EISENSTEIN SERIES AND THE RIEMANN ZETA-FUNCTION By D. ZAGIER* IN THIS PAPER we will consider the functions E(z, p) obtained by setting the complex variable s in the Fisenstein series E(z, s) equal to a zero of the Riemann zeta-function and will show that these functions satisfy a number of remarkable relations. Although many of these relations are con- sequences of more or less well known identities, the interpretation given here seems to be new and of some interest. In particular, looking at the functions E(z, p) leads naturally to the definition of a certain representa- tion of SL,(R) whose spectrum is related to the set of zeroes of the zeta- function. We recall that the Eisenstein series E(z,s) is defined for z= x + iy ¢ H (upper half-plane) and s ¢ C with Re(s) > I by = =. a FeI= SY mra= > > a: yr Ar ea where I'= PSL,(Z), Pa = {4() ln Z}CT If we multiply both « sides of (1) by £(2s) ~® and write m = re, n = rd, we obtain rat q . @) £028) Es) = = > — 2 |mz+n| where 2 indicates summation over all (m,n) € Z{ (0,0) }. The function £(2s)E(z,s) has better analytic properties than E(z,s); in particular, it has a holomorphic continuation to all s except for a simple pole at s = 1. There is thus an immediate connection between the Eisenstein series at s and the Riemann zeta-function at 2s. This relationship has been made use of by many authors and has several nice consequences, two of * Supported by the Sonderforschungsbereich “Theoretische Mathematik” at the University of Bonn,D. ZAGIER which will be mentioned in §1. Our main theme, however, is that there is also a relationship between the Eisenstein series and the zeta function at the same argument. We will give several examples of this in § 2. Each takes the form that a certain linear operator on the space of functions on I'\H, when applied to E(-,s), yields a function of s which is divisible by &(3). Then this operator annihilates all the E(-, p), and it is natural to look for a space & of functions on T'\H which contains all the E(-, p) and which is annihilated by the operators in question. Such a space is defined in § 3. In §4 we show that 6 is the set of K-fixed vectors of a certain G-invariant subspace ¥ of the space of functions on I"\G (where G = PSL,(R), K = PSO(2)). Then Vis a representation of G whose spectrum with respect to the Casimir operator contains p(1—p) discretely with multiplicity (at least) n if p is an n-fold zero of £(s). In particular, if (as seems very unlikely) one could show that 7 is unitarizable, ie. if one could construct a positive definite G-invariant scalar product on’, then the Riemann hypothesis would follow. The paper ends with a discussion of some other representations of G related to ¥ and reformulation in the language of adeles. §1. We begin by reviewing the most important properties of Eisenstein series. a) Analytic continuation and functional equation. The function E(z,s) has a meromorphic continuation to all s, the only singularity for Re(s) >+ being a simple pole at s = 1 whose residue is independent of z: @) res,., E(z,s) = 3 (WzeH). 7 The modified function (o) E*(z,s) = w * T(s) £(2s) E(z,s) is regular except for simple poles at s = 0 and s = | and satisfies the functional equation (5) E*(z,s) = E*(@,1-9). ‘These statements are proved in a way analogous to Riemann’s proof of 276EISENSTEIN SERIES the analytic continuation and functional equation of {(s): we rewrite 1 <1 3 (@o-n dt, where Q (m,n) (ze H) denotes the quadratic form : om (mn) = meta —n1Q,o.0) wv of discriminant —4 and @,(t) = > e the corresponding mneZ theta-series; then the Poisson summation formula implies © ,(+) =t® ,(0) and the functional equation and other properties of E(z,s) follow from this and equation (6). b) “Rankin-Selberg method”. Let F: HC bea I’-invariant function which is of rapid decay as y > @ (i.e. F(x + iy) = 0(y~%) for all N). Let 1 8) C(F:y)= J FO + iy)dx (>) be the constant term of its Fourier expansion and ~ ) 1(F;s) = f C(F:yyy"* dy (Re(s)>1) the Mellin transform of C(F;y). From (1) we obtain (10) Fs) = J F@y'dz= { FQ)EG,s) dz, rw r\w xd) ‘ di where dz denotes the invariant volume element as . Therefore the properties of E(z,s) given in a) imply the corresponding properties of I(F;5): it can be meromorphically continued, has a simple pole at s = 1 with on) res... (Fs) =2 [ F@dz, mw rH and the function (12) I*(F;s) = 27° Xs) £(2s) (Fs) 277D. ZAGIER is regular for s + 0,1 and satisfies (13) I*(F;s) = I*(F; 1-5). ©) Fourier development. The function E*(z,s) defined by (4) has the Fourier ame (14) E*(z,s)=L*(2s)y" + 0" @s— yt +2/y > Be where (as) S@O=7 "TEL GeO), on)= Dd" (ne N,v eC), fin (16) K,) = fe cosh vudu (ve €,t>0). The expansion (14), which can be derived without difficulty from (2), gives another proof of the statements in a); in particular, the functional equation (5) follows from (14) and the functional equations g"(9) = 2" 9), o,(n) =n" o_,(n), K (= K_,(t). Because of the rapid decay of the K-Bessel functions (16), equation (14) also implies the estimates an Gg E@=00 mx@l—) og"y) (n = 0,1,2,-.., 0 = Re(s), s y=Im@) +») for the growth of the Eisenstein series and its derivatives. Finally, it follows from (14) or directly from (1) or (2) that the Eisenstein series E(z,s) are eigenfunctions of both the Laplace operator ra oe? Any ( tet oy ) and the Hecke operators Tn): Fe > > Dy FEz+P) >, sd=n (med) ad>o 278EISENSTEIN SERIES namely (18) A EG@,s)=s(s—1) E@,s), Tin) E(@,s)=n" @,_,(n) E(z,s). We now come to the promised applications of the relationship between E(e,s) and £(2s). The first (which has been observed by several authors and greatly generalized by Jacquet and Shalika [3]) is a simple proof of the non-vanishing of £(s) on the line Re(s) = 1. Indeed, if £ (1 + it) = 0, then (14) implies that the function F(z) = Ee a 7 it) is of rapid decay, does not vanish identically, and has constant term C(F;y) identically equal to 0. But then I(F;s) = 0 for alls, and taking s= 4 —}itin (10) we find { |F(2) dz=0, a contradiction. rn ‘The second “application” is a direct but striking consequence of the Rankin-Selberg method. Let €,¢ I'\Hbe the horocycle T..\(R + iy): it is a closed curve of (hyperbolic) length ; The claim is that, as yO, the curye @, “fills up” I'\H in a very uniform way: not only does ¢, meet any open set Uc I"\H for y sufficiently small, but the fraction of , contained in U tends to vol(U)/vol(I'\H) as y+ 0 and in fact length(#,U) —_ vol(U) ——— + ay?" 0); length (@,) vol(P\H) > oo! moreover, if the error term in this formula can be replaced by O(y’!* ~*) for all U, then the Riemann hypothesis is true! To see this, take F(z) in b) length (#, U) iength (@,) and the Mellin transform I(F,s) of this is holomorphic in Re(s)>4@ to be the characteristic function Xy of U. Then C(F;y) = (where @ is the supremum of the real parts of the zeroes of { (s)) except . : 3 vol(U) for a simple pole of residue x == f F(2)dz=“C0) _ mre [ q ° vol(I"\H) F were sufficiently smooth (say twice differentiable) we could deduce that ats=L.If 1(F; @-+ it) = 0(¢) on any vertical strip Re(s) = @> 4, and the Mellin 1 pore inversion formula would give C(F;y) = « +0(y 209 ). For F=XyD. ZAGIER we can prove only C(F;y)=« + 0(y'”*); conversely, however, if C(F;y) =k +0(y’) then I(F;s)— oT is holomorphic for Re(s)> 1 — @, and if this holds for all F = Xy we obtain @<2(1 — «). §2. In this section we give examples of special properties of the functions E"(z,p) or, more generally, i Oe functions (19) FQ) = F,,, @) = om or E*(,5)| (O< mD, Be, y= = (iply” ToD ta ID)” + =~((D/) rTOt@ = 4 (BI) Pe) HOLD, 280EISENSTEIN SERIES where £, (s) is the Dedekind zeta-function of K and L(s,D) the L-series x ( 2 )n ~*. Since the latter is holomorphic, we deduce that the = hw function 3 E*(zq ,s) is divisible by I'(s)¢ (s), i.e. that it vanishes with multiplicity > n, ata non-trivial zero p of { (s). A similar statement holds for any negative integer D congruent to 0 or | modulo 4 (not necessarily the discriminant of a quadratic field) if we replace £ (8) in the equation _above by the function boy 20) (&D) = oa . , Qmny Qymm>o | where Q, (i = 1,...,h(D)) are representatives for the T' -equivalence classes of binary quadratic forms of discriminant D and T’9, denotes the stabilizer of Q, in I". Again the quotient L(s,D) = ¢ (s,D)/£(s) is entire ([11], Prop. 3, iii), p. 130). This proves Proposition 1; Each of the functions (19) satisfies hd) fe Flzq) =0 for all integers D<0, where Zo,.---sZayp, are the points in T\Hwhich satisfy a quadratic equation with integral coefficients and discriminant D. Notice how strong condition (21) is: the points satisfying some quadratic equation over Z (“points with complex multiplication”) lie dense in I’ \H, so that it is not at all clear a priori that there exists any non-zero continuous function F: T\H— € satisfying eq. (21) for all D<0. Example 2: This is the analogue of Example | for positive discriminants. Let D > 0 be the discriminant of a real quadratic field K and Q,,..-,Q yoy representatives for the I’ -equivalence classes of quadratic forms of discriminant D. To each Q, we associate, not a point z9,€ T’\Has before, but a closed curve Coc F'\H as follows: Let w,, w’,« R be the roots of 281D. ZAGIER the quadratic equation Q, (x, 1) = a,x? + b,x + c, = 0 and let Q, be the semicircle in H with endpoints w, and w’, . The subgroup : oa (22) To=[a(Ho" : ) tueZ,@—Du?=4 ae Feesbi of I’, which is isomorphic to { units of K}/{+ 1} and hence to Z, maps Q, to itself, and Co, is the image P'g\Q, of Q, in P\H. On Co,we have a measure |dg, z|, unique up to a scalar factor, which is invariant under the operation of the group I'y_@ R obtained by replacing Z by R in (22); if we parametize Q, by | w,ip+ w, ee O< ©) ee ape ( P i E%@,9)|dq z|= 9" DY TG) ist a which again is divisible by £(s), and as before we can take for D any positive non-square congruent to 0 or | modulo 4 and get a similar identity with Z, (s) replaced by the function (20). Thus we obtain Proposinion 2: Each of the functions (19) satisfies ND) ae J. Flz)|doz|=0 ar 8 for all non-square integers D <0, where Q, (m,n) = am? + b,mn-+ jn” 282EISENSTEIN SERIES (i=1,...,h(D)) are representatives for the T-equivalence classes of binary quadratic forms of discriminant D, Cg, is the image of {z=x+iy« HJ a,izi + bx + ¢, = O}inT\H, and wn Z| VD ((dx)* + (dy)*) |4,2|= > _ J Jaz?+bz+0, Example 3: The third example comes from the theory of modular forms. Let f(z) be a cusp form of weight k on SL,(Z) which is a normalized eigenfunction of the Hecke operators, i.e. f satisfies f (2#2)- (z+ )'@ (VzeH, (ye SL,(Z)) and has a Fourier development of the form 7 t= ae" oa if (n,m) = I. Define D,(s) by 1 po = TT rey * (l—a2p ")(1—a,8,p ")(I—B2p *) where a, and, are the roots of X‘ —a,X +p" | =O;itis easily checked that _ EQs~ 2k +2) A d= Dias witha, = landa,, =a (Re(s) => 0), Applying the Rankin-Selberg method (10) to the I'-invariant function F(z) = y*|f(z)|? with constant term C(F;y")= 2! al? € J y* |f(@|? E*(z,s)dz = I°(F;s) —4eny POE Osa) TG +k ee wT nett teat =4 TOT +k—10()D, (6+ k—-1). 283D, ZAGIER This formula, which was the original application of the Rankin-Selberg method ({5],[6]), shows that the product {(s)D,(s+k—1) is holo- morphic except for a simple pole at s = 1. It was proved by Shimura [7] and also by the author [11] that in fact D,(s) is an entire function of s. Thus the above integral is divisible by I’ (s) P(s +k — 1)£(s), and we obtain Proposition 3: Each of the functions (19) satisfies (24) J y*ltz)? Flzjdz = 0 rw for every normalized Hecke eigenform f of level 1 and weight k. The statement of Proposition 3 remains true if f is allowed to be a non- holomorphic modular form (Maass wave-form); the proof for k= 0 is given in [12] in this volume and the general case is included in the results of [i] or [4]. Finally, we can extend our list of special properties of the functions (19) by observing that each of these functions is an eigenfunction of the Laplace and Hecke operator (eq. (18)) and hence (trivially) has the Property that (25) A F(z) and T(n)F(z) satisfy Propositions 1-3 for all i>0, nal. Note that for general functions F:E\H > € (not eigenfunctions), eq. (25) expresses a property not contained in Propositions I to 3: for example, eq, (21) for D = — 4 says that F(i) = 0, and this does not imply AF(i)=0. §3. In §2 we proved that the functions E*(z, p), and more generally the functions (19), satisfy a number of special properties. In this section we will both explain and generalize these results by defining in a natural way a space & of functions in I\H which contains the functions (19) and has the same special properties. Let D be an integer congruent to 0 or ! modulo 4. For ®: R>Ca function satisfying certain restrictions (e.g. (27) and (29) below) we define anew function £,@: H + C by 284EISENSTEIN SERIES 2 26) -¥,0(2) = ; > ® ( abl +on+e) (=x+iyeH), bP dace D where the summation extends over all integral binary quadratic forms Q(m,n) = am? + bmn + cn’ of discriminant D. Since Q and —Q occur together in the sum, we may assume that ® is an even function; the factor has then been included in the definition to avoid counting each term twice. The sum (26) converges absolutely for all z ¢ H if we assume that en ®(X) = 0(|x|'*) (|X| +) oa for some © > 0. Moreover, the expression 51+ BREE is unchanged if one acts simultaneously on x + iy H and Q(m,n) = am? + bmn + cn? by an clement ¥ eI’. Hence %,0(¥z) = ¥,@(2), so, is an operator from functions on R satisfying (27) to functions on '\H. Before going on, we need to know something about the growth of ¥,inT\H. If D is not a perfect square, then a + 0 in (26), so (for ® even) © < a(x + b/2a)’— D/4a £,0@) = > : (ys) ant by-2 bie Dimod4s) -i-e S of (ay 24 *8A8"= Pl) ) Ms It ot as y = Im(z) + , where (28) n,(a) = # {b(mod2a)|b? = D(mod4a) }. ee eg 2 n,(a) Since the integral is O(a”'y~*) and 2 —>— converges, we find aia 285D. ZAGIER £,O(2) = Oly"). If D is a square, then the same argument applies to the terms in (26) with a + 0 and we are left with the sum S by : mK + C. 22, Dey oe to estimate. If @ is sufficiently smooth (say twice differentiable), then the inner sum differs by a small amount from the corresponding integral J a@e=y- J coax, and this will be small as y > © only if { (X)dX vanishes. Thus with the requirement a! (29) @isC and f @(X) dX =0 if Disa square as well as (27) we have ¥,@ (2) =O(y~*) as y > © for all D, and so the scalar product of ¥,@ with F inI’\H converges for any F: T\H+>C satisfying F(z) =O(y'~*) for some ¢ > 0 (or even F(z) =O(y)). There- fore the definition of & in the following theorem makes sense. THEOREM: For each integer D ¢ Z, D = 0 or I (mod 4), let even functions ® : R-> C 78 setisfping (27) and (29) —> { functions T\H > € } be the operator defined by (26). Let & be the set of functions F: T\H > C such that a) F(z) =O(y'~*) for some e >0 6) F(z) is orthogonal 10S tm(,), te. { £2) F(z) dz = 0 for all D € Zand all © satisfying (27) and (29). Then i) @ contains the functions (19) ; li) 6 is closed under the action of the Laplace and Hecke operators ; iii) Any F € & satisfies (21), (23) (for all D) and (24) (for all f). Proor. i) The functions (19) satisfy a) because of equation (17), since 0 C a function which is O(y*)as y > © for somea< 1 +e. If Q,(m,n) = am’ + b,mn+c,n" (i = 1,...,h(D)) are representatives for the classes of binary quadratic forms of discriminant D, then any form of discriminant D equals Q,°¥ for a unique i and ¥ To, vr (Lo, = stabilizer of Q, in). Hence wt ? +b, Re(¥z) +e, a im(¥2) ) iN eTQ yh and so : y alz|?+bx+c, f ¥,.@@F@dz = » J o( ET reer, mw aa Fo\n Taking F(z) = {(2s)E(,s) with 10 if we drop the factor +. Then the substitution —ibn+cm 24 +-— introduced in [11], p. 127, maps H to Hand gives —mz+an—3bm ie aaa =ain-m" [0 (52%), Therefore we have i> * dz. ?_ pia mare) vee (30) { (2s) f £0 @E(,s)dz = san [ o( i r\w for 1< Re(s) <1 + , with £(s,D) defined as in (20). Since {(2s) and 287D. ZAGIER {(s,D) have meromorphic continuations to all s and both integrals in (30) converge for 0< Re(s)<1 +, we deduce that the identity is valid in this larger range; the divisibility of {(s,D) by {(s) now implies the orthogonality of the functions (19) with £,@(z). If D is a square, we would have to treat the terms with Q,(n, —m) =0 in the above sum separately (as in [11], pp. 127-128). We prefer a different method, which in fact works for all D. By the Rankin-Selberg method, we know that | %,@ (z) E(z,s) dz equals the Mellin transform of the constant term of ,¢, and writing (2) as 2 © ake+ +n] —D/4a > —) a=) bimodiay pe me aD > + (= x) we see that this constant term is given by b? = D(mod4a) 2. rs a 2 —D/4a carn S nim fol) 0 f D¥¢m’, +4 y fax ¥ D=m'>0, 43 og) Ff D=0, where n_,(a) is defined by (28). The Mellin transform of the first term is (S22): Ff oR os and since > n,(a)a~* = £(s,D)/Z (2s) ([11], Prop. 3, i), p. 130) we recover eq. (30) if D is not a square. The second term vanishes if D = m? + 0 because of the assumption (29), so eq. (30) remains valid in this case. If D = 0, then, using equation (29) and the Poisson summa- 288EISENSTEIN SERIES tion formula, we see that the second term in the formula for C(.¥,;y) equals DoD = a9 [ eenax=7 > Gon, Sy = f (Ke ax where is the Fourier transform of ®. The Mellin transform of this is (s) times the Mellin transform of @, so we obtain 2 BD (2s) j £,® (E(,s)dz = £(s,0) J ® lel Hw + £(8)£02s) fe (y"' dy a for 10, where P,, (t) and F( steed, —t?) denote Legendre and hyper- geometric functions, oe since both of these functions are invariant under s > | —s, we see that (30) is compatible with (and indeed gives another proof of) the functional equation of £(s,D) for D # 0([11], 289D, ZAGIER Prop. 3, ii), p. 130). We can also make the functional equation apparent in the case D = 0 by substituting — 3 for z in the first integral on the right- hand side of (31) and using the identity fe 1 J y” ‘cos 2@Xy dy =—9 ] 2 in the second; this gives 2 re 7 [X|* @ (a aw) pp Pp? (where e(a, b,c) equals | if'a,b,c are integral, 0 otherwise). To prove (36) we must show that D abc ab nfa,b,c) = 1 + (—)e(—.—,-) + p els. P Pe oe (a,b,ceZ, b? — 4ac = Dp?). 291D. ZAGIER For p odd, this follows from the following table, in which yj (m) denotes the exact power of p dividing an integer m. ab %@) %0) 4 | © Pp 0 30 30 0 1 0 0 D 1 >i al 0 1+€) 1 0 i oo ol 1 0 0 0 22 ar! + 1 1 0 >2 2 1 0 1 0 >2. 2 22 1 P 1 1 The proof for p= 2 is similar but there are more cases to be considered. iii) We will show that each of the properties in question is implied by the orthogonality of F with ¥,® for special choices of D and®. For (21) we choose ®(X) = 8(X7+D), where 8 is the Dirac delta-function. From the identity setey Lp jaz’ + bz +e[? 7 ( 1 a we see that the support of £,@ is the set of points in H satisfying some quadratic equation of discriminant D, and an easy calculation shows that mp) 7 38) j F@ £,0(@dz2= 2 rw i= F@,) for any continuous F: !'\H-> €. (Of course, 8 (X? + D) is not a function, and equation (38) must be interpreted in the sense that it holds in the limit n->e if we choose ®(X) = 8,(X? + D) where {8,} is a sequence of smooth, even functions with integral 1 and support tending to {0 }.) Hence any Fe (Im ¢,)* satisfies (21). 292EISENSTEIN SERIES The case D>0, D not a square, is similar; here we choose ®(X) = 5(X). so that ¥, (2) is supported on the semicircles a|z|* + bx += 0 (a,b,ce Z, b” — ac = D), and find no) (39) j Fo) 2,0() d= » f F@)|dq zi, rin AD “ar & : where the equation is to be interpreted in the same way as (38). Thus Fe (Im%,)" implies (23). It remains to prove that any Fe & satisfies equation (24). We follow the proof of the divisibility of } |a,|?/n***”' by £(s) given in [11]. Equa- mm tions (37) and (47) of that paper give the identity dn) (40) > ey he amt (k=1) > Frm Ou (zeH) 2 for all integers m> 0, where r = dim S, (SL, (Z)), (2) = 2, a,(m)e""™ (i=1,...,r) are the normalized Hecke eigenforms of weight k, (f, 8) = J y'i@P dz, and ©,,.00=(X-iy* +(X +i. Thus any n\n function in ¢ is orthogonat to the sum on the left-hand side of (40) and therefore, since the Fourier coefficients a, (m) are linearly independent, to each of the functions y ‘Ve (2))?. Using the computations of [12] and an extension of the Rankin-Selberg method [13], it seems to be possible to prove the orthogonality of Fe & with |f(z)|? also for Maass eigenforms (= non-holomorphic cusp forms which are eigenvalues of the Laplace and Hecke operators) of weight 0. §4. Let G=PSL,(R) and K=SO(2)/{+1} its maximal compact subaroup, and identity the symmetric space G/K with H by gK = ¢ )K<>g-i= +P tn this section we will construct a representation ¥ of G in the space of functions on P'\ G whose space of K-fixed vectors r¥ ise. 293D. ZAGIER Let % =( 6). "Dla,d.ceR} be the 3-dimensional vector space of symmetric real 2 x 2 matrices and 4,C&, the lattice consisting of matrices with a,b,ce Z. The group G acts on %, by g°M =g'Mg(ge G,Me 2), and 4, is stable under the action of the subgroup I’. For M ¢ 2, and ge G, the expression tr(g'Mg) depends only on the right coset gK (since k' =k" for ke K), i.e. only on g-ieH. An easy calculation shows that : +bx+e : : (41) tr(g' Me ae (M=6,,°2)e4,.2=:ieH). 2 alz|’+bx+c , This explains where the strange expression ale OR +e i the defini y tion of &, comes from and also why this expression is invariant under the simultaneous operation of I’ on the upper half-plane (gk -> ¥gK) and on binary quadratic forms (M >(¥')'M¥~™). Using (41), we can rewrite the definition of ¥, as » ©(1(e'Mp). Mer deM= “Dit To pass from functions on H to functions on G, we replace the special function M-> ® (tr(M)) by an arbitrary function ® on the 2-dimensional submanifold £,@(8k) #(D)={(,, 2) «£,|b? — 4ac = D} a Deh b/2 of &,. Thus weextend , to an operator (still denoted. £,) from the space of nice functions on 4 (D) to the space of functions on I*\G by setting (42) £,0@= >) ®(g'Mzg) (g«G), ee where 2 (D)= Z_ (D) A #;. Here “nice” means that ® satisfies the obvious extensions of (27) and (29), i.e. it must be of sufficiently rapid decay in %, (D) and, if D is a square, must be smooth and have zero 10 ‘8 (geGe=+1) integral along each of the lines /,, = aCe 7 294EISENSTEIN SERIES on the ruled surface 2, (D). Itis clear that %,(g) is left I'-invariant, since %; (D) is stable under T and the sum (42) is absolutely convergent. Also, the image of &, is stable under the representation = of G given by right translation, since m(g)°¥,=L,-m(8) (geG), where ’(g)(M) = © (g'Mg). Hence the space (43) ¥ = N (Img,)* Dez of functions F: TG C satisfying an appropriate growth condition and such that (44) J $,(@)F(e)dg=0 (dg=Haarmeasure) r\c for all De Z and all nice functions © on %z (D), is also stable under G."” Also, it is clear that ¥* coincides with the space # defined in §3. In parti- cular, ¥ contains the vectors v, : g > E*(g:i, p) (p a non-trivial, zero of £(8)) and more generally v,,, . > F,._ B'i)(F,_ a8 in (19)). On the other hand, because the function z + E(z,s) is an eigenfunction of the Laplace operator on H, the representation theory of G tells us that (at least for s¢ Z) the smallest G-invariant space of functions on P'\G containing the function g > E(g-i,s) is an irreducible representation isomorphic to the principal series representation & . (Recall that % is the representation of G by right translations on the set of functions f: G > C satisfying £(¢ .-1)8)= |a|” f@) and f],,¢ L7(K)). Thus ¥ contains the principal series representation , for every non-trivial zero p of the Riemann zeta-function. (1) Since the various manifolds #,(D)C are disjoint, we can also define ¥”as (Im )*, where # is the operator from nice functions ‘on £,to functions on I°\G defined by £O(g)= > ig’ Mg), We may also identify #,with the Lie algebra j, ={(." 7,)| a, b, ceR} of G by M > M(; 4); then the operation M + g'Mg of G on, becomes the adjoint representation X +1 Xg of G oni, ,and7‘is the set of functions on P\G orthogonal to all functions of the form g » S ¥(Ad(g)X). where W isa nice function on, 295D. ZAGIER On the other hand, % is unitarizable if and only ifs(1—s)> 0, ic. s¢ (0,1) or Re(s) = 4. Thus the existence of a unitary structure on ¥ would imply the Riemann hypothesis. However, the following argument suggests that it may be unlikely that such a unitary structure can be defined in a natural way. If p is a zero of &(s) of order n>1, then the functions F,,, (m=0,...,n—1) belong to +“, and these functions are not eigenfunctions of A, though the space they generate is stable under A (for example, differentiating (18) with respect to s we find AF,, = p(p—I)F,, + (2p—1)F,.9). Therefore ¥” contains a G-invariant subspace ¥,,, corresponding to the eigenvalue p(1—p) which is reducible but is not a direct sum of irreducible repre- sentations (we have dim ¥," =nand¥,, 3%, 4D...9%, I%o={0} with %,,/%m-; ~%,), and such a representation cannot have a unitary structure. Thus the unitarizability of ~ would imply not only the Riemann hypothesis, but also the simplicity of the zeroes of {(s). Since an analogue of ¥ can be defined for any number field or function field (cf. $ 5), and since there are examples of such fields whose zeta-functions are known to have multiple zeroes, there cannot be any generally appli- cable way of putting a unitary structure on ¥. Of course, this does not preclude the possibility that our particular ¥ (for the fieldQ) has a unitary structure defined in some special way, and indeed, if the zeroes of {(s) are simple and lie on the critical line and if (as seems likely) & is spanned by the v, , then is in fact unitarizable, indeed in infinitely many ways, since we are essentially free to choose the norm of v,. For various reasons, a natural choice seems to be ll v,||=|£*(2)|- Finally, we should mention that the construction of ¥ is closely related to the Weil representation. The functions ¥,@(g) are essentially the Fourier coefficients for a “lifting” operator from functions onT'\G to automorphic forms on the metaplectic group, in analogy with the con- struction of Shintani [9] in the holomorphic case; thus the space Ycan be interpreted as the kernel of the lifting. $5. The proof of the theorem in §3 shows that almost every statement of that theorem can be strengthened to a statement about the individual spaces 296EISENSTEIN SERIES é,=(ImZ,) (Dez) rather than just their intersection &. Thus in part iii) of the theorem, to prove that a function F satisfies (21) or (23) we needed only Fe ¢, for the value of D in question, and it was only for (24) that FeQ¢,, was needed. Similarly in part ii), equation (34) shows that each space #7, is stable under the Laplace operator. The same is not true for the Hecke operators, since T(p) maps Im(¥,) to Im(%,,,) + Im(#,)+ Im(,,,2) but the intersection of the spaces #, for all D with a common squarefree part és stable under the Hecke algebra. Finally—and most interesting— from equation (30) or (32) we see that €, (D+0) contains E*(2, p) whenever p is a zero of £(s,D) (resp-5 E%@z,9)|,_, for0 I*\H and under this projection T(Q)Z(A)\T(A) maps to ¥, . The adelic analogue of Propositions 1 and 2 is the fact that the integral of an Eisenstein series over T(Q)Z(A)\T(A) is a multiple of the zeta-function of E. To prove it, we must recall the definition of the Eisenstein series. Let @ be a Schwartz Bruhat function on A’; then the Eisenstein series E(g,®,s) is defined for ge G(A) and se € with sufficiently large real part by (45) EG,®,s)= > © [ézg]|detzg|¢ dz, LQ\ZA) eeQh{o} where | | denotes the idele norm and dz the Haar measure on Z. (This definition is the analogue of equation (2). The more usual definition of E(g,®,s), analogous to eq. (1), is E(g,®,s)= > £{(¥g,@,), 1ePORGQ) 298EISENSTEIN SERIES where P= {(¢ *)} and £@,,s)=|detg|*, f © [(0,a)g]|a| da, x which is easily seen to agree with (45); note that f(g,@,s) equals £, (2s) times an elementary function of s by Tate theory, and that f((° )g,@,s) = |&{* £@,@,5), so the analogue of f(g,@,s) in the upper half-plane is the function { (2s) Im(g-i)' .) Identifying Q?-{ 0} with E” and observing that the Q-idele norm of det t (te T(A)) equals the E-idele norm of t under the identification of T(A) with A,” we find E(tg,®,s)dt = f > © [gta] |dettg |g at TIQHAN\ TIA) TQ\TIAY eto} =latele f > orgie a EAE EX = |detg |g f ® [egllel. d°e. - e Since ¢-> @{eg] is a Schwartz-Bruhat function on A,, this is precisely the Tate integral for {, (s). (The computation just given is the basis for Harder’s computations of period integrals in this volume as well as for the generalization of the Selberg trace formula in (4].) In particular, it follows that the integral of F(g) = —, E@.®,5)|,_, as over T(Q)Z(A)\T(A)g vanishes if p is a zero of £, (s) of multiplicity >i+1. The natural adelic definition of ¥(E) is thus as the space of functions F: G(Q)Z(A)\G(Q) + € satisfying (46) f F(tg)dt =0 (VgeG(A)) THQIZA)\TIA) as well as some appropriate growth condition. The space ¥(E) then contains irreducible principal series representations corresponding to the zeroes of f(s). Condition (46) is similar to the condition F(ng)dn=0 (WV ge G(A)) NQ\NA) 299D. ZAGIER defining cusp forms (where N is the unipotent radical of a parabolic subgroup of G), so the space ¥“(E) can be thought of as an analogue of the space L3(G(Q)Z(A)\G(A)) of cusp forms. Like L2, it probably has a discrete spectrum. The difference is that, whereas L? has a unitary structure, the corresponding statement for” (E) would imply the Riemann hypothesis and the simplicity of the zeroes of £q(s). We call functions F satisfying (46) toroidal forms (in analogy with the French terminology of “formes paraboliques” for cusp forms) and the ¥(E) toroidal repre- sentations. The calculation given above is unchanged if we replace Q by any global field F and take E to be a quadratic extension of F. In the case where F is the functional field of a curve X over a finite field, there are only finitely many zeroes of {,, (8), their number being equal to the first Betti number of X. Then the K-finite part of our representation 4 = 9 Y(B) isa complex vector space of dimension at least (and hope- fully exactly) equal to this Betti number, and Barry Mazur pointed out that this space might have a natural interpretation as a complex cohomo- logy group H'(X;C). It is not yet clear whether this point of view is tenable. At any rate, however, from conversations with Harder and Deligne it appears that it will at least be possible to show that the dimension of the space in question is finite. BIBLIOGRAPHY {1] GELBART, S. and H. JACQUET, : A relation between automorphic repre- sentations of GL(2) and GL(3). Ann. Se. Ee. Norm. Sup. 11 (1978) 471-542 [2] HECKE,E.: Uber die Kroneckersche Grenzformel fiir reelle quadratische Kérper und die Klassenzahl relativ-abelscher_ Korper, Verhandl. d. Naturforschenden Gesell. i. Basel 28, 363-372 (1917). Mathematische Werke, pp. 198-207. Vandenhoeck & Ruprecht, Géttingen 1970. [3] JACQUET, H. and J. SHALIKA, : A non-vanishing theorem for zeta functions OfGL,. Invent. Math, 38 (1976) 1-16. 300EISENSTEIN SERIES [4] JACQUET, H. and D. ZAGIER, : Eisenstein series and and the Selberg trace formula It. In preparation. [5] RANKIN, R.: Contributions to the theory of Ramanujan’s function t(n) and similar arithmetical functions. I. Proc. Cam. Phil. Soc. 35 (1939) 351-372. [6] SELBERG, A. : Bemerkungen iiber eine Dirichletsche Reihe, die mit der Theorie der Modulformen nahe verbunden ist. Arch. Math. Naturvid. 43 (1940) 47-50. (7] SHIMURA, G.: On the holomorphy of certain Dirichlet series. Proc. Lond. Math, Soc. 31 (1975), 79-98. {8] SHINTANI, T.: On zeta-functions associated with the vector space of quadratic forms J. Fac. Science Univ. Tokyo 22 (1975), 25-65. [9] SHINTANI, T-: On construction of holomorphic cusp forms of half integral weight. Nagoya Math, 3. 58 (1975), 83-126. [10] ZAGIER, D. : A Kronecker limit formula for real quadratic fields. Math. Ann, 213 (1975), 153-184, [11] ZAGIER, D. : Modular forms whose Fourier coefficients involve zeta-functions of quadratic fields. In Modular Functions of One Variable VI, Lecture Notes in Mathematics No. 627, Springer, Berlin-Heidelberg-New York 1977, pp. 107-169. [12] ZAGIER, D.: Eisenstein series and the Selberg trace formula I. This volume, pp. 303-355. (131 ZAGIER, D.: The Rankin-Selberg method for automorphic functions which are not of rapid decay. In preparation. 301