Zeta functions of totally ramified p-covers of the projective line.
HANFENGLI(*) - HUIJUNEZHU(**)
ABSTRACT- In this paper we prove that there exists a Zariski dense open subset U defined over the rationals Q in the space of all one-variable rational functions with prescribed ` poles with fixed orders, such that for every geometric point f inU(Q), the L-function of the exponential sum of f at a prime p has Newton polygon approaching the Hodge polygon as p ap- proaches infinity. As an application to algebraic geometry, we prove that the p-adic Newton polygon of the zeta function of ap-cover of the projective line totally ramified at arbitrary`points with prescribed orders has an asymp- totic generic lower bound.
1. Introduction.
This paper investigates the asymptotics of the zeta functions of p- covers of the projective line which are totally (wildly) ramified at arbi- trary`points. Our approach is via Dwork's method on one-variable ex- ponential sums.
Throughout this paper we fix positive integers `;d1;. . .;d`, and let d:P`
j1dj` 2. For simplicity we assumed2 if`1. LetP1 1, P20, P3;. . .;P` be fixed poles inQof orders d1;. . .;d`, respectively.
Letf be a one-variable function overQwith these prescribed`poles. It can be written in a unique form of partial fractions: f Pd1
i1a1;ixi
P`
j2
Pdj
i1aj;i(x Pj) iwithaj;i2Q. (We assume that f has a vanish- ing constant term because this does not affect thep-adic Newton poly-
(*) Indirizzo dell'A.: Department of Mathematics, University of Toronto, Toronto, ON M5S 3G3, Canada.
E-mail: hli@fields.toronto.edu
(**) Indirizzo dell'A.: Department of Maths and Stats, McMaster University, Hamilton, ON L8S 4K1, Canada.
E-mail: zhu@cal.berkeley.edu
gons of f [13, Introduction]). Let A be the space of aj;i's with Q`
j1aj;dj 60. It is an affine (P`
j1dj)-space over Q. Let the Hodge polygon ofA, denoted by HP(A), be the lower convex graph of the pie- cewise-linear function defined on the interval [0;d] passing through the two endpoints (0;0) and (d;d=2) and assuming every slope in the list below of (horizontal) length 1:
0;. . .;0 z}|{` 1
;1;z}|{. . .` 1;1
; 1
d1;. . .;d1 1 d1
z}|{d1 1
; 1
d2;. . .;d2 1 d2
z}|{d2 1
; . . . ;1
d`;. . .;d` 1 d`
z}|{d` 1 : A non-smooth point on a polygon (as the graph of a piece-wise linear function) is called a vertex. We remark that the classical and geometrical
`Hodge polygon' for any curve (including Artin-Schreier curve as a special case) is the one with end points (0;0) and (d;d=2) and one vertex at (d=2;0).
So the Hodge polygon in our paper is different from the classical Hodge polygon. We anticipate a p-adic arithmetic interpretation of our Hodge polygon, but it remains an open question.
In [13] it is shown that in the case`1 there is a Zariski dense open subsetU defined overQsuch that every geometric closed pointf inU(Q) has p-adic Newton polygon approaching the Hodge polygon as p ap- proaches 1. Wan has proposed conjectures regarding multivariable ex- ponential sums, including the above as a special case (see [10, Conjecture 1.15]). This series of study traces back at least to Katz [4, Introduction], where Katz proposed to study exponential sums in families instead of ex- amining one at a time. He systematically studied families of multivariable Kloosterman sum in [4].
LetQf be the extension field ofQgenerated by coefficients and poles P1;. . .;P` off. For every prime numberp we fix an embeddingQ,!Qp
once and for all. This fixes a placePinQf lying overpof residue degreea for some positive integera. As usual, we letE(x)exp (P1
i0xpi=pi) be thep-adic Artin-Hasse exponential function. Letgbe a root of the p-adic logE(x) with ordp(g)p11. ThenE(g) is a primitivep-th root of unity and we setzp:E(g). LetFpbe the prime field ofpelements. LetFqbe a finite field of pa elements. Fork1, letck:Fqk!Q(zp) be a nontrivial ad- ditive character ofFqk. Henceforth we fix ck()zTrp Fqk=Fp(). LetQ`
j1dj, and all poles and leading coefficients aj;dj of f be p-adic units. Let all coefficients aj;i off be p-adically integral. (These are satisfied when pis large enough.) LetSk(fmodP)P
xck(f(x)modP) where the sum ranges
over allxinFqknfP1;. . .;P`g(wherePjare reductions ofPjmodP). TheL- function off atpis defined as
L(fmodP;T)exp X1
k1
Sk(fmodP)Tk=k
! :
This function lies inZ[zp][T] of degreed. It is independent of the choice ofP (that is, the embedding ofQ,!Qp) forplarge enough, but we remark that its Newton polygon is independent of the choice of P for all p (see [15, Section 1]). One notes immediately that for every primep(coprime to the leading coefficients, the poles and their orders) we have a map NPp() which sends everyp-adic integral pointf ofA(Zp) to the Newton polygon NPp(f) of theL-function of exponential sums offatp. Given anyf 2A(Q), we have forplarge enough thatf 2A(Zp) and hence we obtain the Newton polygon NPp(f) offatp. Presently it is known that NPp(f) lies over HP(A) for everyp. These two polygons do not always coincide. (See [15, Intro- duction].) Some investigation on first slopes suggests the behavior is ex- ceptional ifpis small (see [7, Introduction]). There has been intensive in- vestigation on how the (Archimedean) distance between NPp(f) and HP(A) on the real planeR2varies whenpapproaches infinity. Inspired by Wan's conjecture [10, Conjecture 1.15] (proved in [13] for the one-variable poly- nomial case), we believe that «almost all» points f in A(Q) satisfy limp!1NPp(f)HP(A). Our main result is the following.
THEOREM1.1. LetAbe the coefficients spacefaj;igof the f0s as in the beginning of the paper. There is a Zariski dense open subsetUdefined over Q in A such that for every geometric closed point f in U(Q) one has f 2 U(Zp)for p large enough (only depending on f ), and
p!1lim NPp(f)HP(A):
The two polygons NPp(f) and HP(A) coincide if and only if p1mod lcm(dj) (see [15, Theorem 1.1]). The case`1 is known from [13, Theorem 1.1]. For p61mod lcm(dj), the point f xd1P`
j1(x Pj) dj does not lie inU. This meansU is always a proper subset ofA.
For any f 2A(Fq) and the (generalized) Artin-Schreier curve Cf :yp yf, let NP(Cf;Fq) be the usualp-adic Newton polygon of the numerator of the zeta function of Cf=Fq. If it is shrunk by a factor of 1=(p 1) vertically and horizontally, we denote it byNP(Cp f1;Fq)
COROLLARY1.2. Let notation be as in Theorem1:1and the above. For anyf 2A(Fq)we haveNP(f;Fp 1q)lies overHP(A)with the same endpoints, and
they coincide if and only ifp1mod (lcm(dj)). Moreover, there is a Zariski dense open subsetUdefined overQinAsuch for every geometric closed point f inU(Q)one hasf2 U(Zp)for p large enough (only depending on f), and
p!1lim
NP(Cf;Fq)
p 1 HP(A).
PROOF. This follows from the theorem above and a similar argument as the proof of Corollary 1.3 in [15], which we shall omit here. p REMARK1.3. (1) The results in Theorem 1.1 and Corollary 1.2 do not depend on where those`poles are (as long as they are distinct).
(2) By Deuring-Shafarevic formula (see for instance [3, Corollary 1.5]), one knows that NPp(f) always has slope-0 segment precisely of horizontal length` 1. By symmetry it also has slope-1 segment of the same length.
See Remark 1.4 of [15].
Plan of the paper is as follows: section 2 introduces sheaves of (infinite dimensional)W-modules over some affinoid algebra arising from one-vari- able exponential sums. We consider two Frobenius mapsa1andaa. Section 4 is the main technical part, where major combinatorics of this paper is done. After working out several combinatorial observations we are able to reduce our problem to an analog of the one-variable polynomial case as that in [13]. Now back to Section 3 we improve the key lemma 3.5 of [13] to make the generic Fredholm polynomial straightforward to compute. Section 5 usesp-adic Banach theory to give a new transformation theorem froma1to aafor anya1. This approach is very different from [9] or [14]. It shreds some new light onp-adic approximations ofL-functions of exponential sums and we believe that it will find more application in the future. Finally at the end of section 5 we prove our main result Theorem 1.1.
Acknowledgments. Zhu's research was partially supported by an NSERC Discovery grant and the Harvard University. She thanks Laurent Berger and the Harvard mathematics department for hospitality during her visit in 2003. The authors also thank the referee for comments.
2. Sheaves ofW-modules over affinoid algebra.
The purpose of this section is to generalize the trace formula (see [14, Section 2]) for an exponential sum to that for families of exponential sums.
See [2] for fundamentals in rigid geometry and see [1] for an excellent setup for rigid cohomology related top-adic Dwork theory.
LetO1:Zp[zp] andV1:Qp(zp). Fix a positive integera. LetVabe the unramified extension ofV1of degreeaandOaits ring of integers. Let P^1;. . .;P^` inOa be TeichmuÈller lifts ofP1;. . .;P`inFpa. Similarly letAj;i be that ofaj;iand let~Adenote the sequence ofAj;i(we remark that for most part of the paper~Awill be treated as a variable). Lettbe the lift of Fro- benius toVa which fixesV1. Thent(Aji)Apji. Let 1j`. Pick a root g1=djofginQp(or inZp, all the same) for the rest of the paper, and denote V01:V1(g1=d1;. . .;g1=dj). Let O01 be its ring of integers. Let V0a:VaV01 and letO0abe its ring of integers. Then the affinoid algebraO0ah~Ai(with~A as variables) forms a Banach algebra overO0aunder the supremum norm.
Let 0<r<1 and r2 jV0ajp. Let Ar be the affinoid with ` deleted discs centering atP^1;. . .;P^`each of radiusron the rigid projective line P1 over V0a (as defined in [15]). The topology on Ar is given by the fundamental system of strict neighborhood Ar0 with rr0<1 and r02 jV0ajp. LetAbeArfor some unspecifiedrsufficiently close to 1 (the precise bound on the size of r was discussed in [15, Section 2]). Let H(V0a) be the ring of rigid analytic functions onAoverV0a. Then it is ap- adic Banach space overV0a. It consists of functions in one variableX of the form jP1
i0c1;iXiP`
j2
P1
i1cj;i(X P^j) i where cj;i2V0a and 8j1;limi!1jcj;ijp
ri 0. Its norm is defined as kjk maxj( supijcrj;iijp).
(See [15, Section 2.1].) LetH(V0ah~Ai): H(V0a)^V0aV0ah~Aiwhere^ means p-adic completion after tensoring. It is a p-adic Banach modules over V0ah~Ai with the natural norm on the tensor product of two Banach modules defined by the followings. For any P
vw2 H(V0a)Vah~Ai let kP
vwk inf ( maxi(kvik kwik)), where the inf ranges over all representatives P
iviwi withP
vwP
iviwi. From the p-adic Mittag-Leffler decomposition theorem derived in [15, Section 2.1], we can generalize it to the decomposition of V0ah~Ai as a Banach V0ah~Ai- module. WriteX1XorXj(X P^j) 1 for 2j`. LetZj g1=djXj. Note that~bw f1;Zi1; ;Zi`gi1is a formal basis of the BanachV0ah~Ai- moduleH(V0ah~Ai), that is, everyvinH(V0ah~Ai) can be written uniquely as an infinite sum of cj;iZij's with cj;i2V0ah~Ai and jcj;irijp!0 as i! 1. The Banach moduleH(V0ah~Ai) is orthonormalizable (even though~bwis not its orthonormal basis).
In this paper we extend thet-action so that it acts ongdJ1 trivially for any J. Below we begin to construct the Frobenius operator a1 onH(V0ah~Ai).
Recall thep-adic Artin-Hasse exponential functionE(X). Take expansion of E(gX) at X one gets E(gX)P1
m0lmXm for some lm2 O1. Clearly ordplmpm1 for allm0. In particular, for 0mp 1 the equality
holds andlmm!gm. LetFj(Xj):Qdj
i1E(gAj;iXji). Then Fj(Xj)X1
n0
Fj;n(Aj;1;. . .;Aj;dj)Xnj; whereFj;n:0 forn<0 and forn0
Fj;n:X
lm1 lmdjAmj;11 Amj;ddjj; 1
where the sum ranges over all m1;. . .;mdj 0 andPdj
k1kmkn. It is clear that Fj;n lies in O1[Aj;1;. . .;Aj;dj]. One observes that Fj(Xj)2 2 O1hAj;1;. . .;Aj;djihXji, the affinoid algebra in one variableXj(actually it lies inO1[Aj;1;. . .;Aj;dj]hXji). Taking product overj1;. . .; `, we have that F(X):Q`
j1Fj(Xj) lies inH(OahAi). Let~ ta 1be the push-forward map of ta 1, that is, for any function j, ta 1(j)ta 1jt. For example, ta 1(B=(X P^p))ta 1(B)=(X P) for any^ B2 O1h~Ai and P^ a Teich- muÈller lift of someP. LetUpbe the Dwork operator and letF(X) denote the multiplication map by F(X), as defined in [15, Section 2]. Let a1:ta 1UpF(X) denote the composition map. Thena1is ata 1-linear endomorphism ofH(V0ah~Ai) as a BanachV0ahAi-module.~
LetSbe the affinoid overV0awith affinoid algebraV0ah~Ai. IfLis a sheaf ofp-adic BanachV0ah~Ai-module (with formal basis) anda1is the Frobenius map which ista 1-linear with respect toV0ah~Ai, then we call the pair (L;a1) asheaf ofW-module of infinite rank. Note that the pair (H(V0ah~Ai);a1) can be considered as sections of a sheaf ofV0ah~Ai-module of infinite rank onA.
This is intimately related to Wan'snuclears-module of infinite rank(see [11]) if replacing hissby ourta 1. Wan has definedL-functions of nuclear s-modules and he also showed that it isp-adic meromorphic on the closed unit disc (see Wan's papers [11, 12] which proved Dwork's conjecture).
Finally we defineaa:aa1.
Recall thata1 is ata 1-linear (with respect toV0ah~Ai) completely con- tinuous endomorphism on thep-adic Banach moduleH(V0ah~Ai) overV0ahAi.~ Let 1J1;J`. Write (a1ZiJ)P^J1 P1
n0(ta 1CJn;i1;J)ZnJ1 for someCn;iJ1;J in V0ahAi. The matrix of~ a1, consisting of all these ta 1C?;?J1;J's, is a nuclear matrix (see section 5). This matrix is the subject of the next section. Below we extend Dwork, Monsky and Reich's trace formula to families of one- variable exponential sums.
THEOREM 2.1. Let f Pd1
i1a1;ixiP`
j2Pdi
i1aj;i(x Pj) i2A(Fq) and let ^f be its TeichmuÈller lift with coefficient ajibeing lifted to Aji. Let H(V0ah~Ai)r be the Banach module H(V0ah~Ai) for some suitably chosen
0<r<1with r2 jV0ajpclose enough to1 . Then L(f=Fq;T)det (1 TaajH(V0ahAi))~
det (1 TqaajH(V0ah~Ai))
lies inOah~Ai[T]as a polynomial of degree d. Its TeichmuÈller specializa- tion of~A inOalies inZ[zp][T].
PROOF. The proof is similar to that of [14, Lemma 2.7]. Let Hy:S
0<r<1H(V0ah~Ai)r. Then it is the Monsky-Washnitzer dagger space.
Then aa is a completely continuous endomorphism on Hy and the de- terminant det (1 TaajHy)det (1 TaajH(V0ahAi)~ r) for any r within suitable range in (0;1) is independent of r. Finally one knows that the coefficients are all integral so lie inOaand coefficient ofTmvanishes for all
md. We omit details of the proof. p
3. Explicit approximation of the Frobenius matrix.
This section uses some standard techniques in p-adic approximation and it is very technical. The readers are recommended to skip it at first reading and continue at the next section.
3.1 ±The nuclear matrix.
Let notation be as in the previous section. Assign f(1)0. Let f(Znj)dnjforj2 orndj1forj3. Order the elements in~bw ase1;e2; such thatf(e1)f(e2) . Consider the infinite matrix representing the endomorphisma1of theV0ah~Ai-moduleH(V0ahAi) with respect to the basis~
~bw. This matrix can be written asta 1M, where each entry ista 1CJn;i1;Jfor 1J1;J`.
Our goal of this section is to collect delicate information about entries of the matrixM. Recall the polynomialFJ;nJ inO1[~A] as in (1), which we have already built up some satisfying knowledge. Below we will expressC?;?J1;Jas a polynomial expression in these FJ1;nJ1's. In this paper the formal ex- pansion of CJ?;?1;J will always mean the formal sum in O0a[A] by the com-~ position of (2) and the formula in Lemma 3.1.
Forn;i1, and ifJ1 orJ11 then fori0 or forn0 respec- tively one has
CJn;i1;J gdJi dJn1Hnp;iJ1;J J11;2 gdJi dJn1Pnp
mnCn;mP^np mJ1 Hm;iJ1;J J13 8<
2 :
where C?;?2Zp is defined in [15, Lemma 3.1] and H?;?J1;J2 Oah~Aiis for- mulated in Lemma 3.1 below. Indeed, we recall that Cn;m is actually a rational integer and it only depends onn,mandp.
LEMMA3.1. Let~n:(n1;. . .;n`)2Z`0. (1)For i;n0, then Hn;i1;Jis equal to X F1;n1 X
0mJnJ J61
FJ;mJ nJi 1 mJi 1
P^nJJ mJ 0
B@
1 CA 0
B@
Y
j61;J
Xnj
mj0
Fj;mj nj 1 mj 1
P^njj mj 0
@
1 A
1 A; where the sum ranges over all~n2Z`0such that nn1i P`
j2njand theor depends on J1or J61, respectively.
(2)For J1;J61, one has that Hn;iJ1;Jis equal to X FJ1;nJ1 X
mJ6JJ01
FJ;mJ( 1)mJi nJmJi 1 mJi 1
(P^J P^J1) (nJmJi) 0
BB
@
1 CC A 0
BB
@
X1
m1n1
F1;m1
m1
n1 P^mJ11 n1
!
Y
j61;J1;J
X1
mj0
Fj;mj( 1)mj njmj 1 mj 1
(P^j P^J1) (njmj) 0
@
1 A
1 A where the sum ranges over all~n2Z`0such that nnJ1i P
j6J1njif JJ1and nnJ1 P
j6J1nj if J6J1.
(3)For J161and J1we have that Hn;iJ1;J is equal to X FJ1;nJ1 X
m1n1 i
F1;m1
m1i n1
P^mJ11i n1
!
Y
j6J1;1
X1
mj0
Fj;mj( 1)mj njmj 1 mj 1
(P^j P^J1) (njmj) 0
@
1 A
1 A where the sum ranges over all~n2Z`0such that nnJ1 P
j6J1nj.
PROOF. We shall use «P^j» to mean expansion atP^j. Clearly for anyJ1
one hasFJ1(XJ1)XiJ1 P^J1P1
n0FJ1;nXni.
ForJ2 one has the expansion atP^1 1:
FJ(XJ)XJi X1
m0
FJ;m(X 1(1 P^JX 1) 1)mi
^
P1X1
m0
FJ;m X1
kmi
k 1 mi 1
P^kJ (mi)X k
X1
n0
(Xn
m0
FJ;m ni 1 mi 1
P^n mJ )X n i:
ForJ161 andJ61;J1, its expansion atP^J1 is:
FJ(XJ)XJi X1
m0
FJ;m(XJ11 (P^J P^J1)) (mi)
^
PJ1X1
m0
FJ;m( 1)miX1
n0
nmi 1 mi 1
(P^J P^J1) (nmi)XJ1n
X1
n0
X1
m0
FJ;m( 1)mi nmi 1 mi 1
(P^J P^J1) (nmi)
! XJ1n:
ForJ161 andJ1 then one has FJ(X)XJi P^J1X1
n0
X1
mn i
FJ;m mi n
P^mi nJ1
! XJ1n: ByF(X)XJi (FJ(XJ)XJi)Q
j6JFj(Xj), and Key Computational Lemma of [15], one can compute and obtain (F(X)XJi)P^J1 for the case J11 or J161 presented respectively in the two formulas in our assertion. This
proves the lemma. p
REMARK3.2. If we are dealing with the case of unique pole at1then one sees easily thatC1;1?;?lies inO01[~A]. This greatly reduces the complexity of situation.
The following results were presented in [15]. See Section 3 and in particular, Theorem 3.7 of [15] for a proof. We shall usetJ1 to denote the lower bound in Lemma 3.3 c).
LEMMA3.3. Let notation be as above.
(a) For all J and nJwe haveordp(FJ;nJ)dpnJdJe1dJ(pnJ1): (b) For all J1;J, and all n;i we haveordp(Hn;iJ1;J)dJn i
1(p 1): (c) For any J1and any n we haveordp(Cn;?J1;?)dnJ
1 or ndJ1
1 depending on J11;2 or 3J1`. Moreover, ordp(CJn;i1;J1) dnp idJ
1e=(p 1) or
d(n 1)pdJ (i 1)
1 e=(p 1)depending on J11;2or3J1`, respectively.
3.2 ±Approximation by truncation.
From the previous subsection one has noticed an unpleasant feature of C?;?J1;Jfor the purpose of approximation byFJ1;nJ1's. First, in the sum for (2) whenJ13, the range ofmis too «large». Second,H?;?J1;Jof Lemma 3.1 is generally an infinite sum of FJ1;nJ1's. In this subsection we will define an approximation in terms of truncated finite sum ofFJ1;nJ1's. Below we prove two lemmas which will be used for approximation in Lemma 4.3.
For any integer 0<tp, let tCn;iJ1;J be the same as CJn;i1;J except for J13 its sum ranges over all m in the sub-interval [(n 1)p1;
(n 1)pt].
LEMMA3.4. Let3J1`,1J`. Let ndJ1and idJ. (1)For p large enough, one has
ordp(CJn;i1;J pCn;iJ1;J)>n 1 dJ1 d
p 1: 3
(2)There is a constantb>0depending only on d such that for tb one has
ordp(pCn;iJ1;J tCn;iJ1;J)>n 1 dJ1
d p 1: 4
PROOF. (1) By [15, Lemma 3.1], one knows that for anym(n 1)p one has ordp(Cn;m)1 and hence ordp(Cn;iJ1;J)1(diJ dnJ
1)p11. For ndJ1 and for p large enough one has 1(diJ dnJ
1)p11>ndJ1
1 pd1. Combining these two inequalities, one concludes.
(2) We may assumeJ13. Then for any 1vp, by Lemma 3.3, ordp(H(nJ1;J1)pv;i)(n 1)pv i
dJ1(p 1) > n dJ1
i dJ
1
p 1n 1 dJ1
d p 1;
ifvbfor someb>0 only depending ond. Therefore, ordp(pCn;iJ1;J tCn;iJ1;J)>n 1
dJ1 d p 1:
This finishes our proof. p
Fixbfor the rest of the paper. We will truncate the infinite expansion of H?;?J1;J. Letw>0 be any integer. ForJ11;2 letwHnp;iJ1;Jbe the sub-sum in Hnp;iJ1;Jwhere~n(n1;. . .;n`) are such thatnJ1 npandnj lie the interval [ w;w] forj6J1. Similarly, forJ13 letwHm;iJ1;Jbe the sub-sum ofHm;iJ1;J where ~n ranges over the finite set of vectors (n1;. . .;n`) such that nJ1 (n 1)pandnjlie in the interval [ w;w] forj6J1. ConsiderbCn;iJ1;J as a polynomial expression inH?;?J1;J's, then we setwKn;iJ1;J:bCn;iJ1;J(wHn;iJ1;J):
LEMMA3.5. There is a constantadepending only on d such that ordp(bCn;iJ1;J aKn;iJ1;J)>n 1
dJ1 d p 1: 5
PROOF. This part is similar to Lemma 3.4 2), so we omit its proof. p 3.3 ±Minimal weight terms.
Theweightof a monomial (with nonzero coefficient) (Q`
j1
Qdj
i1Akj;ij;i) in O0a[~A] is defined asP`
j1Pdj
i1ikj;i:For example, the weight ofAa1;2Ab1;3is equal to 2a3b. We will later utilize the simple observation that every monomial inFJ;nJ is of weightnJ.
We call those entries withJ1Jthe diagonal one (or blocks). As we have seen in Lemma 3.1, the off-diagonal entries are less manageable while the diagonal entries behave well in principle. For any integer 0<tp, let tM:(tCn;iJ1;J) with respect to the basis arranged in the same order as that for M. Consider the diagonal blocks, consisting of
pC?;?J;J's. Despite pC?;?J;J lives in O0ah~Ai, its minimal weight terms live in P^ZJgi ndJO1[~AJ].
LEMMA3.6. Let p>djfor all j. The minimal weight monomials ofpCn;iJ;J (with J1;2) live in the termgi nd1F1;np iwhere d1>n;i0unless n0 and i>0. For J3and n2, the minimal weight monomials ofpCn;iJ;J
live in the term
gi ndJCn;(n 1)p1P^pJ 1FJ;(n 1)p (i 1)
where dJ>n;i1.
PROOF. This follows from Lemma 3.1. We omit its proof. p Given akkmatrixM:(mij)1i;jkwith a given formal expansion of mij2 O0ah~Ai, theformal expansionof detM means the formal expansion asP
s2Sksgn(s)Qk
n1mij where the product is expanded according to the given formal expansion mij. For example, if mijCJ?;?1;J then its formal expansion is given by composition of (2) and formulas in Lemma 3.1.
LEMMA3.7. Let notation be as above and let p>djfor all j. Then in the formal expansion ofdet (pM)[k] in O0ah~Ai, all minimal weight terms are fromQ`
J1detpCn;iJ;J(with n;i1in a suitable range for J1;2and with n;i2for J3) of the diagonal blocks.
PROOF. We will show that picking an arbitrary entry on the diagonal block, every off-diagonal entry on the same row has strictly higher minimal weight among its monomials.
Let~AJstand for the vector (AJ;1;. . .;AJ;dJ). As we have noticed earlier the polynomialFJ;nJinO1[~AJ] has every monomial of equal weightnJfor any J. For simplicity we assumen;i1 here. Using data from Lemma 3.1, we find all minimal weight monomials inH?;?J1;J's illustrated below by an arrow:
Hnp;i1;1 !F1;np i;Hnp;i1;J2!F1;npi;Hnp;i2;2 !F2;np i;Hnp;i2;J62!F2;np. One also notes that forJ13 one has thatH(nJ13;J1)p1;i!FJ1;(n 1)p (i 1)ifJ1J, and H(nJ13;J1)p1;i!FJ1;(n 1)p1ifJ16J. One notices from (2) and the above that the minimal weight monomials of pCn;iJ1;J live in Hnp;iJ1;J if J11;2 and in H(nJ1;J1)p1;iif 3J1`:
Recall that forJ1 the range foriisi0. In all other cases the range isi1. From the above we conclude our claim in the beginning of the proof. Consequently, all minimal weight monomials in the formal expan- sion of the determinant detM[k]come from the diagonal blocks. By Lemma 3.6,C1;10;iandC1;iJ;J(withJ3) both have their minimal weight equal to 0 if i0 and >0 if i>0. Then it is not hard to conclude that the minimal weight monomials of det (Cn;i1;1)n;i0 (resp. det (CJ;Jn;i)n;i1) are from det (C1;1n;i)n;i1(resp. det (CJ;Jn;i)n;i2). p
For 1J`, letD[k]J :det (FJ;ip j)1i;jk2 O1[A1;. . .;Ad].
PROPOSITION3.8. Letp > djfor allj. The minimal weight monomials ofdet ((pCi;jJ;J)1i;jk)forJ 1,2(resp.det ((pCi;jJ;J)2i;jk)forJ 3) lie in D[k]J (resp.D[kJ 1]). Every monomial ofD[k]J (resp.D[kJ 1]) corresponds to a monomial in the formal expansion ofdet ((pCi;jJ;J)1i;jk)forJ 1,2(resp.
det ((pCi;jJ;J)2i;jk) for J 3) by the same permutation s2Sk in the natural way.
PROOF. It follows from Lemmas 3.6 and 3.7 above. p
3.4 ±Local at each pole.
For ease of notation, we drop the subindexJ for the rest of this sub- section. One should understand that d;Ai;Fip j;Dn stand for dJ;AJ;i;FJ;ip j;D[n]J , respectively. Let 1nd 1 and let Sn be the permutation group. LetDn:det (Fip j)1i;jn2 O1[A1;. . .;Ad]. Then we have the formal expansion ofD[n]:
D[n]X
s2Sn
sgn(s)X Yn
i1
gs;i;
where the second P
runs over all termsgs;i of the polynomialFip s(i) in O1[A1;. . .;Ad].
PROPOSITION3.9. Let1nd. Then there is a unique monomial in the above formal expansion of D[n] with highest lexicographic order (according to Ad;. . .; A1). Moreover, the p-adic order of this monomial (with coefficient) is minimal among thep-adic orders of all monomials in the above formal expansion.
REMARK3.10. We shall fix the uniques0found in the proposition for the rest of the paper. The minimal p-adic order of this monomial (with coefficient) is equal to (7) while every row achieves its minimal order in Lemma 3.3c). We shall use this fact later.
PROOF. Denote byrthe least non-negative residue ofpmodd. Recall the n by n matrix rn: frijg1i;jn where rij:ddri jd e (ri j). The properties of this matrix can be found in [13, Lemma 3.1]. LetQn
i1hs;ibe a highest-lexicographic-order-monomial in the formal expansion of D[n].
Then hs;i must be the highest-lexicographic-order-monomial in Fip s(i), which is easily seen to becip s(i)Aipdds(i)orcip s(i)Abdipds(i)cAd ri;s(i)depending on ri;s(i)0 or not, wherecip s(i)2V1. We show first that
s(i)kfor any 1i;knwithrik0:
6
Suppose that (6) does not hold. Pick a pair (i;k) among the pairs failing (6) such thatjs(i) kjis minimal. Says(j)k. Define another permutation s02Snbys0(i)kands0(j)s(i) whiles0(s)s(s) for all others. Denote by hs0;t the highest-lexicographic-order-monomial in Ftp s0(t). Then it is easy to see that the lexicographic order ofQn
t1hs0;t is strictly higher than that ofQn
t1hs;t, which is a contradiction. Therefore (6) holds.
Notice that for permutationss002Snsatisfying (6), the degree ofAdin Qn
t1hs00;t does not depend on the choice ofs00, wherehs00;t is the highest- lexicographic-order-monomial inFtp s00(t). Then the proof of [13, Lemma 3.2] shows that there exists a unique s02Sn such that Qn
t1hs0;t has highest lexicographic order among the corresponding monomials for all s002Snsatisfying (6). In fact,s0is exactly the permutation in [13, Lemma 3.2]. By the above discussion, this monomialQn
t1hs0;t also has the unique highest lexicographic order in the formal expansion ofD[n].
Next we show that thep-adic order ofQn
i1hs0;iis minimal (among the p-adic orders of the monomials in the formal expansion of D[n]). Let Qn
i1gs;ibe an arbitrary monomial in the formal expansion of D[n]. Then clearly ordp(gs;i) dipds(i)e pi s(i)rd i;s(i)for all 1in. Sincehs0;iis the highest-lexicographic-order-monomial in Fip s0(i), one sees easily that ordp(hs0;i)pi s0(i)rd i;s0(i) for all 1in. From (6) it is easy to see that ri;j ri;s0(i)j s0(i) for all 1i;jn. It follows that ordp(Qn
i1gs;i) ordp(Qn
i1hs0;i). p
REMARK3.11. In the proof of Proposition 3.9 we have noticed that the s0is exactly the permutation in [13, Lemma 3.2]. Therefore, one can always taket00, that is,fn0(~A)60 in [13, Lemma 3.5].
4. Newton polygon ofa1.
Recall that HP(A) lives on the real plane over the interval [0;d]. Be- cause of Remark 1.3, one only has to consider the part of NPp(f) with slope
<1, that is, to consider the part of NPp(f) over the interval [0;d `]. This part is our focus of this section. Suppose for some 1kd `, the point
(k;c0) is a vertex on HP(A). Then one notices that c0X2
J1
XkJ
i1
i=dJX`
J3
XkJ
i1
(i 1)=dJ
for a sequence of nonnegative integers k1;. . .;k` such that k1. . .k`k. This sequence is unique because (k;c0) is a vertex.
From now on we fix such ak.
For our purpose we also fix the residue classes ofpmoddJfor allJ. Let rJ;ij be the least nonnegative residue of (ip j)moddJ. Let s0 be the permutation in Sk which is the union of those permutations found in Proposition 3.9 locally at each pole PJ. Let sJ1 be the rational number defined by
sJ1 :(p 1)kJ1(kJ11)=2
dJ1
PkJ1
i1rJ1;i;s0(i) dJ1 7
where and is taken according to J11;2 or J13. Let s0:s1 s`. Clearlys0 c0(p 1)<kd `.
Letaandbbe the integers chosen in Lemmas 3.4 and 3.5 (they depend only ond). LetQ0:Q(g1=d1;. . .;g1=d`).
LEMMA4.1. For any J11;2,1J`and for any n;i in their range, for p large enough, there is a polynomial Gn;iJ1;JinQ0(~P)[~A]such that
aKJn;i1;Jg(p 1)n=dJ1UJ1;nGn;iJ1;Jmodg(p 1)n=dJ1d1: For the case J13, one has a similar Gn;iJ1;Jsuch that
aKJn;i1;Jg(p 1)(n 1)=dJ1UJ1;nGn;iJ1;Jmodg(p 1)(n 1)=dJ1d1;
where UJ1;n is a p-adic unit depending only on the the row index(J1;n).
PROOF. We use the same technique as [13], so we only outline our proof here for the case J1J1. Let nj2[ a;a] for j6J1. Let nJ1npP
j6J1nj i. For any~n(n1;. . .;n`) in this range, we have Fj;nj gnjdjQjmodgnjdjd1
and
FJ1;nJ1 gnJdJ11VJ1QJ1modgnJdJ11d1;