3 . Now Z π dx dy Ress=1 ET (z, s) 2 3 Γ\H y Z π dx dy y s δ(y ≤ T ) 2 = Ress=1 3 y Γ∞ \H vol(Γ\H) = T s−1 π Ress=1 3 s−1 π = . 3 = (16.10) (16.11) (16.12) (16.13) We can compute vol(G(Z)\G(R)) for many G similarly (such as G split of rank one). P Suppose ϕ(z) is a maass form. Then ϕ(z) = r∈Z fr (y)e(rx), and ∆ϕ(z) = X − y 2 fr00 (y) + 4πr2 y 2 fr (y) e(rx) r∈Z 54 (16.14) implying fr (y)(4πr2 y 2 − λϕ ) − y 2 fr00 (y) = 0. This and fr (y) = O(y A ) for some A implies √ fr (y) = aϕ (r) yKv (2π|r|y) for r 6= 0, where λϕ = 1 4 (16.15) − v2. We say that ϕ is even if ϕ(x + iy) =Rϕ(−x + iy) and odd if ϕ(x + iy) = −ϕ(−x + iy). We say 1 that ϕ is a maass cusp form if for all y, 0 ϕ(x + iy) dx = 0. Let L20 (Γ\H) be the space of cuspidal functions. We have Z hf, gi = f (z)g(z) Γ\H dx dy . y2 (16.16) Lemma 16.2. For f, ∈ L2 (Γ\H), h∆f, yi = h∆f, gi. R Proof. We need Γ\H ω = 0 for ω = (fxx + fyy )g − (gxx + gyy )f dx dy. But ω = d (fx g − gx f )dy + (fy g − f gy )dx . By Stokes’ theorem, we’re done. We also have that for f > 0, Z hf, ∆f i = −f (fxx + fyy ) dx dy (16.17) (fx2 + fy2 ) dx dy (16.18) Γ\H Z = Γ\H >0 (16.19) because the integrands differ by d(f fx dy + f fy dx). As a result (and some spectral theory), we get that L20 (Γ\H) = M (16.20) Cϕi i∈N ∆ϕi = λi ϕi . As y → ∞, E(z, s) ∼ Ay s + By 1−s so ϕ(r) ∈ Cc∞ (R), define 2 dx dy Γ\H |E(z, s)| y 2 R (16.21) doesn’t quite converge. But for 1 Eϕ (z) = E z, + ir ϕ(r) dr. 2 R (16.22) Eis = hEϕi : ϕi ∈ Cc∞ (R)i. (16.23) Z Then Eϕ ∈ L2 (Γ\H). Define Then we have: 55 Theorem 16.3. L2 (Γ\H) = L20 (Γ\H) ⊕ Eis ⊕ C · 1. Theorem 16.4. For y ∈ R+ , {x + iy : 0 ≤ x ≤ 1} = Iy → Γ\H; then Iy equidistribute to Equivalently, Z 1 Z f (x + iy) dx → f Γ\H 0 dx dy y2 dx dy . y2 (16.24) as y → 0 for each f ∈ Cc∞ (Γ\H). Proof. Use the decomposition. For f = 1, this is clear. For f ∈ L20 , the left hand side is identically zero, and Z f Γ\H dx dy = hf, 1i = 0. y2 (16.25) Finally, Z 0 1 1 1 Eϕ (x + iy) dx = ϕ(r) E x + iy, + ir dx 2 R 0 Z 1 1 = ϕ(r) f1 (r)y 2 +ir + f2 (r)y 2 −ir dr Z Z (16.26) (16.27) R 1 = O(y 2 ) (16.28) as y → 0. Fix Me We have more quadratic stuff now. Reorganize? (16) √ Suppose √D is squarefree, negative, and for simplicity D ≡ 1 (mod 4). Let K = Q( D) and OK = Z[ 1+2 D ]. Let IK be the group of fractional ideals, so K × \IK = Cl(K). Consider triples (I, α, β) for I ⊆ K a fractional ideal, α and β generating I, and =(αβ) > 0. We can also mod this out by K × , acting by scaling all of I, α, and β. Call the quotient RD . Now we have a map ψD : RD → H by (I, α, β) 7→ αβ . Then ψD is SL2 (Z)-equivariant, with SL2 (Z) acting on RD by change of basis. Lemma 16.5. z ∈ im ψD iff z = √ −b+ D 2a for a, b, c ∈ Z with b2 − 4ac = D. Proof. For the “if” implication, take β = 2a and α = −b + √ D. Check that α and β span an ideal. √ 1+ D ∈ span(α, β), 2 b2 −D that 4a ∈ Z. For the “only if” implication, after scaling, we can assume that β = 1. Then so α = √ −b+ D 2a for some a ∈ N and b ∈ Q. Then √ 1+ D 2 α ∈ span(α, 1) shows Corollary 16.6. |HD | = |Cl(K)|. Fix I1 , . . . , Ir representatives in Cl(K). Then for <(s) > 1, X z∈HD r X s X α E(z, s) = = β i=1 56 (16.29) αβ with the inner sum ranging over α, β ∈ Ir , α (mod β), and =(αβ) > 0. Now =( αβ ) = =( |β| 2 ). We get that r X X i=1 √ !s r s X X N (Ii ) D 2 =D N (βIi−1 )−s |β|2 i=1 β∈Ii s X = D2 N (I)−s (16.30) (16.31) IOK s 2 = D ζK (s). (16.32) ∼ We have π : P GL+ → H by g 7→ g · i. Let 2 (R)/P SO(2) − T ⊆ P GL+ 2 (R) x 0 × = : x, y ∈ R , xy > 0 . 0 y (16.33) Lemma 16.7. For g ∈ SL2 (R), π(gT ) is a geodesic, and all geodesics arise this way. Proof. For g = e, π(T ) = {ix : x > 0} which is a geodesic. In general, π(gT ) = gπ(T ) and z 7→ gz is an isometry. If g = ac db and tx = ( x0 10 ), then π(g · tx ) = We get a geodesic joining a c axi + b ax b . ·i= cx d cxi + d (16.34) and db . For D > 0 and a, b, c ∈ Z with b2 − 4ac = D, let γ[a,b,c] be the geodesic joining √ √ √ −b− D =b+ D −b− D T . Now gT = (gT g −1 )g; we have . This is given by π 2a 2a 2a √ −b+ D 2a and √ √ √ 1 x − y −b 2c x+y 1 0 x 0 2a b + √D −b + D −b − D √ = √ − . 0 y 0 1 2a 2a −2a −b + D 4a D 2 2 D 2a b (16.35) √ −2c . Then M 2 Let M[a,b,c] = −b [a,b,c] = D · I. For K = Q( D), we get i[a,b,c] : K ,→ M2 (Q) by 2a b √ u + v D 7→ u + vM[a,b,c] . Observe: √ D+ • i−1 2 [a,b,c] (M2 (Z)) = Z[ this is OK . D ], the unique quadratic ring of discriminant D. If D is squarefree, • For g ∈ SL2 (Z), gM[a,b,c] g −1 = Mg[a,b,c] . • If M ∈ M2 (Z) with M 2 = D, then M = −b −2c 2a b ∼ for some a, b, c ∈ Z with b2 − 4ac = D. −1 Given [a, b, c], consider θ[a,b,c] : K − → Q2 by λ 7→ i[a,b,c] (λ) ( 10 ). Let I[a,b,c] = θ[a,b,c] (Z2 ). Then I is a fractional ideal of K. For g ∈ SL2 (Z), 57 1 −1 1 ig[a,b,c] (λ) = gi[a,b,c] (λ)g . 0 0 (16.36) We know that there exists λ ∈ K × such that λg −1 ( 10 ) = ( 10 ). Then I[a,b,c] λ = Ig[a,b,c] . So the ideal class of I[a,b,c] in Cl(K) is independent of G-orbit. Conversely, if I OK is an ideal, and α, β is a basis of I, we get i : K ,→ M2 (Q) by a α β i(λ) = λ(aα + bβ). b (16.37) As I is an ideal, i−1 (M2 (Z)) = OK and θi−1 (Z2 ) = α−1 I. This provides a bijection SL2 (Z)\{(a, b, c) : b2 − 4ac = D} ↔ Cl(K). (16.38) Let [ SD = T[a,b,c] Z[a,b,c] (16.39) SL2 (Z)\[a,b,c] Z[a,b,c] = √ √ −b + D −b − D . 2a 2a (16.40) Q Q Let A be the adele ring 0p Qp × R, where 0 consists of sequences (ap ) such that ap ∈ Zp for all but finitely many p. Then Q ,→ A and A/Q is compact. We have ϕ + + → P GL+ P GL+ 2 (Q)\P GL2 (A) 2 (Z)\P GL2 (R)/P SO(2) − .Y P GL2 (Zp ) × P SO(2) (16.41) p by ϕ(g)∞ = g, ϕ(g)p = 1. Theorem 16.8 (Strong approximation). ϕ is an isomorphism. + + Proof. For (Q) Q injectivity, suppose g ∈ P GL2 (R) such that ϕ(g) = 1. Then there exists hQ ∈ P GL2−1 and k ∈ p P GL2 (Zp ) × P SO(2) such that ϕ(g) = hQ k. The for every p, hQ · kp = 1 so hQ = kp ∈ P GL2 (Zp ) implying hQ ∈ P GL+ 2 (Z). Now considering the infinite place, g = hQ k∞ is trivial in + P GL+ (Z)\P GL (R)/P SO(2). 2 2 For surjectivity, it is enough to show that for all primes p, 1 P GL2 (Zp ) · P GL2 Z = P GL2 (Qp ). p (16.42) Suppose g = ac db ∈ P GL2 (Qp ). We have ϕm : Qp → Qp /pm Zp ≈ Z[ p1 ]/pm Zp . We can pick A B ∈ P GL (Z[ 1 ]) such that h + pm Z = ϕ (g) for m 0. So g h−1 ∈ P GL (Z ). hp = C 2 p p m p p 2 p D p 58 Now define TD = {(x, y) : x2 − Dy 2 is a unit} with (x1 , y1 ) · (x2 , y2 ) = (x1 y1 + Dx2 y2 , x1 y2 + x2 y1 ). (16.43) For example, TD (Q) ≈ K × and TD (A) ≈ A× K = 0 Y Kp× × (K ⊗Q R)× (16.44) p √ where K = Q( D). We have .Y ∼ → Cl(K). TD (Z)p − TD (Q)\TD (Af ) (16.45) p We have i[a,b,c] : TD (A) → GL2 (A) → P GL+ 2 (A). Examine TD (A) · Z[a,b,c] : P GL+ 2 (Q) (t∞ , t2 , . . . , tp , . . .) √ √ −b − D −b + D , 12 , . . . , 1p , . . . P GL2 (Zp ) (16.46) 2a 2a ∞ Claim. TD (A) · Z[a,b,c] = SD . √ D . For z ∈ H, We can also do the case D < 0 adelically. For b2 − 4ac = D, let Z[a,b,c] = −b+ 2a + −1 Kz = stabz ⊆ P GL2 (R). So stabi = P SO(2) and stabg(i) = gP SO(2)g . Therefore √ KZ[a,b,c] x D −b = −y 0 2a y b = xI + √ D −2a | {z 1 2a √b √ 0 D 2a D 2c . −b } y x (16.47) (16.48) M[a,b,c] √ √ We get i[a,b,c] : K ,→ M2 (Q) for K = Q( D) by u + v D 7→ uI + vM[a,b,c] . We have TD (A) · Z[a,b,c] = HD . For points on a sphere: Let H be the quaternions (over R). We have a quadratic form hz, wi = + wz). Let V ⊆ H be the subset of purely imaginary quaternions; that is, those α such that α = −α. Then S 2 ⊆ V as (a, b, c) = ai + bj + ck, as the elements of norm 1. 1 2 (zw ∼ Let P1 (H× ) = H× /Z(H× ) − → SO(3) by ϕ ϕ(α)β = αβα−1 = αβα . |α|2 (16.49) For β ∈ V \ {0}. Z(β) = R ⊕ Rβ ≈ C by u + vβ 7→ u + v|β|i. Now consider {(a, b, c) : a2 + b2 + c2 = D, a, b, c, D ∈ Z}. 59 (16.50) √ √ We have stab ai+bj+ck −D) = P(R ⊕ R(ai + bj + ck)). We obtain i : K ,→ H for K = Q( [a,b,c] D √ √ D+ −D −1 by u + v −D 7→ u + v(ai + bj + ck), and i[a,b,c] (H(Z)) = Z[ ]. 2 Our analogies are: P GL2 H Heegner points/geodesics hyperbolic Laplacian maass forms P(H× ) S2 elements of fixed norm Laplacian harmonic polynomials √ ) are coefficients of theta functions Recall that for harmonic polynomials P , a2 +b2 +c2 =D P ( (a,b,c) D (modular forms of half-integral weight). So we might expect that for ϕ a maass form, we would get a “weight 12 ” maass form ϕ e with P Z X ϕ(z) ∼ ϕ(z) dz. (16.51) SD z∈HD Fix Me One of these should have a tilde. (17) Fix Me Now we’re doing L-functions of two modular forms. (18) Suppose f (z) ∈ Sk (Γ) and g(z) ∈ Sk (Γ) are Hecke cusp forms. Then Z f (z)g(z)y k E(z, s) As = Γ\H dx dy y2 (16.52) is meromorphic in C, and is holomorphic for <(s) > 1. We have that Z f (z)g(z)y k Ress=1 As = Γ\H = ( 0 3 dx dy π y2 f= 6 c·g f = g. 3 π hf, f i (16.53) (16.54) But also Z f (z)g(z)y k · As = Γ\H Z X =(γz)s γ∈Γ∞ \Γ dx dy y2 f (z)g(z)y k+s−2 dx dy = Γ∞ \H Z ∞Z 1 = 0 Z 1 2 X 0 n>0 ∞ = 0 X af (n)e(nz) X ag (m)e(−mz)y k+s−2 dx dy (16.55) (16.56) (16.57) m>0 af (n)ag (n)e−4πny y k+s−2 dy (16.58) n>0 X af (n)ag (n) = Γ(k + s − 1) (4πn)k+s−1 (16.59) = L(f × g, s)(4π)−(k+s−1) Γ(k + s − 1) (16.60) n>0 60 where X af (n)ag (n) L(f × g, s) = n>0 nk+s−1 . (16.61) In the case f = g, L(f × f, s) is meromorphic in C is a unique pole at s = 1. Also L(f × f, s) has polynomial growth in vertical strips. We also have a functional equation relating L(f × g, s) to L(f × g, 1 − s). Theorem 16.9. There exists c such that X af (n)2 ∼ cX. nk−1 (16.62) n≤X In fact c = Ress=1 L(f × f, s). P a (n)N Remark. If n>0 f ns were reasonably behaved for every N , we would know Ramanujan. Recall that L(f, s) = 1 Y p (1 − αf,p p−s )(1 −1 −s − αf,p p ) . (16.63) We can do something similar for L(f × g, s): ! X af (pm )ag (pm ) L(f × g, s) = pm(k+s−1) p m>0 ! P r1 −r2 P s1 −s2 Y X r1 +r2 =m αf,p α g,p s1 +s2 =m = m(s−1) p p m>0 Y (16.64) (16.65) !−1 = Y −s (1 − αf,p αg,p p )(1 − −1 −s αf,p αg,p p )(1 − −1 αf,p αg,p p−s )(1 − −1 −1 −s αf,p αg,p p ) . p (16.66) The Langlands philosophy: think of α f,p 0 0 α−1 f,p as a (semisimple) conjugacy class in GL2 , or SL2 . We have a block diagonal action GL2 × GL2 → GL4 by (v, w) 7→ v ⊗ w, preserving conjugacy classes, and αf,p −1 αf,p × αg,p * −1 αg,p 7→ !+ αf,p αg,p .. . . (16.67) For the case f = g, !−1 L(f × f, s) = ζ(s) Y −2 −s 2 (1 − αf,p p−s )(1 − αf,p p )(1 − p−s ) p 61 . (16.68) We have a symmetric square map Sym2 : GL2 → GL3 with α α−1 *α2 1 7 → + α−2 . (16.69) More generally, we have SymN : GL2 → GLN +1 , and an L-function !−1 N L(Sym f, s) = Y N −2 −s −N −s N −s (1 − αf,p p )(1 − αf,p p ) · · · (1 − αf,p p ) . (16.70) p It’s conjectured that this extends holomorphically to C. Fix Me Back to maass forms, maybe? (19) θ sin θ 1x Let G= SL2 (R), K =SO2 (R) = kθ = −cos sin θ cos θ : θ ∈ R/2πZ , N = {nx = ( 0 1 ) : x ∈ R}, 1 2 and A = ay = y −0 1 : y ∈ R+ a maximal torus. Then we have the Iwasawa decomposition: 0 y 2 1 1 y 2 xy − 2 = x + iy. G = N AK, G/K ≈ N A ≈ H by g 7→ g · i, and nx ay = −1 0 y 2 Recall the Laplacian on G, ∆G = −y 2 ∂2 ∂2 + ∂x2 ∂y 2 +y ∂2 . ∂x∂θ (16.71) Any f ∈ C ∞ (G) can be written f (nx ay kθ) = X fk (x + iy)e(kθ). (16.72) k∈Z Applying ∆G , we have X ∂2 ∂ 2 ∂2 (∆G f )(nx ay kθ ) = + + iky −y fk (x + iy)e(kθ). 2 2 ∂x ∂y ∂x k∈Z | {z } (16.73) ∆k For Γ ⊆ G discrete, we need to study L2 (Γ\G) under the right action of G, fh (g) = f (gh). Under ∆G , if f ∈ C ∞ (Γ\G) with ∆G f = λf , then ∆k fk = λfk . Now for γ ∈ Γ given by γ = ac db , If for z = x + iy, we let a b c d az+b cz+d 1 y2 0 1 xy − 2 1 y− 2 ! =y − 12 ay ax + b . cy cx + d (16.74) = x0 + iy 0 , then get 1 √ 0 y cx+d |cz+d| cy |cz+d| 0 y x0 0 1 | cy − |cz+d| . cx+d |cz+d| {z kz 62 ! } (16.75) Now for g = hx ay kθ , we have f (g) = f (γg) (16.76) = f (γax hy kθ ) (16.77) = f (hx0 ay0 kz kθ ) X az + b cz + d k e(kθ). = fk cz + d |cz + d| (16.78) (16.79) k∈Z For f ∈ C(H) and γ = a b c d , define f |k γ to now be f |k γ = f az + b cz + d cz + d |cz + d| k . (16.80) We say that f is a maass form of weight k and level N if for every γ ∈ Γ0 (N ), f |k γ = f , ∆k f = λf for some λ, and f (x + iy) = O(y A ) for some A. k Theorem 16.10. For f (z) ∈ Mk (Γ0 (N )), let F (z) = y 2 f (z). Then F is a maass form of weight k and level N . k Proof. We have F (x + iy) = O(y 2 ), ∆k f = k2 (1 − k2 )f , and F |k γ = F for γ ∈ Γ0 (N ). 1 Let θ1 (z) = y 4 θ(z). Then for γ ∈ Γ0 (4), define θ1 (γz) θ1 (z) 1 cz + d 2 c = . d |cz + d| d J(γ, z) = (16.81) (16.82) For k ∈ 21 Z, f is a maass form of weight k and level N if f (x + iy) = O(y A ) for some A, ∂ f (γz) = J(γ, z)2k f (z), and f is an eigenfunction of ∆ 1 = ∆ + iy 2 ∂x . 2 For any s > 0, ∆k ys = s(1 − s)ys, so for s > 1, letting =s (z) = y s , we can define Ek (s, z) = = 1 2 X (=s |k γ)(z) (16.83) γ∈Γ∞ \Γ0 (4) 2k X y s (cz + d)k 2k c . |cz + d|k+2s d d (16.84) (c,d)=1 4|c For f ∈ Sk (Γ) and g ∈ S` (Γ), consider Z f (z)g(z)Ek−` (s, z)y Γ\H 63 k+` 2 dx dy . y2 (16.85) Take k = Z 1 2 and N = 4. Then for n 6= 0, we get iy+1 E 1 (s)e(−nx) dx = y s X Z 2 iy (c,d)=1 iy+1 d iy c d k k (cz + d) 2 −s (cz + d)− 2 −s e(−nx) dx (16.86) 4|c = ys X c−2s c6=0 X nd c Z i+∞ k k d z 2 +s z 2 −s e(−nx) dx e − c d i−∞ d (c) | {z } (16.87) call this W 1 ,s (4π|n|y) 4 2 = y W 1 ,s (4π|n|y) X −2s c 4 K 1 (−n, 0; c). (16.88) 2 c6=0 For c = 4p, K 1 (−n, 0; c) = K 1 (np−1 , 0; 4)S(n · 4−1 , 0; p). We want to determine 2 2 X x ax S(a, 0; p) = e . p p (16.89) x (p) Then S(ab, 0; p) = S(a, 0; p) b p for (b, p) = 1. And, taking α to be any nonsquare modulo p, X x x e S(1, 0; p) = p p x (p) 2 2 1 X 1 X x αx = − e e 2 p 2 p x (p) x (p) 1 √ 1 √ = p p − − p p 2 2 √ = p p. (16.90) (16.91) (16.92) (16.93) We end up with Z iy+1 iy E 1 (z, s)e(−nx) dx = W 1 ,s (4π|n|y) 2 X 4 c6=0 c n c2s . (16.94) | {z } L(χn ,2s) 17 Subconvexity of L-Functions Recall that X s E(z, s) = |D| 2 L(s, χD )ζ(s) z∈HD Also, we have 64 (?). (17.1) L2 (SL2 (Z)\H) = 1 ⊕ L20 (SL2 (Z)\H) ⊕ Eis, (17.2) Eis = hEϕ : ϕ ∈ Cc∞ (R)i. (17.3) 1 E z, + it ϕ(t) dt. Eϕ (z) = 2 R (17.4) where Here Z Corollary 17.1. To conclude equidistribution, we need that as |D| → ∞, 1 X 1 E z, + it → 0. |HD | 2 (17.5) z∈HD Observe that the residue at s = 1 of (?) gives 3 π |HD | 1 = |D| 2 L(1, χD ). Theorem 17.2. L(1, χD ) = |D|o(1) . Hence to prove the corollary, we need to show that there exists δ > 0 such that for every t, 1 L( 12 + it, χD ) = O(|D| 4 −δ ). This is a subconvexity problem. For n ∈ N, χ : (Z/mZ)× → C× is called a Dirichlet character to modulus m. We say that χ is primitive if there does not exist a proper divisor n of m such that χ factors through (Z/nZ)× . In this case, we say that m is the conductor of χ. There are exactly m characters to modulus m. The number of primitive characters is X ϕ(d)µ n d d|m If m is squarefree, the above equals m1− . Q = Y (ϕ(pr ) − ϕ(pr−1 )). (17.6) pr kn p|m (p − 2). If 2km, this number is zero, otherwise it is Observe that if D is odd and squarefree, then χD (n) = P n Define τ (x) = n (m) χ(n)e m , the Gauss sum for χ. √ Lemma 17.3. If χ is primitive, then |τ (χ)| = m. n D is primitive. Proof. Consider χ : Z/mZ → C by extension by zero. Then χ b(r) = nr n (m) χ(n)e m . (mod m c ) such that P If (r, m) = c 6= 1, then since χ is primitive, there exists s ≡ 1 Then n 7→ ns gives χ b(r) = χ(s)b χ(r) so χ b(r) = 0. χ(s) 6= 1. On the other hand, if (r, m) = 1, n 7→ nr gives χ b(r) = χ(r)τ (χ). Now Plancherel implies X n (m) |χ(n)|2 = 1 X |b χ(r)|2 m r (m) 65 (17.7) so ϕ(m) = ϕ(m)|τ (χ)|2 , m giving |τ (χ)| = √ m. We can now write χ(r) = nr 1 X χ(n)e τ (χ) m (17.8) nr τ (χ)χ(−1) X χ(n)e m m (17.9) n (m) = n (m) because τ (χ) = τ (χ)χ(−1). Twisted Poisson summation formula: For f ∈ C 2 (R), with f (x) = O((1+|x|)−2 ), and χ primitive of conductor m, then X χ(n)f (n) = n∈Z X χ(n) X f (km + n) (17.10) k∈Z n (m) X χ(n) X k kn fb = e − m m m k∈Z n (m) τ (χ)χ(−1) X k . = χ(k)fb m m (17.11) (17.12) k∈Z Suppose χ is even (meaning χ(−1) = 1). Then let θχ (t) = X 2 χ(n)e−πn t . (17.13) n∈Z Twisted Poisson summation yields τ (χ) θχ (t) = √ θχ m t For L(χ, s) = is nontrivial. χ(n) n∈N ns P 1 m2 t . (17.14) converges absolutely for <(s) > 1, and conditionally for <(s) > 0 if χ −1 Assume that χ is nontrivial. Then θχ (t) = O(e−t ) as t → ∞, and θχ (t) = O( e√tt ) as t → 0. R∞ s Then 0 θχ (t)t 2 dtt is entire in s, and if <(s) > 1, is equal to XZ n∈Z 0 Applying t 7→ 1 , m2 t ∞ 2 s χ(n)e−πn t t 2 s s dt = 2π − 2 L(χ, s)Γ = Λ(χ, s). t 2 we obtain Λ(χ, s) = τ (χ)m−s Λ(χ, 1 − s). Corollary 17.4. L(χ, s) is entire, with trivial zeros at the nonpositive even integers. 66 (17.15) Now suppose that χ is odd, meaning χ(−1) = −1. Then define X θχ (t) = 2 χ(n)ne−πn t . (17.16) n∈Z We end up with s+1 Λ(χ, s) = π Γ L(χ, s) 2 Λ(χ, s) = (−i)τ (χ)m−s Λ(χ, 1 − s). − s+1 2 L(χ, s) has an Euler product Q p (1 (17.17) (17.18) − χ(p)p−s )−1 in the domain of convergence. Convexity bounds: let s = σ + it. For σ > 1, L(χ, σ + it) = O(1). For σ < 0, the functional 1 1 equation gives L(χ, σ + it) = O((1 + |t|) 2 −σ , m 2 −σ ). The Phragmen-Lindelof principle shows that 1 L χ, + it 2 Define Snχ = Pm i=1 χ(i). 1 1 = O ((1 + |t|) 4 + m 4 + ). (17.19) Then by summation by parts, X ∞ 1 1 1 χ L χ, = Sn √ − √ 2 n n−1 n=1 ∞ X 1 − 52 χ = Sn ) . 3 + O(n 2n 2 n=1 (17.20) (17.21) We have |Snχ | ≤ ϕ(m) ≤ m by the orthogonality relations (so Snχ = Snχ (m) ). Also, we have that √ √ for every r, |b χ(r)| ≤ m, as well as χ b(0) = 0, so |Snχ | m log m. Finally, we have the trivial bound |Snχ | ≤ n. Now √ m log m ∞ X X 1 n L χ, 3 + 2 2n 2 √ n=1 1 m log m √ m log m (17.22) 3 n2 1 m 4 (log m) 2 . (17.23) 1 1 To prove subconvexity, we need a bound of the form |Snχ | m 2 −δ for n ∼ m 2 . 1 1 Return to the Riemann O (|t| 4 + ). P∞ zeta1function. The convexity bound for ζ( 2 + it) as t → ∞ isP Observe that if t 6= 0, n=1 1 +it converges conditionally. For subconvexity, we need nr=1 rit n2 √ 1 −δ 2 t for n ∼ t. Let 67 M X M RN = nit (17.24) n=N M −N X it =N = N it n=0 M −N X 1+ n it N 1 n (17.25) n 2 n 3 +O(( N ) )) eit( N − 2 ( N ) . (17.26) n=0 P −N it(( n )− 1 ( n )2 ) N 2 N = o(1). Then we need M to be small. n=0 e PN Model problem: for α, β ∈ (0, 1], we want to bound n=0 e(αn + βn2 ). We have We can ignore the third tem if tn3 N3 2 N N X X e(αn + βn2 ) = e(α(n − m))e(β(n − m)(n + m)) n=0 (17.27) m,n=0 = 2N −|c| N X X e(αc) c=−N e(βcd) (17.28) d=|c| d≡c (2) N X min c=−N 1 ,N |1 − e(2βc)| N log N. (17.29) (17.30) As an exercise, one can show that for k ∈ N, N X e(nk α) N 1− 1 2k + . (17.31) n=1 Returning to Snχ for n ∼ √ m, consider X χ (m) This implies |Snχ | √ χ 2 χ (m) |Sn | , P |Snχ |2 = n X X We get χ(i)χ(j) (17.32) χ (m) i,j=1 = m · #{i ≤ n : (i, m) = 1} (17.33) mn. (17.34) mn, which is not useful. Now try 68 X |Snχ |4 = X X χ(i1 i2 )χ(j1 j2 ) (17.35) χ (m) i1 ,i2 ,j1 ,j2 χ (m) m · #{i1 , i2 , j1 , j2 ∈ [1, m] : i1 i2 ≡ j1 j2 (mod m)} (17.36) 2 n X m τr (c1 )τr (c2 ) (17.37) c1 ,c2 =1 c1 ≡c2 (m) 1+ m 1 1 So |Snχ | m 4 + n 2 + 18 n 1 n4 n + m 2 . (17.38) , and we only obtain the trivial bound. m4 The Burgess Estimate P 1 2 We want to bound SχN = 2N n=N χ(n) for χ a primitive character modulo q and N ∼ q . Here is the idea: if ab is small compared to N , SχN = 2N X χ(n + ab) + O(ab) (18.1) n=N = χ(a) 2N X χ(an + b) + O(ab). (18.2) n=N Let A, B ∈ N with AB small compared to N . Then ABSχN B X 2N A X X χ(n + ab) + O(A2 B 2 ) a=1 b=1 n=N A B X X = χ(a) a=1 (18.3) χ(an + b) + O(A2 B 2 ) b=1 B 2N X X χ(an ≤ + b) + O(A2 B 2 ) a=1 n=N b=1 B X X χ(y + b) + O(A2 B 2 ) ≤ v(y) A X y (q) (18.4) (18.5) (18.6) b=1 for v(y) = #{1 ≤ a ≤ A, N ≤ n ≤ 2N : an = y}. We have X v(y) = AN (18.7) y (q) X v(y)2 = #{(a1 , n1 ), (a2 , n2 ) : a1 n2 ≡ a2 n1 y (q) 69 (mod q)}. (18.8) If AN < q, congruence implies equality, so X 2 v(y) ≤ AN X σ0 (r) (18.9) r=1 y (q) AN q . (18.10) Hölder: pick r ∈ N. Then 1− 1 1 B 2r 2r1 r 2r X X X X 2 2 2 v(y) v(q) . A B + χ(y + b) ABSχN y (q) b=1 y (q) (18.11) y (q) Therefore 1 ABSχN q (AN )1− 2r 1 2r r X X Y y + bi . χ y (q) 1≤b1 ,...,br ≤B i=1 y + ci | {z } 1≤c1 ,...,cr ≤B (18.12) Q(y) P 1 The Weil bound gives y (q) χ(Q(y)) q 2 unless χ(Q(y)) ≡ 1. This occurs if ord χ = k ord Q(y) is a kth power. The number of bi , ci for which this happens is B r , since bi , ci take on at most 2r k ≤ r values. So now 1 1 1 ABSχN A2 B 2 + q (AN )1− 2r (B r q + B 2r q 2 ) 2r 2 2 1 1− 2r A B + q (AN ) 1 2r−1 1 2 (B q 1 2r (18.13) 1 4r + Bq ). (18.14) 1 For B = q 2r and A = N 2r+1 q 2(2r+1) , −1 2 − r+1 SχN q N N 2r+1 q 2r(2r+1) . r+1 (18.15) 1 1 This is useful if N > q 4r . We can choose r large enough if N > q 4 +δ . If N ∼ q 2 , we can pick 1 1 r = 2, lowering our bound by q 24 . (We could get q 16 savings if we do more work to improve the initial O(ab) bound.) 19 Amplification Recall that we have X |SχN |4 N 2 q 1+ . χ (q) Here are some ideas to improve this: 70 (19.1) • Use higher moments (such as replacing 4 with 6), but this usually doesn’t help much. • Shorten the character sum; this works in certain cases, like q highly composite, but not if q is prime. • Use amplification. We’ll explain this last part. Let L be a small power of N . Then look at X χ (q) L 2 X |SχN |4 c` χ(`) (19.2) `=1 for (c` ) general in C. We hope for a bound of the form N 2 q 1+ PL `=1 |c` | 2, which we call (?). Expanding our sum, X `1 ,`2 2N X c`1 c`2 X χ(m1 n2 n3 n4 `1 `2 ) = n1 ,n2 ,n3 ,n4 =N χ (q) N ≤n ,n ,n ,n ≤2N 1 2 3 4 c`1 c`2 ϕ(q)# n1 n2 `1 ≡ n3 n4 `2 . X (19.3) `1 ,`2 If we assume N 2 L < q, we get 2 N X #{n1 n2 `1 = n3 n4 `2 } σ0 (r) (19.4) . (19.5) r=1 `1 |r (`1 ,`2 ) N 2+ `1 (`1 ,`2 ) So our sum is ϕ(a)N 2+ L X |c` |2 . (19.6) `=1 1 But we want a similar bound for N 2 L > q. Suppose that (?) holds for N ∼ q 2 and L = q δ for 2 q 1+ some δ > 0. Then take c` = χ(`). Then |SχN |4 N L . We now deal with X L 2 X χ 6= 1|SχN |4 c` χ(`) . (19.7) `=1 Fact. (r, q) 6= 1 0 X χ(r) = −1 (r, q) = 1, r 6≡ 1 (mod q) χ6=1 ϕ(q) − 1 r ≡ 1 (mod q). 71 (19.8) So we get L X c`1 c`2 (ϕ(q) − 1)#{N ≤ n1 , n2 , n3 , n4 ≤ 2N : n1 n3 `2 = n2 n4 `1 } (19.9) `1 ,`2 =1 − #{N ≤ n1 , n2 , n3 , n4 ≤ 2N : n1 n3 `2 6= n2 n4 `1 } . (19.10) Remark. For n1 , n2 , n3 , n4 , n1 n2 n3 n4 has one of ϕ(q) values.If `1 = `2 , we have at least 2N 2 solutions. Now n1 n3 `2 = n2 n4 `1 =⇒ ∃ h : −L ≤ h ≤ L such that n1 n3 `2 − n2 n4 `1 = hq. We want to understand the asymptotics of the shifted divisor sum 2 4N X σ0 (a)σ0 (b). (19.11) a,b=N 2 `2 a−`1 b<qh A model problem is given by X X σ0 (a)σ0 (a + 1) = X X σ0 (a2 + a). (19.12) a=1 a=1 Non-shifted sums are easier. For example, fixing σ > 1 and σ 0 < 1, X Z ζ(s)2 xs σ0 (n) = <(s)=σ n≤x Z ds s (19.13) ds ζ(s)2 xs + Ress=1 s s (19.14) ζ(s)2 xs = <(s)=σ 0 = x log x + x + O(x1−δ ). (19.15) 1 Alternatively, count lattice points under a hyperbola. This gives a bound of O(x 2 ), and we can improve this further with more work. The shifted convolution problem: let f, g be (cuspidal) maass forms, and we want to understand the sum X X λf (n)λg (n). (19.16) m,n=1 `1 m−`2 n=h Morally speaking, X σ0 (m)σ0 (n) ≈ P |n|≤X σ0 (an2 + bn + c). (19.17) |n|≤X m,n≤X `1 m−`2 n=k Theorem 19.1 (Bykovskii). X 2 σ0 (n2 + D) = c1 (D)X log X + c0 (D)X + O,D (X 3 + ). 72 Proof. σ0 (b2 + D) = #{a, c ∈ N : b2 + D = ac}. Let K(D) = {(a, b, c) ∈ Z : b2 − ac =√ −D}. −1 b −a Γ = SL2 (Z) acts on K(D) by γ · cb −a γ . Also K(D) → H by (a, b, c) 7→ b+ a−D = −b = γ c −b √ pc D Z(a,b,c) . We have |Za,b,c | S= a and =(Za,b,c ) = a . By finiteness of the class group, there exist Z1 , . . . , Zλ(D) such that i Γ/ΓZi = im(K(D)). q P 2 +D 2 Choose h ∈ C(R+ ) such that n∈Z σ0 (n + D)h( n D ) converges absolutely. We can rewrite this sum as ∞ X X √ ac D h b∈Z a,c=1 λ(D) X = 1 X h |ΓZk | σ∈Γ k=1 |σZk | =(σzk ) (19.18) λ(D) 1 ϕ(h, Zk ); |ΓZk | k=1 X |σz| ϕ(h, z) = . h =(σz) X = (19.19) (19.20) σ∈Γ Observe that ϕ is Γ-invariant. Idea: L2 (Γ\H) = C · 1 ⊕ Eis ⊕ L20 (Γ\H) with L20 (Γ\H) = ∞ M (19.21) Cuj i=1 ∆uj = λj uj X √ uj (x + iy) = ρj (n) yKirj (2π|n|y)e(nx) (19.22) (19.23) n∈Z 1 + rj2 = λj . 4 (19.24) Now we have ϕ(h, z) = c0 + X hϕ, uj iuj (z) + (Eis part). (19.25) j But ϕ 6∈ L2 . Define u (z, s, s0 ) = X =(σz) s |σz| σ∈Γ (=(σz))s0 . (19.26) If <(s0 ) > 0, then u(z, s, s0 ) ∈ L2 (Γ\H). Now u (z, s, s0 ) = X σ∈Γ∞ \Γ =(σz)s+s0 X n∈Z 73 s ((=(σz))2 + (<(σz) + n)2 )− 2 . (19.27) Applying Poisson summation with respect to x to √ X s+1 (=(σz)) σ∈Γ∞ \Γ 1 n∈Z (y 2 +(x+n)2 ) 2s P , we get πΓ s−1 2 (=(σz))1−s Γ 2s (19.28) ! s s−1 1−s 2π 2 X + |m| 2 (=(σz)) 2 K s−1 (2π|m|=(σz))e(m=(σz)) 2 Γ 2s m6=0 (19.29) √ s+1 πΓ s−1 2(2π)−s0 i− 2 X 2 E(1 + s0 , z) + |m|−1−s ρm (z, s, s0 ) = √ Γ 2s πΓ 2s m6=0 (19.30) where ρm,s,s0 (z) = X (2π|m|=(σz)) σ∈Γ∞ \Γ | s+1 +s0 2 K s−1 (2π|m|=(σz)) e(m<(σz)). {z 2 } (19.31) Fm (s,s0 ,=(σz)) Now Z hρm,s,s0 , 1i = Fm (s, s0 , y)e(mx) dx (19.32) Γ∞ \H =0 (19.33) √ hρm,s,s0 , uj i = pj (m) 2πm Z Γ∞ \H ρm,s,s0 , E √ 1 + it, • = 2πm 2 Γ Fm (s, s0 , y)Kirj (2π|m|y) 1 1 2 2π 2 −it d 1 (m) − it ζ(1 − 2it) 2 +it dy 3 (19.34) y2 Z ··· . (19.35) We end up with 1 u(z, s, s0 ) = · · · + L uj , + s0 b(uj ) + Eis (19.36) 2 j Z ∞ ∞ X 1 1 12 log u 2 ϕ(h, z) = h(u) + c1 + c2 (n)s η(z) du + b0,h (rij )L uj , uj (z) + Eis. π 2 1 X j=1 (19.37) Kronecker’s limit formula: E(z, s) = 1 3 6 (s − 1) + (γ − log(z) − log |y 2 η(z)|) + O((s − 1)2 ). π π Fi Me Where does this proof actually end, if ever? (20) Now for f, g ∈ Sk (Γ), consider L(f × g, 12 ). We need to understand 74 (19.38) X λf (n)λg (n) (19.39) m,n≤X `1 m−`2 n=h where f (z) = P∞ m=1 m k−1 2 λf (m)e(mz). Now for N = `1 `2 , define X ph,s = =(γz)s e(−h<(γ1 γ2 )) (19.40) γ∈Γ∞ \Γ0 (N ) V (z) = f (`1 z)g(`2 z)y k . (19.41) Then V (z) is Γ0 (N )-equivariant. We have Z hph,s , V (z)iN = Γh,s V (z) Γ0 (N )\H Z dx dy y2 (19.42) y k+s f (`1 z)g(`2 z)e(−hx) = Γ∞ \H Z y k+s = Γ∞ \H X dx dy y2 λf (m)λg (n)(mn) k−1 2 (19.43) e((m`1 − n`2 − h)x)e2π(m`1 +n`2 )y m,n dx dy y2 (19.44) = Γ(k + s − 1) (2π)k+s−1 X λf (m)λg (n) m,n `1 m−`2 n=h √ mn `1 m + `2 n k−1 , (`1 m + `2 n)s . {z | (19.45) } D(f,g,s) We have Z X s e D(f, g, s)G(s)x ds = <(s)=2 λf (m)λg (n) m,n `1 m+`2 n=h √ mn `1 m + `2 n k−1 G `1 m + `2 n x . (19.46) Now hph,s , 1i = 0 Z hph,s , uj i = (19.47) Γ\H z s−1 ρj (h) = Γ |h|s−1 So for 1 4 dx dy y2 ! s − 21 + itj Γ 2 e(−hx)y s uj (z) + t2j = λj , we obtain 75 (19.48) s− 1 2 − itj 2 ! . (19.49) ph,s (z) = ∞ X j=1 z s−1 ρj (h) Y Γ uj (z) |h|s−1 ± s− ! 1 2 ± itj 1 + Eis. (19.50) To continue this beyond s = 1, we need <(s − 12 + itj ) > 0. Since the uj are self-adjoint under ∆, + t2j = λj > 0. We need that for some δ > 0, λj > δ for every uj and for every N . 1 4 Selberg’s eigenvalue conjecture: λj ≥ 41 ; equivalently tj ∈ R. We have D(f, g, s) = X hy k f (`1 z)g(`2 z), uj (z)i. (19.51) j Understanding Z s e D(f, g, s)G(s)x ds (19.52) <(s)=1−δ requires knowing the growth of D(f, g, s) on vertical strips to be controlled. We have D(f, g, s) = hph,s , V i. By Plancherel, this is X hph, suj ihV, uj i + Eis = X 2s−1 ρj (h) Y j Suppose s = 1 2 j |h|s−1 Γ s− 1 2 ± ± itj 2 ! hV, uj i. (19.53) + iR. Then Γ iR + itj 2 π|t| iR − itj Γ : Γ(it) ≈ e− 2 . 2 (19.54) Consider the following: • |R| > |tj | : e− • |R| < |tj | : e− π|R| 2 . π|tj | 2 . What now needs to be shown is: Proposition 19.2 (Sarnak’s triple product estimate). k k hV, uj i = hy 2 f (z)y 2 g(z), uj (z)i = O(e− π|tj | 2 |t|A ). (19.55) Recall that Dh (f, g, s) = (2π)s+k−1 (` 1 `2 s− 12 |h| 1 )k− 2 2s · ∞ X Γ ρj (−h) j=1 76 s− 12 +itj 2 1 s− 2 −itj Γ 2 Γ(s + k − 1) hV, uj i + Eis. (19.56) Suppose s = σ + iR. Then Γ(s) ∼ e− |R+tj |+|R−tj |−2R −π 2 RA e πR 2 Rσ−1 so the absolute value of the Γ factors is ∼ . In order to control the growth of the terms with small tj , we need |ρj (−h)|hV, uj i |tj |A (19.57) for some A > 0. As an aside, applying ∆ A times, ∆(A) V = ∞ X A j cj λj uj P + Eis for cj = hV, uj i, so (A) c2j t2A V, ∆(A) V i. j ≤ h∆ (19.58) j=1 So we already know polynomial decay of the cj . But this is not strong enough! πt Lemma 19.3. e 2 (tj N )−1− |ρj (1)| e πtj 2 (tj N ) . Proof. 1 = huj , uj i Z dx dy π uj (z)2 E(z, s) 2 = Ress=1 3 y Γ\H Z ∞Z 1 π dx dy = Ress=1 uj (x + iy)2 y s 2 3 y 0 0 Z ∞ ∞ X dy |ρj (n)|2 Kiy (2π|n|y)y s = Ress=1 y n=1 0 ∞ X |ρj (n)|2 1 1 = Ress=1 Γ + it − it Γ j j . ns 2 2 n=1 {z } | (19.59) (19.60) (19.61) (19.62) (19.63) ∼e−πtj Let L(s) = P∞ n=1 |ρj (n)|2 ns X n and c = Ress=1 L(s). Let h ∈ Cc∞ ( 21 , 1). Then 2 |ρj (n)| h n x Z L(s)e h(s)xs s = (19.64) <(s)=2 Z = ce h(1)x + <(s)= 12 1 L(s)e h(s)xs dx. 1 (19.65) By the convexity bound, the second integral is x 2 P O ((tj N ) 2 + ). Also ρj (n) = ρj (1)aj (n) for mn aj (n) the Hecke eigenvalues of uj . Then aj (n)aj (m) = d|(m,n) aj d2 . Now we have 77 X X !2 aj (n)2 X X = n=1 (aj (n)aj (m))2 (19.66) n,m=1 X ≤ mn 2 aj d m,n=1 · d(m, n) (19.67) d|(m,n) 2 ≤ X X |aj (n)|2 σ0 (n) n=1 X σ0 (m) . m m≤X | {z } (19.68) (log 3X)2 PX 2 2 2 2 2 Pick X c12 (tj N )1+ . Then n=1 aj (n) ∼ cX, so c X cX (log 3X) implying c P 1 2 (log 3X)2 so that c (tj N ) . For the lower bound on c, X n=1 |aj (n)| ≥ 1 so c ≥ X . In fact (we won’t prove this): |ρj (1)| ∼ e 20 πtj 2 |tj N |o(1) . Applications of Subconvexity • Let f, g be distinct Hecke cusp forms. We want to know the smallest n = N (f, g) such that λf (n) 6= λg (n). Suppose f ∈ S2 (Γ0 (Nf )) and g ∈ S2 (Γ0 (Ng )). Then for q = e(z), f (z)dz − g(z)dz = X (λf (n) − λg (n))q n−1 dq (20.1) n≥1 so N (f, g) is the order of vanishing of (f (z) − g(z)dz at ∞, implying N (f, g) ≤ deg Ωx0 [Nf ,Ng ] ≤ 2gen(· · · ) + 1 = [Nf , Ng ]1+o(1) . (20.2) Now suppose V ∈ Cc∞ ( 21 , 1). Consider X (f, g, N ) = X λf (n)λg (n)V n n N . P P P Note that if N = N (f, g), then (f, g, N ) = (f, f, N ). We have L(f ×g, s) = n so that X 1 (f, g, N ) = 2πi Z L(f × g, s)Ve (s)N s d (20.3) λf (n)λg (n) , ns (20.4) (2) Z = δf =g Ve (1)N Ress=1 L(f × f, s) + ( 21 ) L(f × g, s)Ve (s)N s ds. 1 (20.5) The convexity bound is L(f × g, 21 + it) (|t| + [Nf , Ng ]) 2 + . Also Ress=1 L(f × f, s) = [Nf , Ng ]o(1) . We end up with N (f, g) [Nf , Ng ]1+ . A subconvexity bound L(f ×g, 21 +it) 1 (|t| + [Nf , Ng ]) 2 +−δ would give N (f, g) [Nf , Ng ]1+−2δ . 78 1 . Fact (Kowalki-Michel-Vanderkam). This holds for δ = 12 − 82 √ • Let K = Q( −D) for D positive and squarefree. Recall √ the Minkowski bound: for every √ 2 a ∈ Cl(K), there exists I OK with [I] = a and N (I) ≤ π D. As an aside, I = (2a, −b+ D) √ √ √ D ∈ H with h(zI ) = 2aD . So constant D is optimal. has N (I) = 2a, and have zI = −b+ 2a Fix a character ψ : Cl(K) → C× . We want to know N (ψ); that is, we want to determine the existence of I with N (I) = N (ψ) and ψ(I) 6= 1. Let θψ (z) = X IOK • ψ(I)e(N (I)z) ∈ M1 Γ0 (D), . D (20.6) θψ is cuspidal unless ψ is real. We have L(ψ, s) = ζK (s) = X ψ(I) N (I)s IOK X 1 IOK N (I)s (20.7) . (20.8) Taking X = N (ψ), X N (I)≤X V N (I) X = X ψ(I)V N (I)≤X N (I) X (20.9) so that Z L(ψ, s)Ve (s)X s ds = (2) Z ζK (s)Ve (s)X s ds. (20.10) (2) 1 1 A subconvexity bound of the form L(ψ, 21 ) D 4 +−δ implies N (ψ) D 2 +−2δ . It turns 1 out that this is possible for δ = 24001 . Suppose Cl(K) is cyclic. We want to know the smallest N (I) among I generating Cl(K). Let G = Cl(K) and δG (a) be 1 if a generates G and 0 otherwise. Then δG (a) = X µ(d) X ψ(a). d d d||G| (20.11) ψ =1 • Quantum Chaos: Let X be a (compact) hyperbolic manifold, ∆x a (the?) Fix Me (21) LaplaceBeltrami operator, and u1 , u2 , . . . , uj , . . . with ∆x uj = λj uj . We have λj → ∞ and that the uj span L2 . Consider the measure d|uj |2 on X. It is conjectured that d|uj |2 converges to the hyperbolic measure on X. Arithmetic Q.U.E.: In L2 (Γ\H), u1 , . . . , uj , . . . Hecke maass forms. Then: Theorem 20.1. For ϕ ∈ Cc∞ (Γ\H), Z Γ\H dx dy j→∞ ϕ(z)|uj (z)|2 2 −−−→ y 79 Z ϕ(z) Γ\H dx dy . y2 (20.12) An example is given by |E( 12 + it, z)|2 dxy2dy as t → ∞. Proof. We have Z u1 (z)E Γ\H 1 + it, z 2 1 1 dx dy u1 (z)E + it, z y 2 −it 2 2 y Γ∞ \H 1 1 L u1 , + it Gt = L u1 , 2 2 dx dy = y2 Z (20.13) (20.14) 1 for Gt consisting of gamma functions. So Gt ∼ |t|− 2 . So decay to 0 of the integral amounts to a subconvex bound on L(u1 , 12 + it). For the case of cusp forms, a subconvex bound on L(f ⊗ g ⊗ h, 12 + it) would imply the result, but we don’t know this. 80 To do. . . 2 1 (p. 4): There’s a gap here. Fill in? 2 2 (p. 14): Rectangular contour. 2 3 (p. 15): Rectangle 2 4 (p. 16): Fix Me Fix the (?). 2 5 (p. 23): Fundamental domain picture 2 6 (p. 23): Actually determine the genus. 2 7 (p. 26): Fix Me Maybe deal with the overset alignments? 2 8 (p. 26): Fix Me We may have confused notation. 2 9 (p. 29): Fix Me This part appears in the notes after equidistribution. May want to organize differently? 2 10 (p. 35): Fix Me Are they both d0 , or does d appear? 2 11 (p. 41): Fix Me Disgtinguish between p and P . 2 12 (p. 50): Can the (?) be referenced? 2 13 (p. 50): What is the set the sum runs over? 2 14 (p. 52): Fix Me Should all of 2πi be raised to power? 2 15 (p. 53): Fix Me More stuff on this topic appears later. Reorganize? 2 16 (p. 56): Fix Me We have more quadratic stuff now. Reorganize? 2 17 (p. 60): Fix Me One of these should have a tilde. 2 18 (p. 60): Fix Me Now we’re doing L-functions of two modular forms. 2 19 (p. 62): Fix Me Back to maass forms, maybe? 2 20 (p. 74): Fi Me Where does this proof actually end, if ever? 2 21 (p. 79): Fix Me Which word should go here? 81