Effective Convergence Bounds for Frobenius Structures on Connections
KIRANS. KEDLAYA(*) - JANTUITMAN(**)
ABSTRACT- Consider a meromorphic connection onP1over ap-adic field. Inmany cases, such as those arising from Picard-Fuchs equations or Gauss-Manin con- nections, this connection admits a Frobenius structure defined over a suitable rigid analytic subspace. We give an effective convergence bound for this Fro- benius structure by studying the effect of changing the Frobenius lift. We also give an example indicating that our bound is optimal.
1. Introduction
In recent years, much work has gone into usingp-adic cohomology as an effective tool for numerical computation of zeta functions (and some re- lated quantities) of algebraic varieties over finite fields. One important technique in this field is thedeformation methodof Lauder, inwhich one computes the zeta function of a variety by fitting it into a one-parameter family of varieties, constructing the associated Picard-Fuchs equation or Gauss-Manin connection, then exploiting the existence of a Frobenius structure on this differential equation to reduce the problem to another member of the family (the initial condition, so to speak). Another important technique is thefibration method, also introduced by Lauder, in which one uses similar techniques to compute the zeta function of the total space of a one-parameter family, again starting from a single fiber. See [4, 5] for further discussion.
(*) Indirizzo dell'A.: Department of Mathematics, University of California, SanDiego, 9500 GilmanDrive #0112, La Jolla, CA 92093-0112, USA.
E-mail: kedlaya@ucsd.edu
(**) Indirizzo dell'A.: Mathematical Institute, University of Oxford, 24-29 St.
Giles', Oxford OX1 3LB. UK.
E-mail: tuitman@maths.ox.ac.uk
To execute the deformationand fibrationmethods inpractice, it is necessary to have not just the existence of a Frobenius structure, but explicit bounds on its convergence within a given residue disc. Concretely, these bounds are needed to enable the reconstruction of a rational function from a power series expansion, by bounding the degrees of its zero and pole divisors. One can often obtain crude bounds by direct calculations, but it is essential to have more accurate bounds in order to limit the required intermediate precision needed to achieve a final result to a given level of accuracy.
A technique for obtaining accurate bounds has been suggested by the first author in the preprint [3], under the assumptions (satisfied in many cases in practice) that the differential equation has at most one singularity in any residue disc, and that the exponents of the local monodromy at such a singularity are p-adically integral. The idea is to exploit the parallel transport of Frobenius structures between two choices of a Frobenius lift, to reduce the questionof convergence withina givenresidue disc to the same question with the Frobenius lift centered around the singularity in the disc, a questionwhich canbe solved rather easily.
However, the bound given in [3, Theorem 6.5.10] is not best possible. A stronger bound was claimed in the original (2008) manuscript of [3], but the second author discovered that the proof was incomplete, as it relied on some unjustified assertions about the convergence of solutions of p-adic differential equations. In this paper, we give a corrected version of the original argument, thus giving a stronger version of [3, Theorem 6.5.10].
We also provide a numerical example which indicates that the resulting bound is sharp.
2. The theorem
We first introduce some notation and terminology.
Letpdenote a prime,na positive integer, andFqthe finite field with qpn elements. We write Qq for the unique unramified extension of degree n of the field of p-adic numbers Qp, an d Zq for the ring of in- tegers ofQq. LetUbe anopendense subscheme ofP1Qq with nonempty complementZ. Suppose thatE is a vector bundle onU equipped with a connectionr, an d lettdenote some coordinate on P1Qq.
We writesfor the standardp-th power Frobenius lift onP1Qqthat is, the (semilinear) map that lifts thep-th power Frobenius map onP1Fqand satisfies s(t)tp.
Let V denote the rigid analytic subspace ofP1Qq which is the comple- ment of the union of the open disks of radius 1 around the points ofZ, an d Oy(U) the ring of functions that converge on some strict neighbourhood of V. A Frobenius structure on E with respect to s is anisomorphism F :sE ! E of vector bundles with connection defined on some strict neighbourhood ofV.
We fix a basis [v1;. . .;vr] of sections of E on U, define matrices N2Mr(O(U)) andF2Mr(Oy(U)) such that
rvjXr
i1
Nijvidt;
FvjXr
i1
Fijvi;
and call these the matrices ofrandF. Note, however, thatrandFare not O(U)- andOy(U)-linear, respectively. Instead,rsatisfies the Leibniz rule, andF iss-semilinear as a map fromEto itself.
SinceFis a morphism of vector bundles with connection, it is horizontal with respect to r. This implies that the matrices N and F satisfy the differential equation
NFdF
dt ds(t)
dt Fs(N)ptp 1Fs(N):
1
Now letzbe a geometric point ofZ, and suppose that the entries of Nhave at most a simple pole atz. Whenris a Gauss-Manin connection, by the regularity theorem we canalways choose the basis [v1;. . .;vr] so that this is the case (where the choice will in general depend onz). The exponents of r at z with respect to [v1;. . .;vr] are defined as the eigenvalues of the matrix (t z)Njtz. When r is a Gauss-Manin con- nection or admits a Frobenius structure, these are known to be rational numbers.
Letj:jdenote the norm onOy(U) induced by the supremum norm onV, andvp(:) the corresponding discrete valuation, so thatj:j p vp(:). Extend both of these to Mr(Oy(U)) inthe usual way, i.e., as the maximum and minimum over the entries, respectively.
THEOREM2.1. Let z be an unramified geometric point of Z, and as- sume that Z does not contain anyother points with the same reduction modulo p. Suppose that[v1;. . .;vr]is a basis ofEwith respect to which the matrix N ofrhas at most a simple pole at z, and the exponentsfl1;. . .;lrg
of r at z are contained inQ\Zp. Assume thatE admits a Frobenius structureFwith respect tos, and letFbe the matrix ofFwith respect to the basis[v1;. . .;vr]. For i2N, put
f(i)maxf(vp(F)vp(F 1))dlogp(i)e;
(r 1)vp(N)(vp(F)vp(F 1))blogp(i)cg;
and define
c 0 if vp(N)0
minf0;if(i):i2Ng if vp(N)50:
(
For m2N, put
g(m)maxfi2Njivp(F)cf(i)5mg;
and define
a1 b pmin
i flig max
i fligc;
a2
0 if N does not have a pole at z;
0 if z2 f0;1g;
g(m) otherwise:
8>
<
>:
Then F is congruent modulo pm to a matrix of rational functions of order greater than or equal to (a1pa2) at z (that is, the entries of the difference between the two matrices all have p-adic valuation at least m).
The proof proceeds inseveral steps. We start with the following lemma.
LEMMA2.2. Let N P1
i 1Nitibe an rr matrix such that tN converges on the open unit disk and N 1is a nilpotent matrix. LetF P1
i 1Fitibe an rr matrix that converges on some open annulus of outer radius1.
Suppose that N;Fsatisfyequation(1).ThenFi0for all i50, so thatF converges on the whole open unit disk.
PROOF. See [2, Proposition17:5:1]. p
When the exponents of N at 0 are not necessarily zero, this can be generalized as follows.
LEMMA2.3. Let N P1
i 1Nitibe an rr matrix such that tN converges on the open unit disk and the eigenvalues l1;. . .;lr of N 1 are rational numbers with denominators coprime to p. LetF P1
i 1Fiti be an rr matrix that converges on some open annulus of outer radius1. Suppose that N;Fsatisfyequation(1).ThenFi0whenever
i5pmin
j fljg max
j fljg:
PROOF. First we may adjoint1=k forkcoprime top(if necessary), to reduce to the case wherel1;. . .;lr2Z. Inthat case, by applying so-called shearing transformations, one can find an invertiblerrmatrixW over Qq(t) such that the matrix
N0W 1NWW 1dW dt
still has (at most) a simple pole att0, but now with all exponents equal to 0. Moreover, one can ensure that tbW andt aW 1 do not have a pole at t0, foraminjfljgandbmaxjfljg. More details onthis canbe found in [3, Lemma 5.1.6]. If we change basis to the basis given by the columns of W, then
N!N0;
F!F0W 1Fs(W):
Now Lemma 2.2 canbe applied to the pairN0;F0, so thatF0i0 for alli50.
SinceFWF0s(W 1), this implies thatFi0 for alli5pa b. p Recall that we have chosenthe standardp-th power Frobenius lifts.
However, we could just as well have chosena different lift. The following lemma allows one to change from one Frobenius lift to another.
LEMMA 2.4. Let D denote the differential operator on E defined by rv Dvdt. Suppose thatEadmits a Frobenius structureF1:s1E ! E with respect to a Frobenius lifts1, and lets2be some other Frobenius lift.
ThenEalso admits a Frobenius structureF2:s2E ! Ewith respect tos2, defined by
F2(v)X1
i0
(s2(t) s1(t))iF1 Di i! (v)
:
PROOF. See [2, Proposition17:3:1]. p
Finally, we need a bound on the matrices of the differential opera- tors Di
i! that appear inLemma 2.4.
LEMMA2.5. LetD(i) be the matrix of the differential operatorDi i! with respect to the basis[v1;. . .;vr], that is,
Di
i! vkXr
j1
D(i)jkvj:
Then we have
vp(D(i))f(i);
where f(i)is defined as in Theorem2.1.
PROOF. Lethdenote a generic point of the disk of radius 1 aroundz.
One can verify that the Taylor series T( th;vj)X1
i0
( th)iDi(vj) i!
defines a horizontal section of r that meetsvj at the point h. Form the matrixMwhosej-th columnconsists of the expressionofT( th;vj) in terms of the basis [v1;. . .;vr], thenexpandMP1
i0Mi(t h)i. Sin cerdoes not have any singularities in the open disk of radius 1 aroundh, it follows from [2, Theorem 18.3.3] that
minfvp(M0);. . .;vp(Mi)g (vp(F)vp(F 1))dlogp(i)e;
and from [2, Remark 18.3.4] (withqp) that
minfvp(M0);. . .;vp(Mi)g (r 1)vp(N)(vp(F)vp(F 1))blogp(i)c:
SincejD(i)jattains its maximum ath, an d
Mi( 1)iD(i)(h)(terms coming fromD(j) withj5i);
we deduce the bound by induction oni. p
Now we finally get to the proof of Theorem 2.1.
PROOF OFTHEOREM2.1. We first note that in casez0 orz 1, the claim is clear from Lemma 2.3.
Suppose next that z is a point at which N has no pole, so that a1a20. If we use the Frobenius lifts0withs0(t z)(t z)p, thenby Lemma 2.3 again (applied after translatingzto the origin), the Frobenius matrix F0 with respect to s0 is holomorphic at z. By Lemma 2.4 (with s1s0;s2s),Fis also holomorphic atz, proving the claim in this case.
Finally, suppose thatNdoes have a pole atz. Inthis case, Lemma 2.3 implies thatF0has order at least a1 atz. We may againuse Lemma 2.4 (with s1 s0;s2s) to convert back to the original Frobenius lift; this gives us the identity
FX1
i0
piuiF0s0(D(i));
wherepu(t z)ps(z) tp(withvp(u)0), andD(i) againdenotes the matrix of the differential operatorDi
i! with respect to the basis [v1;. . .;vr].
In this identity, the summand at indexihas order at least a1 piatz, an d p-adic valuationat leastivp(F0)f(i) by Lemma 2.5. This will give the desired bound once we check that vp(F0)vp(F)c. To see this, apply Lemma 2.4 withs1ands2interchanged to obtain
F0X1
i0
pi
i!( u)iFs(i!D(i)):
Ifvp(N)50, we get the claim by invoking Lemma 2.5 again; ifvp(N)0, we instead note thatvp(pi=i!) an dvp(i!D(i)) are both nonnegative. p The following corollary is oftenuseful whenthe matrix N of r with respect to some basis does not have a simple pole atz.
COROLLARY 2.6. Suppose that [v1;. . .; vr] is a basis for E as in Theorem 2.1, and let [w1;. . .; wr] be another basis for E, such that vj Pr
i1Wijwj withW2Mr(Qq(t)). Then the matrixF0 ofF with respect to[w1;. . .; wr]is congruent modulopmvp(W)vp(W 1)to a matrix of rational functions of order greater than or equal to
(a1pa2(m))ordz(W)pordz(W 1) at z.
PROOF. The matrixF0satisfies
F0WFs(W) 1: p
REMARK2.7. Insome special cases Theorem 2.1 canstill be improved a little.
1. Ifs(z)zp(such azis called aTeichmuÈller lift), thens(t) s0(t) is divisible byt zinthe proof of Theorem 2.1. So whenwe apply Lemma 2.4, some cancellation occurs, and modulopm the matrixFhas order greater thanor equal to (a1(p 1)a2(m)) atz.
2. Suppose thatz60;1. If we denote the residue matrix (t z)Njtz ofNatzbyRz, and the identity matrix byI, thenthe leading term inthe Laurent series expansion of the matrixD(i) ofDi
i! at (t z) is givenby
(Rz (i 1)I). . .(Rz I)Rz (t z) i
i! :
Inmany cases there exists S2GLr(Qq) such that S 1RzS is diagonal.
Writingl1;. . .;lrfor the entries on the diagonal, the leading term can then be writtenas
S
l1(l1 1). . .(l1 (i 1))
i! 0
.. .
0 lr(lr 1). . .(lr (i 1))
i!
0 BB BB
@
1 CC CC
AS 1(t z) i:
However, sincel1;. . .;lr2Q\Zpby assumption, the matrix in the middle is easily seento have entries inQ\Zpas well. This means that the val- uation of the leading term in the the Laurent series expansion ofD(i)atzis bounded byvp(S)vp(S 1). Now whenwe apply Lemma 2.4, we see that if g(m)vp(F)cvp(S)vp(S 1)m, thenmodulopmthe matrixFhas order greater thanor equal to (a1p(a2(m) 1)) atz. This is related to the improvement upon [2, Theorem 18.2.1] given by the theorem of Dwork and Robba on which it is based [1].
3. An example: a family of elliptic curves
We consider the family given by the affine equation y2x31(t1)(x2x);
The closure of the zero locus of this equationinP2QP1Qdefines a family X=Uof elliptic curves overUP1Q f 2;2g.
The relative algebraic de Rham cohomology H1dR(X=U) is a vector bundle onU, and carries a natural Gauss-Manin connectionr. Moreover, H1dR(X=U) is of rank 2, and a basis is given by
dx y ;xdx
y
:
Let p be anodd prime number. The space H1dR(X=U) equipped with r coincides with the relative rigid cohomologyH1rig(Xp=Up) of the reduction Xp=UpofX=Umodulop, and therefore it admits a Frobenius structureF with respect to the standard liftsof thep-th power Frobenius. LetNandF be the matrices ofrandFwith respect to the above basis, respectively. We compute
N 1 t2 4
t 2
1 2
t 23
2 1
2 t 21
2 0
BB
@
1 CC A:
It is known thatvp(F)0, andvp(F 1) 1 inthis case, and clearly vp(N)0, so inTheorem 2.1 we have
g(m)maxfi2Nji blogp(i)c5mg:
3.1 ± z2
At z2 the exponents are f 1=4;1=4g. The residue matrix R2 is di- agonalizable by integral matrices forp65. So by remark 2.7, forp65 the bound from Theorem 2.1 for the order ofFmodulopmcanbe improved to
p1 4
p(g(m) 1)
, while forp5 it remains (15g(m)).
Experimentally, we find that for p3 the order is boun ded by 1 3(m 1), for p5 it is boun ded by 1 5(m 1), and forp7 it is bounded by 2 7(m 1), all for m up to 250 an d with equality for many m.
3.2 ± z 2
Atz 2 the exponents aref0;0g, so the bound from Theorem 2.1 for the order ofFmodulopmis givenby pg(m).
Experimentally, we find that for p3 the bound is sharp for m1;2;3;6;8;17;25;52;78;159;239, for p5 for m1;2;3;4;5;10, 15;20;24;49;74;99;123;248, and for p7 for m1;2;3;4;5;6;7;14, 21;28;35;42;48;97;146;195;244, all formup to 250.
Acknowledgments. Kedlaya was supported by NSF (CAREER grant DMS-0545904, grant DMS-1101343), DARPA (grant HR0011-09-1-0048), MIT (NEC Fund, Cecil and Ida Green professorship), and UCSD (Stefan E. Warschawski professorship). Tuitmanwas supported by the European Research Council (grant 204083).
REFERENCES
[1] B. DWORK - P. ROBBA, Effective p-adic bounds for solutions of homo- geneous linear differential equations. Tran s. Amer. Math. Soc., 259 (2) (1980), pp. 559-577.
[2] K. S. KEDLAYA,p-adic Differential Equations. Cambridge University Press, 2010.
[3] K. S. KEDLAYA, Effective p-adic cohomologyfor cyclic cubic threefolds. In Computational Algebraic and Analytic Geometry of Low-dimensional Varieties.
Amer. Math. Soc., 2012. Available at http://math.mit.edu/kedlaya/papers/.
[4] A. LAUDER,Rigid cohomologyand p-adic point counting. J. TheÂor. Nombres Bordeaux,17(2005), pp. 169-180.
[5] A. LAUDER, A recursive method for computing zeta functions of varieties.
LMS J. Comput. Math.,9(2006), pp. 222-269.
Manoscritto pervenuto in redazione il 15 novembre 2011.