In this chapter, we will construct, for every n ≥ 3, n-dimensional objects with Ggeom ⊂ Garith ⊂ SL(n), but only when n is prime will we be able to prove that Ggeom = Garith = SL(n).
Theorem 21.1. Let k be a finite field of characteristic p, Ã a nontrivial additive character of k. Let f (x) = ∑ ai=–b Aixi ∈ k [x, 1/x] be a Laurent polynomial of “bidegree” (a, b), with a, b both ≥ 1 and both prime to p. Assume further that f (x) is Artin-Schreier reduced. Thus Aa A–b ≠ 0, and Ai ≠ 0 implies that i is prime to p. We have the following results.
(1) The object N := LÃ(f(x))(1/2)[1] ∈ Parith is pure of weight zero, geometrically irreducible, and of “dimension” a + b.
(2) If (–1)aa Aa = (–1)b+1bA–b, then Ggeom,N ⊂ SL(a + b). In general, the Tannakian determinant “det” (N) is geometrically isomorphic to δα, for α = (–1)b+1bA–b/(–1)aaAa. And setting
β := det(Frobk,|!(N)),
we have the arithmetic determinant formula
“det”(N) = βdeg ⊗ δα.
(3) If a = b, then N is not Lie-self-dual, i.e., G0geom, N lies in no orthogonal group SO(a + b), and, if a + b is even, it lies in no symplectic group Sp(a + b).
(4) Suppose in addition that gcd{i|Ai ≠ 0} = 1. Then N is geometrically Lie-irreducible, i.e., G0geom, N is an irreducible subgroup of SL(a + b) (in the given a + b-dimensional representation).
Proof. The sheaf LÃ(f(x)) is lisse of rank one (and a fortiori geometrically irreducible) and pure of weight zero on Gm, with Euler characteristic –Swan0 – Swan∞ = –a – b, so (1) is obvious. Let us admit the truth of (2) for the moment. To show that N is not Lie-self-dual when a ≠ b, it suffices to show that for any integer d ≥ 1 prime to p, the d’th power direct image [d]?N is not geometrically isomorphic to its Tannakian dual. But this object has different Swan conductors at 0 and ∞, namely a and b respectively, while its Tannakian dual [x 1/x] ?D([d]?N) has Swan conductors b and a respectively at 0 and ∞. If in addition gcd{i|Ai ≠ 0} = 1, the geometric Lie-irreducibility of N follows from Corollary 8.3. Indeed, this gcd = 1 hypothesis and the fact that f is Artin-Schreier reduced together insure that for any a ≠ 1 in k×, N is not geometrically isomorphic to [x ax] ?N, cf. the proof of Theorem 14.2.
We now turn to the calculation of the Tannakian determinant “det”(N). We will compute, for every finite extension field E/k, and every character χ of E×,
det(FrobE,χ|!(N)) = det(FrobE|H0(A1 ⊗k k, j0!(N ⊗ Lχ))).
Because N is totally wild at both 0 and ∞, this is
= det(FrobE|H0c(Gm ⊗k k, N ⊗ Lχ))
= det(FrobE|H1c(Gm ⊗k k, LÃ(f(x)) ⊗Lχ)(1/2))
= √#E–a–b det(FrobE|H1c(Gm ⊗k k, LÃ(f(x)) ⊗Lχ)).
We follow the method of Hasse-Davenport [D-H]. The L-function on Gm/E with coefficients in LÃ(f(x)) ⊗Lχ has the additive expression
where the inner sum is over the space of effective divisors of degree d on Gm/E. Concretely, the effective divisors D of given degree d are the monic polynomials of degree d in E[x] of the form fD := xd – s1xd–1 + …+ (–1)dsd, with sd invertible. The term χ(D) for this divisor is χ(sd). The term Ã(D) for this divisor is Ã(∑ai=–b AiNi), for Ni the sum of the i’th powers of the roots of the corresponding polynomial fD. [N.B. Because fD has an invertible constant term, all its roots are invertible, so Ni makes sense for negative i as well.] Comparing the additive expression for the L-function with its cohomological expression
L(Gm/E, T, LÃ(f(x))⊗Lχ) = det(1–TFrobE|H1c(Gm⊗kk, LÃ(f(x))⊗Lχ)),
which shows the L function to be a polynomial in T of degree a + b, and equating coefficients of Ta+b, we get the identity
With the convention that for i > 0, s–i is the i’th elementary symmetric function of the inverses of the roots, we have
s–i = sa+ b–i ∕sa+ b.
And by the relation between Newton symmetric functions and elementary symmetric functions, we have
Na = (–1)a+ 1asa + pa(s1, …, sa–1),
where pa(s1, …, sa–1) is isobaric of weight a and does not involve sa. Similarly,
N–b = (–1)b + 1bs–b + pb(s–1, …, s–(b–1)),
where pb(s–1, …, s–(b–1))is isobaric of weight b and does not involve s–b. The terms Ni with 0 ≤ i < a are polynomials in the variables s1, …, si, and the terms N–j with 0 < j < b are polynomials in s–1, …, s–j. Thus we get an expression for
det(–FrobE|H1c(Gm ⊗k k, LÃ(f(x)) ⊗Lχ)
of the form
with
R := P (s1, …, sa–1)+ Q(s–1, …, s–(b–1)).
Making use of the relations s–i = sa+b–i/sa+ b, we see that the term Q(s–1, …, s–(b–1)) is itelf a polynomial in sa+ b–1, sa+ b–2, …., sa+ 1 and 1/sa+ b. Thus R does not involve the variable sa. Using the relation s–b = sa/sa+ b, we see that the only occurrence of the variable sa is in the two terms
Aa(–1)a+ 1asa + A–b(–1)b+ 1bsa/sa+ b.
So we can factor out the sum over sa, and get an expression of det(–FrobE) as the product of
times
This last sum vanishes unless
sa+ b = A–b(–1)b+ 1b/Aa(–1)aa,
in which case this sum is equal to #E. Defining α = A–b(–1)b+ 1b/Aa(–1)aa, we get an expression for det(–FrobE) as
χ(α) × (an expression independent ofχ).
Putting this all together, we get
det(–FrobE,χ|!(N)) = χ(α)S(E, N)
with S(E, N) the exponential sum in a+ b–2 variables s1, …, sa–1, sa+ 1, …, sa+ b–1 given by
Taking χ = , we see that
S(E, N) = det(–FrobE, |!(N)).
Thus we find
det(FrobE,χ|!(N)) = χ(α)det(FrobE,|!(N)),
so defining
β := det(Frobk, |!(N)),
we get the arithmetic determinant formula
“det”(N) = βdeg ⊗ δα.
The following lemma gives us some control over the quantity β := det(Frobk, |!(N)).
Lemma 21.2. Let k be a finite field of characteristic p, Ã a nontrivial additive character of k. Let f(x) = ∑ai=–bAixi ∈ k[x, 1/x] be a Laurent polynomial of “bidegree” (a, b), with a, b both ≥ 1 and both prime to p. The ratio
det(Frobk|H1c(Gm/k, LÃ(f(x))))2/(#k)a+ b
is a root of unity of order dividing 2p if p is odd, and the ratio is ±1 if p = 2. If a + b is even, the same is true for the ratio
det(Frobk|H1c(Gm/k, LÃ(f(x))))/(#k)(a+ b)/2.
If a + b is odd, the same is true of the ratio
det(Frobk|H1c(Gm/k, LÃ(f(x))))/(G(Ã, χ2)(#k)(a+ b –1)/2),
with G(Ã, χ2) the quadratic Gauss sum.
Proof. The exponential sum expression for det(Frobk|H1c(Gm/k, LÃ(f(x)))) derived in the proof of the previous theorem shows that this determinant lies in the ring Z[ζp] of integers in the cyclotomic field Q(ζp). Hence the ratios in question all lie in Q(ζp) as well. We are asserting that the ratios in question are each roots of unity in Q(ζp). For this, it suffices to see that each ratio has absolute value one at every place, finite or infinite, of Q(ζp). At archimedean places, this is the fact that H1c(Gm/k, LÃ(f(x))) is pure of weight one, and of dimension a + b. At ℓ-adic places, ℓ≠ p, both the determinant, calculated now via ℓ-adic cohomology, and the quantities #k and G(Ã, χ2), are ℓ-adic units. At the unique place lying over p, both this determinant and its complex conjugate have the same p-adic absolute value as each other, as the two quantities are Galois conjugate in Qp(ζp). But the two groups H1c(Gm/k, LÃ(f(x))) and H1c(Gm/k, LÃ(f(x))) are dually paired to Qℓ(–1), so the product of the determinant with its complex conjugate is (#k)a + b. Therefore the square of the determinant, divided by (#k)a + b, is a p-adic unit as well. The square of either of the second two ratios is ±1 times the first ratio, so it too is a p-adic unit.
Theorem 21.3. We have the following results concerning the object N of Theorem 21.1.
(1) The quantity
β := det(Frobk, |!(N))
is a root of unity of order dividing 2p, if either a + b is even or if p is odd and –1 is a square in k. If p and a + b are both odd and –1 is a nonsquare in k, then β is a root of unity of order dividing 4p.
(2) If (–1)aaAa = (–1)b + 1bA–b, then for any a + b’th root ° of 1/β, the object °deg ⊗ N has Ggeom ⊂ Garith ⊂ SL(a + b).
(3) If in addition a ≠ b and gcd{i|Ai ≠ 0} = 1 and a+ b is a prime number ≥ 3, then the object °deg ⊗ N has Ggeom = Garith = SL(a + b).
(4) If the object °deg ⊗ N has Ggeom = Garith = SL(a + b) (e.g., if a+ b is prime), then for any r ≥ 2 distinct characters χ1, …, χr of k×, the object ⊕ri=1°deg ⊗ N ⊗ Lχi has
Proof. Assertion (1) is immediate from the corollary, because G(Ã, χ2)2 = χ2(–1)#k. Assertion (2) is immediate from assertion (2) of Theorem 21.1. To prove (3), we argue as follows. By parts (3) and (4) of Theorem 21.1, G0geom is an irreducible subgroup of SL(a + b) which lies in no orthogonal subgroup SO(a + b). By Gabber’s theorem on prime-dimensional representations [Ka-ESDE, 1.6], the only connected irreducible subgroups of SL(a+ b), for a+ b prime, are SL(a+ b), SO(a+ b), the image of SL(2) in Syma+ b –1(std2), and, if a+ b = 7, the group G2 in its seven-dimensional representation. All of these except for SL(a + b) itself lie in an orthogonal subgroup SO(a + b).
Assertion (4) results from the Goursat-Kolchin-Ribet Theorem 13.5. We have a priori inclusions for this object
so it suffices to show that Ggeom =∏ri=1 SL(a + b). We must show that for i ≠ j, and for any one-dimensional object L ∈ Pgeom, there is no geometric isomorphism between N ⊗ Lχi and either N ⊗ Lχj ?mid L or its Tannakian dual. To see this, we argue as follows. The fact that N ⊗ Lχi and N ⊗ Lχj and its Tannakian dual all have trivial determinants forces L⊗(a+ b) to be geometrically trivial, which in turn implies that L is punctual, so some δ² with² ∈ . Thus we must show that LÃ(f(x)) ⊗Lχi is not geometrically isomorphic to either LÃ(f(²x)) ⊗Lχj or to LÃ(–f(²/x)) ⊗ Lχj, for any ². The second is impossible, because of the asymmetry of the Swan conductors. For the first, their ratio is LÃ(f(²x)–f(x)) ⊗Lχj/χi. If ² ≠ 1, this ratio is wildly ramified at either 0 or ∞ or both, thanks to the hypothesis that f is Artin-Schreier reduced and has gcd{i|Ai ≠ 0} = 1. So their ratio is not geometrically trivial. If ² = 1, their ratio is Lχj/χi, which is nontrivially ramified at both 0 and ∞, so again is not geometrically trivial.
To end this chapter, we give another family of examples where, for every n ≥ 3 we have Ggeom ⊂ Garith ⊂ SL(n), but where, once again, only when n is prime can we prove that Ggeom = Garith = SL(n). We work over a finite field k, with à a nontrivial additive character of k. For each integer m ≥ 1, we denote by Klm the Kloosterman sheaf of rank m Klm(Ã; , , …, ). Recall that its complex conjugate Klm is given by Klm = [x (–1)mx] ?Klm.
Theorem 21.4. Fix strictly positive integers a ≠ b. Denote by F the lisse sheaf on Gm given by
F := Kla ⊗ [x 1/x] ? Klb = Kla ⊗ [x (–1)b/x] ?Klb.
Denote by N ∈ Parith the object F((a + b – 1)/2)[1]. Then we have the following results.
(1) The object N ∈ Parith is pure of weight zero, geometrically irreducible, and of “dimension” a + b.
(2) For a multiplicative character χ of k×, consider the Kloosterman sheaf Kla+ b(Ã; χ,.., χ, , …, ) of rank a + b with characters χ repeated a times, and repeated b times. Then we have the identity
det(1 – TFrobk,χ|!(N))
= det(1 – TFrobk,(–1)b |Kla+ b(Ã; χ,.., χ, , …, )((a + b – 1)/2)).
(3) We have Ggeom,N ⊂ Garith,N ⊂ SL(a + b).
(4) The object N ∈ Parith is geometrically Lie-irreducible, i.e., G0geom, N is an irreducible subgroup of SL(a + b).
(5) The object N ∈ Parith is not geometrically Lie-self-dual, i.e., G0geom, N lies in no orthogonal group SO(a + b) and, if a + b is even, it lies in no symplectic group Sp(a + b).
(6) If a + b is a prime number, then Ggeom = Garith = SL(a + b).
(7) Suppose that Ggeom = Garith = SL(a + b), and that gcd(a, b) = 1 (e.g., both hold if a + b is prime). For any r ≥ 2 distinct characters χ1, …, χr of k×, put Ni := N ⊗ Lχi. Then the object ⊕ri=1Ni has
Proof. (1) One knows that Klm is pure of weight m – 1, so F is pure of weight a + b – 2, and hence N is pure of weight zero. To show that N, or equivalently F, is geometrically irreducible, we repeat the argument given in Remark 18.14. By purity, F is geometrically semisimple, so it suffices to observe that its local monodromy at 0 (or at ∞, either one will do) is indecomposable. Its local monodromy at 0 is Unip(a) ⊗ Wildb,1, where Wildb,1 is totally wild of rank b and Swan conductor one, and the indecomposability follows from [Ka-RLS, 3.1.7]. Its local monodromy at ∞ is Wilda,1 ⊗ Unip(b). Thus the dimension of N is Swan0 + Swan∞ = a + b.
To prove (2), it suffices to prove, over all finite extensions of k, the identity of traces
Trace(Frobk,χ|!(N)) = Trace(Frobk,(–1)b |Kla+ b(Ã; χ,.., χ, , …, )((a+ b–1)/2)).
Because F is totally wild at both 0 and ∞, there are no bad characters, so we have
Trace(Frobk,χ|!(N)) := Trace(Frobk|H0c(Gm ⊗ k, N ⊗ Lχ))
= Trace(Frobk|H1c(Gm ⊗ k, F ⊗Lχ)((a + b – 1)/2)).
So what we must show is the identity
Trace(Frobk|H1c(Gm ⊗ k, F ⊗Lχ))
= Trace(Frobk,(–1)b |Kla+ b(Ã; χ,.., χ, , …, )).
On the left-hand side, the H1c is the only nonvanishing cohomology group, so by the Lefschetz trace formula [Gr-Rat] the left-hand side is
This in turn is
which is precisely the trace of Frobk,(–1)b on the convolution of Kla(Ã; χ, …, χ) with Klb, and that convolution is precisely the Kloosterman sheaf |Kla+ b(Ã; χ,.., χ, , …, ), cf. [Ka-GKM, 5.5].
To prove (3), it suffices to prove, thanks to (2), that for every χ we have the identity
det(Frobk,(–1)b |Kla + b(Ã; χ,.., χ, , …, )) = (#k)(a+ b)(a+ b–1)/2.
This identity is a special case of the arithmetic determinant formula for Kloosterman sheaves given in [Ka-GKM, 7.4.1.3 and 7.4.1.4]. [There the formula comes out as(#k)(a+ b)(a+ b–1)/2 times the additional factor χa((–1)a+ b–1)χa((–1)b), but this factor is 1.]
To prove (4), we must show that N is not geometrically isomorphic to any nontrivial multiplicative translate of itself. But already this is true for its local monodromy at 0, Unip(a) ⊗ Wildb,1, indeed it is true for Wildb,1 itself, cf. [Ka-GKM, 4.1.6 (3)].
We prove (5) exactly as in Theorem 21.1. We simply observe that N has different Swan conductors a and b at 0 and ∞ respectively, as do all its direct images by Kummer maps.
We then get (6) by Gabber’s classification of prime-dimensional representations [Ka-ESDE, 1.6], exactly as in the proof of Theorem 21.3.
To prove (7), we use Goursat-Kolchin-Ribet. We have a priori inclusions for the object ⊕ri=1Ni,
so it suffices to show that Ggeom,⊕ri=1Ni = ∏ri=1 SL(a + b). We must show that for i ≠ j, and for any one-dimensional object L ∈ Pgeom, there is no geometric isomorphism between N ⊗ Lχi and either (N ⊗ Lχj) ?mid L or its Tannakian dual. To see this, we argue as follows. The fact that N ⊗ Lχi and N ⊗ Lχj and its Tannakian dual all have trivial determinants forces the Tannakian tensor power L⊗(a+ b) to be geometrically trivial, which in turn implies that L is punctual, so some δ² with² ∈ . Thus we must show that for i ≠ j, N ⊗ Lχi is not geomerically isomorphic to any multiplicative translate of either N ⊗ Lχj or of its Tannakian dual. The second is impossible, because of the asymmetry of Swan conductors. To rule out the first, we argue as follows. If N ⊗ Lχi is geometrically isomorphic to some multiplicative translate [x ²x] ?(N ⊗ Lχj), then comparing local monodromies at 0 and ∞ respectively we find geometric isomorphisms
Wildb,1 ⊗Lχi ≅ [x ²x] ?(Wildb,1 ⊗Lχj)
and
Wilda,1 ⊗Lχi ≅ [x x] ?(Wilda,1 ⊗Lχj).
Suppose first that a = 1. Then Wilda,1 ≅ LÃ(x) as I(∞) representation, so the second isomorphism asserts that LÃ(x) ⊗Lχi ≅ LÃ(²x) ⊗Lχj as I(∞) representation, i.e., that LÃ((1–²)x) ⊗Lχi/χj is the trivial character of I(∞). This is nonsense: if² ≠ 1, this character has Swan conductor one, and if ² = 1, it is the nontrivial character Lχi/χj. Similarly, we deal with the case b = 1.
If both a, b ≥ 2, then both Wilda,1 and Wildb,1 have trivial determinants. Equating determinants in the two isomorphisms, we find equalities χbi = χbj and χai = χaj. As gcd(a, b) = 1, we infer that χi = χj, contradiction.
3.14.245.167