This page intentionally left blank

C A M B R I D G E S T U D I E S I N A DVA N C E D M AT H E M AT I C S 1 2 5
Editorial Board
B . B O L L O B S , W. F U LTO N , A . K ATO K , F. K I RWA N ,
P. S A R NA K , B . S I M O N , B . TOTA RO

p-adic Differential Equations

Over the last 50 years the theory of p-adic differential equations has grown into an
active area of research in its own right, and has important applications to number
theory and to computer science. This book, the first comprehensive and unified
introduction to the subject, improves and simplifies existing results as well as
including original material.
Based on a course given by the author at MIT, this modern treatment is accessible to
graduate students and researchers. Exercises are included at the end of each chapter to
help the reader review the material, and the author also provides detailed references
to the literature to aid further study.

K I R A N S . K E D L AYA is Associate Professor of Mathematics at the

Massachusetts Institute of Technology.

CAMBRIDGE STUDIES IN ADVANCED MATHEMATICS

Editorial Board:
B. Bollobs, W. Fulton, A. Katok, F. Kirwan, P. Sarnak, B. Simon, B. Totaro
All the titles listed below can be obtained from good booksellers or from Cambridge
University Press. For a complete series listing visit:
http://www.cambridge.org/series/sSeries.asp?code=CSAM
Already published
73
74
75
76
77
78
79
81
82
83
84
85
86
87
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124

B. Bollobs Random graphs (2nd edition)

R. M. Dudley Real analysis and probability (2nd edition)
T. Sheil-Small Complex polynomials
C. Voisin Hodge theory and complex algebraic geometry, I
C. Voisin Hodge theory and complex algebraic geometry, II
V. Paulsen Completely bounded maps and operator algebras
F. Gesztesy & H. Holden Soliton equations and their algebro-geometric solutions, I
S. Mukai An introduction to invariants and moduli
G. Tourlakis Lectures in logic and set theory, I
G. Tourlakis Lectures in logic and set theory, II
R. A. Bailey Association schemes
J. Carlson, S. Mller-Stach & C. Peters Period mappings and period domains
J. J. Duistermaat & J. A. C. Kolk Multidimensional real analysis, I
J. J. Duistermaat & J. A. C. Kolk Multidimensional real analysis, II
M. C. Golumbic & A. N. Trenk Tolerance graphs
L. H. Harper Global methods for combinatorial isoperimetric problems
I. Moerdijk & J. Mrcun Introduction to foliations and Lie groupoids
J. Kollr, K. E. Smith & A. Corti Rational and nearly rational varieties
D. Applebaum Lvy processes and stochastic calculus (1st edition)
B. Conrad Modular forms and the Ramanujan conjecture
M. Schechter An introduction to nonlinear analysis
R. Carter Lie algebras of finite and affine type
H. L. Montgomery & R. C. Vaughan Multiplicative number theory, I
I. Chavel Riemannian geometry (2nd edition)
D. Goldfeld Automorphic forms and L-functions for the group GL(n,R)
M. B. Marcus & J. Rosen Markov processes, Gaussian processes, and local times
P. Gille & T. Szamuely Central simple algebras and Galois cohomology
J. Bertoin Random fragmentation and coagulation processes
E. Frenkel Langlands correspondence for loop groups
A. Ambrosetti & A. Malchiodi Nonlinear analysis and semilinear elliptic problems
T. Tao & V. H. Vu Additive combinatorics
E. B. Davies Linear operators and their spectra
K. Kodaira Complex analysis
T. Ceccherini-Silberstein, F. Scarabotti & F. Tolli Harmonic analysis on finite groups
H. Geiges An introduction to contact topology
J. Faraut Analysis on Lie groups: An introduction
E. Park Complex topological K-theory
D. W. Stroock Partial differential equations for probabilists
A. Kirillov, Jr An introduction to Lie groups and Lie algebras
F. Gesztesy et al. Soliton equations and their algebro-geometric solutions, II
E. de Faria & W. de Melo Mathematical tools for one-dimensional dynamics
D. Applebaum Lvy processes and stochastic calculus (2nd edition)
T. Szamuely Galois groups and fundamental groups
G. W. Anderson, A. Guionnet & O. Zeitouni An introduction to random matrices
C. Perez-Garcia & W. H. Schikhof Locally convex spaces over non-Archimedean valued fields
P. K. Friz & N. B. Victoir Multidimensional stochastic processes as rough paths
T. Ceccherini-Silberstein, F. Scarabotti & F. Tolli Representation theory of the symmetric groups
S. Kalikow & R. McCutcheon An outline of ergodic theory
G. F. Lawler & V. Limic Random walk: A modern introduction
K. Lux & H. Pahlings Representations of groups

p-adic Differential Equations

K I R A N S . K E D L AYA
Massachusetts Institute of Technology

CAMBRIDGE UNIVERSITY PRESS

Cambridge, New York, Melbourne, Madrid, Cape Town, Singapore,

So Paulo, Delhi, Dubai, Tokyo
Cambridge University Press
The Edinburgh Building, Cambridge CB2 8RU, UK
Published in the United States of America by Cambridge University Press, New York
www.cambridge.org
Information on this title: www.cambridge.org/9780521768795
K. S. Kedlaya 2010
This publication is in copyright. Subject to statutory exception and to the
provision of relevant collective licensing agreements, no reproduction of any part
may take place without the written permission of Cambridge University Press.
First published in print format 2010
ISBN-13

978-0-511-74472-3

eBook (EBL)

ISBN-13

978-0-521-76879-5

Hardback

Cambridge University Press has no responsibility for the persistence or accuracy

of urls for external or third-party internet websites referred to in this publication,
and does not guarantee that any content on such websites is, or will remain,
accurate or appropriate.

Contents

Preface
0
0.1
0.2
0.3
0.4

Part I
1
1.1
1.2
1.3
1.4
1.5

2
2.1
2.2
2.3

3
3.1
3.2

Introductory remarks
Why p-adic differential equations?
Zeta functions of varieties
Zeta functions and p-adic differential equations
A word of caution
Notes
Exercises
Tools of p-adic Analysis
Norms on algebraic structures
Norms on abelian groups
Valuations and nonarchimedean norms
Norms on modules
Examples of nonarchimedean norms
Spherical completeness
Notes
Exercises
Newton polygons
Introduction to Newton polygons
Slope factorizations and a master factorization theorem
Applications to nonarchimedean field theory
Notes
Exercises
Ramification theory
Defect
Unramified extensions

page xiii
1
1
3
5
7
8
9
11
13
13
16
17
25
28
31
33
35
35
38
41
42
43
45
46
47
v

vi
3.3
3.4

4
4.1
4.2
4.3
4.4
4.5

Contents
Tamely ramified extensions
The case of local fields
Notes
Exercises
Matrix analysis
Singular values and eigenvalues (archimedean case)
Perturbations (archimedean case)
Singular values and eigenvalues (nonarchimedean case)
Perturbations (nonarchimedean case)
Horns inequalities
Notes
Exercises

Part II Differential Algebra

5
5.1
5.2
5.3
5.4
5.5
5.6
5.7
5.8

6
6.1
6.2
6.3
6.4
6.5
6.6
6.7
6.8
6.9

7
7.1

Formalism of differential algebra

Differential rings and differential modules
Differential modules and differential systems
Operations on differential modules
Cyclic vectors
Differential polynomials
Differential equations
Cyclic vectors: a mixed blessing
Taylor series
Notes
Exercises
Metric properties of differential modules
Spectral radii of bounded endomorphisms
Spectral radii of differential operators
A coordinate-free approach
Newton polygons for twisted polynomials
Twisted polynomials and spectral radii
The visible decomposition theorem
Matrices and the visible spectrum
A refined visible decomposition theorem
Changing the constant field
Notes
Exercises
Regular singularities
Irregularity

49
52
53
54
55
56
60
62
68
71
72
74
75
77
77
80
81
84
85
87
87
90
91
91
93
93
95
102
104
105
107
109
112
114
116
117
118
119

Contents
7.2
7.3
7.4
7.5

Exponents in the complex analytic setting

Formal solutions of regular differential equations
Index and irregularity
The TurrittinLeveltHukuhara decomposition theorem
Notes
Exercises

Part III
8
8.1
8.2
8.3
8.4
8.5
8.6

9
9.1
9.2
9.3
9.4
9.5
9.6
9.7
9.8
9.9
9.10

10
10.1
10.2
10.3
10.4
10.5
10.6
10.7
10.8

p-adic Differential Equations on Discs and Annuli

Rings of functions on discs and annuli

Power series on closed discs and annuli
Gauss norms and Newton polygons
Factorization results
Open discs and annuli
Analytic elements
More approximation arguments
Notes
Exercises
Radius and generic radius of convergence
Differential modules have no torsion
Antidifferentiation
Radius of convergence on a disc
Generic radius of convergence
Some examples in rank 1
Transfer theorems
Geometric interpretation
Subsidiary radii
Another example in rank 1
Comparison with the coordinate-free definition
Notes
Exercises
Frobenius pullback and pushforward
Why Frobenius descent?
pth powers and roots
Frobenius pullback and pushforward operations
Frobenius antecedents
Frobenius descendants and subsidiary radii
Decomposition by spectral radius
Integrality of the generic radius
Off-center Frobenius antecedents and descendants

vii
120
123
126
127
129
130
133
135
136
138
140
143
144
147
149
150
151
152
153
154
155
157
158
160
162
162
164
165
166
168
168
169
170
172
174
176
180
181

viii

11
11.1
11.2
11.3
11.4
11.5
11.6
11.7
11.8
11.9

12
12.1
12.2
12.3
12.4
12.5
12.6
12.7
12.8

13
13.1
13.2
13.3
13.4
13.5
13.6
13.7

Contents
Notes
Exercises
Variation of generic and subsidiary radii
Harmonicity of the valuation function
Variation of Newton polygons
Variation of subsidiary radii: statements
Convexity for the generic radius
Measuring small radii
Larger radii
Monotonicity
Radius versus generic radius
Subsidiary radii as radii of optimal convergence
Notes
Exercises
Decomposition by subsidiary radii
Metrical detection of units
Decomposition over a closed disc
Decomposition over a closed annulus
Decomposition over an open disc or annulus
Partial decomposition over a closed disc or annulus
Modules solvable at a boundary
Solvable modules of rank 1
Clean modules
Notes
Exercises
p-adic exponents
p-adic Liouville numbers
p-adic regular singularities
The Robba condition
Abstract p-adic exponents
Exponents for annuli
The p-adic Fuchs theorem for annuli
Transfer to a regular singularity
Notes
Exercises

182
183
184
185
186
189
190
191
193
195
197
198
199
200
201
202
203
207
209
210
211
212
214
216
216
218
218
221
222
223
225
231
234
237
238

Part IV Difference Algebra and Frobenius Modules

241

14
14.1

243
243

Formalism of difference algebra

Difference algebra

Contents
14.2
14.3
14.4
14.5
14.6

15
15.1
15.2
15.3
15.4

16
16.1
16.2
16.3
16.4
16.5

Twisted polynomials
Difference-closed fields
Difference algebra over a complete field
Hodge and Newton polygons
The DieudonnManin classification theorem
Notes
Exercises
Frobenius modules
A multitude of rings
Frobenius lifts
Generic versus special Frobenius lifts
A reverse filtration
Notes
Exercises
Frobenius modules over the Robba ring
Frobenius modules on open discs
More on the Robba ring
Pure difference modules
The slope filtration theorem
Proof of the slope filtration theorem
Notes
Exercises

ix
246
247
248
254
256
258
260
262
262
264
266
269
271
272
273
273
275
277
279
281
284
286

Part V Frobenius Structures

289

17
17.1
17.2

291
291

17.3
17.4
17.5

18
18.1
18.2
18.3
18.4
18.5

Frobenius structures on differential modules

Frobenius structures
Frobenius structures and the generic radius of
convergence
Independence from the Frobenius lift
Slope filtrations and differential structures
Extension of Frobenius structures
Notes
Exercises
Effective convergence bounds
A first bound
Effective bounds for solvable modules
Better bounds using Frobenius structures
Logarithmic growth
Nonzero exponents

294
296
298
298
299
300
301
301
302
306
308
310

19
19.1
19.2
19.3
19.4

20
20.1
20.2
20.3
20.4
20.5

21
21.1
21.2
21.3
21.4
21.5

Contents
Notes
Exercises
Galois representations and differential modules
Representations and differential modules
Finite representations and overconvergent differential
modules
The unit-root p-adic local monodromy theorem
Ramification and differential slopes
Notes
Exercises
The p-adic local monodromy theorem
Statement of the theorem
An example
Descent of sections
Local duality
When the residue field is imperfect
Notes
Exercises
The p-adic local monodromy theorem: proof
Running hypotheses
Modules of differential slope 0
Modules of rank 1
Modules of rank prime to p
The general case
Notes
Exercises

Part VI
22
22.1
22.2
22.3
23
23.1
23.2
23.3

Areas of Application

PicardFuchs modules
Origin of PicardFuchs modules
Frobenius structures on PicardFuchs modules
Relationship to zeta functions
Notes
Rigid cohomology
Isocrystals on the affine line
Crystalline and rigid cohomology
Machine computations
Notes

310
311
313
314
316
318
321
323
325
326
326
328
329
332
333
335
337
338
338
339
341
342
343
343
344

345
347
347
348
349
350
352
352
353
354
355

Contents
24
24.1
24.2
24.3
24.4
24.5

xi

p-adic Hodge theory

A few rings
(, )-modules
Galois cohomology
Differential equations from (, )-modules
Beyond Galois representations
Notes

357
357
359
361
362
363
364

References
Notation
Index

365
374
376

Preface

This book is an outgrowth of a course, taught by the author at MIT during fall 2007, on p-adic ordinary differential equations. The target audience
was graduate students with some prior background in algebraic number theory, including exposure to p-adic numbers, but not necessarily with any
background in p-adic analytic geometry (of either the Tate or Berkovich
flavors).
Custom would dictate that ordinarily this preface would continue with an
explanation of what p-adic differential equations are, and why they matter. Since we have included a whole chapter on this topic (Chapter 0), we
will devote this preface instead to a discussion of the origin of the book, its
general structure, and what makes it different from previous books on the
subject.
The subject of p-adic differential equations has been treated in several previous books. Two that we used in preparing the MIT course, and to which
we make frequent reference in the text, are those of Dwork, Gerotto, and
Sullivan [80] and of Christol [42]. Another existing book is that of Dwork [78],
but it is not a general treatise; rather, it focuses in detail on hypergeometric
functions.
However, this book develops the theory of p-adic differential equations in
a manner that differs significantly from most prior literature. Key differences
include the following.
We limit our use of cyclic vectors. This requires an initial investment in
the study of matrix inequalities (Chapter 4) and lattice approximation
arguments (especially Lemma 8.6.1), but it pays off in significantly
stronger results.
We introduce the notion of a Frobenius descendant (Chapter 10). This
complements the older construction of Frobenius antecedents, particularly in dealing with certain boundary cases where the antecedent
method does not apply.
xiii

xiv

Preface

As a result, we end up with some improvements of existing results, including the following. (Some of these can also be found in an upcoming book of
Christol [46], whose development we learned about only after this book was
mostly complete.)
We refine the Frobenius antecedent theorem of Christol and Dwork
(Theorem 10.4.2).
We extend some results of Christol and Dwork, on the variation of the
generic radius of convergence, to subsidiary radii (Theorem 11.3.2).
We extend Youngs geometric interpretation of subsidiary generic radii
of convergence beyond the range of applicability of Newton polygons
(Theorem 11.9.2).
We give quantitative versions of the ChristolMebkhout decomposition
theorem for differential modules on an annulus that are applicable even
when the modules are not solvable at a boundary (Theorems 12.2.2
and 12.3.1).
We give a somewhat simplified treatment of the theory of p-adic
exponents (Theorems 13.5.5, 13.5.6, and 13.6.1).
We sharpen the bound in the Christol transfer theorem to a disc containing a regular singularity with exponents in Z p (Theorem 13.7.1).
We give a general version of the DieudonnManin classification
theorem for difference modules over a complete nonarchimedean field
(Theorem 14.6.3).
We give improvements on the ChristolDworkRobba effective bounds
for solutions of p-adic differential equations (Theorems 18.2.1 and
18.5.1) and some related bounds that apply in the presence of a
Frobenius structure (Theorem 18.3.3). The latter can be used to recover
a theorem of Chiarellotto and Tsuzuki concerning the logarithmic
growth of solutions of differential equations with Frobenius structure
(Theorem 18.4.5).
We state a relative version of the p-adic local monodromy theorem,
formerly Crews conjecture (Theorem 20.1.4), and describe in detail
how it may be derived either from the p-adic index theory of Christol
and Mebkhout, which we treat in detail in Chapter 13, or from the slope
theory for Frobenius modules of Kedlaya, which we only sketch, in
Chapter 16.
Some of the new results are relevant in theory (in the study of higherdimensional p-adic differential equations, largely in the context of the
semistable reduction problem for overconvergent F-isocrystals, for which
see [138] and [143]) or in practice (in the explicit computation of solutions
of p-adic differential equations, e.g., for the machine computation of zeta

Preface

xv

functions of particular varieties, for which see [139]). There is also some relevance, entirely outside number theory, to the study of flat connections on
complex analytic varieties (see [144]).
Although some applications involve higher-dimensional p-adic analytic
spaces, this book treats exclusively p-adic ordinary differential equations. In
joint work with Liang Xiao [145], we have developed some extensions to
higher-dimensional spaces.
Each individual chapter of this book exhibits the following basic structure.
Before the body of the chapter, we give a brief introduction explaining what
is to be discussed and often setting some running notations or hypotheses.
After the body of the chapter, we typically include a section of afternotes, in
which we provide detailed references for results in that chapter, fill in historical
details, and add additional comments. (This practice is modeled on that in [94],
although we do not carry it out quite as fully.) Note that we have a habit of
attributing to various authors slightly stronger versions of their theorems than
the ones they originally stated; to avoid complicating the discussion in the
text, we resolve these misattributions in the afternotes instead. At the end of
a chapter we typically include a few exercises; a fair number of these request
proofs of results which are stated and used in the text but whose proofs pose
no unusual difficulties.
The chapters themselves are grouped into several parts, which we now
describe briefly. (Chapter 0, being introductory, does not fit into this grouping.)
Part I is preliminary, collecting some basic tools of p-adic analysis. However, it also includes some facts of matrix analysis (the study of the variation
of numerical invariants attached to matrices as a function of the matrix entries)
which may not be familiar to the typical reader.
Part II introduces some formalism of differential algebra, such as differential
rings and modules, twisted polynomials, and cyclic vectors, and applies these
to fields equipped with a nonarchimedean norm.
Part III begins the study of p-adic differential equations in earnest, developing some basic theory for differential modules on rings and annuli, including
the ChristolDwork theory of variation of the generic radius of convergence
and the ChristolMebkhout decomposition theory. We also include a treatment of p-adic exponents, culminating in the ChristolMebkhout structure
theorem for p-adic differential modules on an annulus satisfying the Robba
condition (i.e., having intrinsic generic radius of convergence everywhere
equal to 1).
Part IV introduces some formalism of difference algebra, and presents (without full proofs) the theory of slope filtrations for Frobenius modules over the
Robba ring.

xvi

Preface

Part V introduces the concept of a Frobenius structure on a p-adic differential module, to the point of stating the p-adic local monodromy theorem
and sketching briefly the proof techniques using either p-adic exponents or
Frobenius slope filtrations. We also discuss effective convergence bounds for
solutions of p-adic differential equations.
Part VI consists of a series of brief discussions of several areas of application of the theory of p-adic differential equations. These are somewhat more
didactic, and much less formal, than in the other parts; they are meant primarily
as suggestions for further reading.
The following diagram indicates the logical dependencies of the chapters. To
keep the diagram manageable, we have grouped together some chapters (13
and 912) and omitted Chapter 0 and the chapters of Part VI. The reader should
be aware that there is one forward reference, from Chapter 13 to Chapter 18,
but the graph remains acyclic. (There are some additional forward references
between Chapters 1 and 2, but these should not cause any difficulty.)
GFED
@ABC
13 o
H

/.-,
()*+
4I O




/.-,
()*+
()*+
/.-,
()*+
 /.-,
7I O
5 o
6O o








/.-,
()*+
HIJK
ONML
89:;
 ?>=<
912 o
8 @_ o @ 
13
  >^ >>
@@
F

>>
 @@ 
 
>>
 @@@@
 
>>
>>
 @

 


>>
89:;
89:;
?>=<
89:;
15 o   nn6 ?>=<
16
14 o  ?>=<
>>
 n nn
>>

n

n
>>
nn

n
n
>>
 nnnnn  
>
n
no
?>=<
89:;
?>=<
89:;
89:;
?>=<
?>=<
89:;
89:;
?>=<
o
o
o
19
18
17
20
21

As noted above we have not assumed that the reader is familiar with rigid
analytic geometry and so have phrased all statements more concretely in terms
of rings and modules. Although we expect that the typical reader has at least
a passing familiarity with p-adic numbers, for completeness we include a
rapid development of the algebra of complete rings and fields in the first few
chapters of the book. This development, when read on its own, may appear
somewhat idiosyncratic; its design is justified by the reuse of some material in
later chapters.

Preface

xvii

We would like to think that the background needed is that of a two-semester

undergraduate abstract algebra course. However, some basic notions from
commutative algebra do occasionally intervene, including flat modules, exact
sequences, and the snake lemma. It may be helpful to have a well-indexed text
on commutative algebra within arms reach; we like Eisenbuds book [84], but
the far slimmer Atiyah and Macdonald [9] should also suffice.
The author would like to thank the participants of the MIT course 18.787
(Topics in number theory, fall 2007) for numerous comments on the lecture notes which ultimately became this book. Particular thanks are due to Ben
Brubaker and David Speyer for giving guest lectures, and to Chris Davis, Hansheng Diao, David Harvey, Raju Krishnamoorthy, Ruochuan Liu, Eric Rosen,
and especially Liang Xiao for providing feedback. Additional feedback was
provided by Francesco Baldassarri, Laurent Berger, Bruno Chiarellotto, Gilles
Christol, Ricardo Garca Lpez, Tim Gowers, and Andrea Pulita.
During the preparation of the course and of this book, the author was supported by a National Science Foundation CAREER grant (DMS-0545904), a
Sloan Research Fellowship, MITs NEC Research Support Fund, and the MIT
Cecil and Ida Green Career Development Professorship.

0
Introductory remarks

The theory of ordinary differential equations is a fundamental instrument of

continuous mathematics, in which the central objects of study are functions
involving real numbers. It is not immediately apparent that this theory has
anything useful to say about discrete mathematics in general or number theory
in particular.
In this book we consider ordinary differential equations in which the role
of the real numbers is instead played by the field of p-adic numbers, for some
prime number p. The p-adics form a number system with enough formal similarities to the real numbers to permit meaningful analogues of notions from
calculus, such as continuity and differentiability. However, the p-adics incorporate data from arithmetic in a fundamental way; two numbers are p-adically
close together if their difference is divisible by a large power of p.
In this chapter, we first indicate briefly some ways in which p-adic differential equations appear in number theory. We then focus on an example of Dwork,
in which the p-adic behavior of Gausss hypergeometric differential equation
relates to the manifestly number-theoretic topic of the number of points on an
elliptic curve over a finite field.
Since this chapter is meant only as an introduction, it is full of statements
for which we give references instead of proofs. This practice is not typical of
the rest of this book, except for the discussions in Part VI.

0.1 Why p-adic differential equations?

Although the very existence of a highly developed theory of p-adic ordinary
differential equations is not entirely well known even within number theory,
the subject is actually almost 50 years old. Here are circumstances, past and
present, in which it arises; some of these will be taken up again in Part VI.
1

Introductory remarks

Variation of zeta functions (see Chapter 22). The original circumstance in

which p-adic differential equations appeared in number theory was Dworks
work on the variation of zeta functions of algebraic varieties over finite fields.
Roughly speaking, solving certain p-adic differential equations can give rise
to explicit formulas for the numbers of points on varieties over finite fields.
In contrast with methods involving tale cohomology, methods for studying zeta functions based on p-adic analysis (including p-adic cohomology)
lend themselves well to numerical computation. The interest in computing
zeta functions for varieties where straightforward point-counting is impossible
(e.g., curves over extremely large finite fields) has been driven by applications in computer science, the principal example being cryptography based on
elliptic or hyperelliptic curves.
p-adic cohomology (see Chapter 23). Dworks work suggested, but did not
immediately lead to, a proper analogue of tale cohomology based on p-adic
analytic techniques. Such an analogue was eventually developed by Berthelot (on the basis of some work of Monsky and Washnitzer, and also ideas
of Grothendieck); it is called rigid cohomology (see the notes at the end of this
chapter for the origin of the word rigid). The development of rigid cohomology has lagged somewhat behind that of tale cohomology, partly owing to the
emergence of some thorny problems related to the construction of a good category of coefficients. These problems, which have only recently been resolved,
are rather closely related to questions concerning p-adic differential equations;
in fact, some results presented in this book have been used to address them.
p-adic Hodge theory (see Chapter 24). The subject of p-adic Hodge theory
aims to do for the cohomology of varieties over p-adic fields what ordinary
Hodge theory does for the cohomology of varieties over C: that is, it aims
to provide a better understanding of the cohomology of a variety in its own
right, independently of the geometry of the variety. In the p-adic case, the
cohomology in question is often tale cohomology, which carries the structure
of a Galois representation.
The study of such representations, pioneered by Fontaine, involves a number
of exotic auxiliary rings (rings of p-adic periods), which serve their intended
purposes but are otherwise a bit mysterious. More recently, the work of Berger
has connected much of the theory to the study of p-adic differential equations;
notably, a key result that was originally intended for use in p-adic cohomology (the p-adic local monodromy theorem) turned out to imply an important
conjecture about Galois representations, Fontaines conjecture on potential
semistability.
Ramification theory (see Chapter 19). There are some interesting analogies between properties of differential equations over C with meromorphic

0.2 Zeta functions of varieties

singularities and properties of wildly ramified Galois representations of

p-adic fields. At some level, this is suggested by the parallel formulation of
the Langlands conjectures in the number field and function field cases. One
can use p-adic differential equations to interpolate between the two situations,
by associating differential equations with Galois representations (as in the previous item) and then using differential invariants (for example, irregularity) to
recover Galois invariants (for example, Artin and Swan conductors).
For representations of the tale fundamental group of a variety over a field
of positive characteristic of dimension greater than 1, it is difficult to construct meaningful Galois-theoretic numerical invariants. Recent work of Abbes
and Saito [1, 2] provides satisfactory definitions, but the resulting quantities
are quite difficult to calculate. One can alternatively use p-adic differential
equations to define invariants which can be somewhat easier to deal with; for
instance, one can define a differential Swan conductor which is guaranteed
to be an integer [133], whereas this is not clearly the case for the Abbes
Saito conductor. One can then equate the two conductors, deducing integrality
for the AbbesSaito conductor; this has been carried out by Chiarellotto and
Pulita [40] for one-dimensional representations and by L. Xiao [219] in the
general case.

0.2 Zeta functions of varieties

For the rest of this introduction, we return to Dworks original example showing the role of p-adic differential equations and their solutions in number
theory. This example refers to elliptic curves, for which see Silvermans book
[200] for background.
Definition 0.2.1. For in some field K , let E be the elliptic curve over K
defined by the equation
E : y 2 = x(x 1)(x )
in the projective plane. Remember that there is one point O = [0 : 1 : 0] at
infinity. There is a natural commutative group law on E (K ), with identity
element O, characterized by the property that three points add to zero if and
only if they are collinear. (It is better to say that three points add to zero if
they are the three intersections of E with some line, as this correctly permits
degenerate cases. For instance, if two of the points coincide, the line must be
the tangent to E at that point.)
For elliptic curves over finite fields, one has the following result of Hasse,
which generalizes some observations, made by Gauss and others, for certain
special cases.

Introductory remarks

Theorem 0.2.2 (Hasse). Suppose that belongs to a finite field Fq . If we write

#E (Fq ) = q + 1 aq (), then |aq ()| 2 q.

Proof. See [200, Theorem V.1.1].
Hasses theorem was later vastly generalized as follows, originally as a set
of conjectures by Weil. (Despite no longer being conjectural, these are still
commonly referred to as the Weil conjectures.)
Definition 0.2.3. For X an algebraic variety over Fq , the zeta function of X is
defined as the formal power series


 Tn
X (T ) = exp
#X (Fq n ) ;
n
n=1

another way to write this, which makes it look more like a typical zeta
function, is

X (T ) =
(1 T deg(x) )1 ,
x

where x runs over the Galois orbits of X (Fq ) and deg(x) is the size of the orbit
x. (If you prefer algebro-geometric terminology, you may run x over closed
points of the scheme X , in which case deg(x) denotes the degree of the residue
field of x over Fq .)
Example 0.2.4. For X = E one can verify that
X (T ) =

1 aq ()T + qT 2
,
(1 T )(1 qT )

using properties of the Tate module of E ; see [200, Theorem V.2.2].

A statement of the Weil conjectures is given in the following theorem.
Theorem 0.2.5 (Dwork, Grothendieck, Deligne, et al.). Let X be an algebraic variety over Fq . Then X (T ) represents a rational function of T .
Moreover, if X is smooth and proper of dimension d, we can write
X (T ) =

P1 (T ) P2d1 (T )
,
P0 (T ) P2d (T )

where each Pi (T ) has integer coefficients, satisfies Pi (0) = 1, and has all
roots in C on the circle |T | = q i/2 .
Proof. The proof of this theorem is a sufficiently massive undertaking that
even a reference is not reasonable here; instead, we give [107, Appendix C] as
a source of references. (Another useful exposition is [178].)

0.3 Zeta functions and p-adic differential equations

Remark 0.2.6. It is worth pointing out that the first complete proof of
Theorem 0.2.5 used the fact that for any prime   = p one has

(1)i Trace(F n , Heti (X, Q )),
#X (Fq n ) =
i

where Heti (X, Q ) is the i-th tale cohomology group of X (or rather, the base
change of X to Fq ) with coefficients in Q . This is an instance of the Lefschetz
trace formula in tale cohomology.

0.3 Zeta functions and p-adic differential equations

Remark 0.3.1. The interpretation of Theorem 0.2.5 in terms of tale cohomology (Remark 0.2.6) is all well and good, but there are several downsides.
An important one is that tale cohomology is not explicitly computable; for
instance, it is not straightforward to describe tale cohomology to a computer
well enough that the computer can make calculations. (The main problem is
that while one can write down tale cocycles, it is very hard to tell whether any
given cocycle is a coboundary.)
Another important downside is that you do not get every good information
about what happens to X when you vary X . This is where p-adic differential
equations enter the picture. It was observed by Dwork that if one has a family
of algebraic varieties defined over Q, the same differential equations appear
on the one hand when one studies the variation of complex periods and on the
other hand when one studies the variation of zeta functions over F p .
Here is an explicit example due to Dwork.
Definition 0.3.2. Recall that the hypergeometric series
F(a, b; c; z) =


a(a + 1) (a + i)b(b + 1) (b + i)
i=0

c(c + 1) (c + i)i!

zi

(0.3.2.1)

satisfies the hypergeometric differential equation

z(1 z)y  + (c (a + b + 1)z)y  aby = 0.

(0.3.2.2)

Set
(z) = F(1/2, 1/2; 1; z).
Over C, is related to an elliptic integral, for instance, by the formula

d
2 /2

(0 < < 1).
() =
0
1 sin2

Introductory remarks

(One can extend this to complex , but care needs to be taken with the branch
cuts.) This elliptic integral can be viewed as a period integral for the curve E ,
i.e., one is integrating some meromorphic differential form on E around some
loop (or more properly, around some homology class).
Let p be an odd prime. We now try to interpret (z) as a function of a
p-adic variable rather than a complex variable. Be aware that this means that z
can take any value in a field with a norm extending the p-adic norm on Q, not
just in Q p itself. (For the moment, you can imagine z running over a completed
algebraic closure of Q p .)
Lemma 0.3.3. The series (z) converges p-adically for |z| < 1.
Proof. Exercise.
Dwork discovered that a closely related function admits a sort of analytic
continuation.
Definition 0.3.4. Define the Igusa polynomial
H (z) =

( p1)/2
 
i=0

( p 1)/2 2 i
z.
i

Modulo p, the roots of H (z) are the values of F p for which E is a supersingular elliptic curve, i.e., for which aq () 0 (mod p). (In fact, the roots
of H (z) all belong to F p2 , by a theorem of Deuring; see [200, Theorem V.3.1].)
Dworks analytic continuation result is the following.

i
Theorem 0.3.5 (Dwork). There exists a series (z) =
i=0 Pi (z)/H (z) ,
with each Pi (z) Q p [z], converging uniformly for those z satisfying |z| 1
and |H (z)| = 1 and such that
(z) = (1)( p1)/2

(z)
(z p )

(|z| < 1).

Proof. See [213, 7].

Remark 0.3.6. Note that itself satisfies a differential equation derived
from the hypergeometric equation. We will see such equations again once
we introduce the notion of a Frobenius structure on a differential equation,
in Chapter 17.
In terms of the function , we can compute zeta functions in the Legendre
family as follows.

0.4 A word of caution

Definition 0.3.7. Let Zq be the unique unramified extension of Z p with residue

field Fq . For Fq , let [] be the unique qth root of 1 in Zq congruent to
mod p (the Teichmller lift of ).
Theorem 0.3.8 (Dwork). If q = pa and Fq is not a root of H (z), then
T 2 aq ()T + q = (T u)(T q/u),
where
u = ([])([] p ) ([] p

a1

).

That is, the quantity u is the unit root (meaning the root of valuation 0) of
the polynomial T 2 aq ()T +q occurring (up to reversal) in the zeta function.
Proof. See [213, 7].

0.4 A word of caution

Example 0.4.1. Before we embark on the study of p-adic ordinary differential
equations, a cautionary note is in order concerning the rather innocuouslooking differential equation y  = y. Over R or C, this equation is nonsingular
everywhere and its solutions y = ce x are defined everywhere.
Over a p-adic field, things are quite different. As a power series around
x = 0, we have
y=c


xn
n!
n=0

and the denominators hurt us rather than helping. In fact, the series only converges for |x| < p1/( p1) (assuming that we are normalizing in such a way
that | p| = p 1 ). For comparison, note that the logarithmic series

log

 xn
1
=
1x
n
n=1

converges for |x| < 1.

Remark 0.4.2. The conclusion to be drawn from the previous example is that
there is no fundamental theorem of ordinary differential equations over the
p-adics! In fact, the hypergeometric differential equation in the previous example was somewhat special; the fact that it had a solution in a disc where it had
no singularities was not a foregone conclusion. One of Dworks discoveries
is that this typically happens for differential equations that come from
geometry, such as the PicardFuchs equations, which arise from integrals

Introductory remarks

of algebraic functions (e.g., elliptic integrals). Another of Dworks discoveries

is that, using similar techniques to those used to study obstructions to solving
complex differential equations in singular discs, one can quantify the obstruction to solving a p-adic differential equation in a nonsingular disc. We will
carry this out later in the book.

Notes
For detailed notes on the topics discussed in Section 0.1, see the notes for the
chapters referenced.
We again mention [107, Appendix C] and [178] as starting points for further
reading about the Weil conjectures.
The notion of an analytic function specified in terms of a uniform limit
of rational functions with poles prescribed to certain regions is the original
such notion, introduced by Krasner. For this book, we will restrict our consideration of p-adic analysis to working with complete rings in this fashion,
without attempting to introduce any notion of nonarchimedean analytic geometry. However, it must be noted that it is much better in the long run to work
in terms of analytic geometry; for example, it is prohibitively difficult to deal
with partial differential equations without doing so.
That said, there are several ways to develop a theory of analytic spaces over a
nonarchimedean field. The traditional method is Tates theory of rigid analytic
spaces, so-called because one develops everything rigidly by imitating the
theory of schemes in algebraic geometry but using rings of convergent power
series instead of polynomials. The canonical foundational reference for rigid
geometry is the book of Bosch, Gntzer, and Remmert [31], but novices may
find the text of Fresnel and van der Put [93] or the lecture notes of Bosch [30]
more readable. A more recent method, which in some ways is more robust,
is Berkovichs theory of nonarchimedean analytic spaces (commonly called
Berkovich spaces), as introduced in [19] and further developed in [20]. For
both points of view, see also the lecture notes of Conrad [59].
Dworks original analysis of the Legendre family of elliptic curves using
the associated hypergeometric equation (this analysis expands earlier work of
Tate) appears in [74, 8]. The treatment in [213] is more overtly related to
p-adic cohomology.
The family of hypergeometric equations with a, b, c QZ p is rich enough
that one could devote an entire book to the study of its p-adic properties.
Indeed, Dwork did exactly this; the result was [78].
It is possible to resurrect in part the fundamental theorem of ordinary differential equations in the p-adic setting. The best results in that direction seem

Exercises

to be those of Priess-Crampe and Ribenboim [183]. One consequence of their

work is the fact that a differential equation over Q p has a solution if and only
if it has a sufficiently good approximate solution; this amounts to a differential version of Hensels lemma. We too will need noncommutative forms of
Hensels lemma; see Theorem 2.2.2.
Christol [45] has given an interesting retrospective on some of the key
ideas of Dwork, including generic points, the transfer principle, and Frobenius
structures, which resonate throughout this book.

Exercises
The reader new to p-adic numbers should postpone doing these exercises until
he or she has read Part I.
(1) Prove directly from the definition that the series F(a, b; c; z) converges
p-adically for |z| < 1 whenever a, b, c are rational numbers with
denominators not divisible by p. This implies Lemma 0.3.3.
(2) Using the fact that (z) satisfies the hypergeometric equation, write down
a nontrivial differential equation with coefficients in Q(z) satisfied by the
function (z).
(3) Check that the usual formula
lim inf |an |1/n
n

n
for the radius of convergence of the power series
n=0 an z still works
over a nonarchimedean field. That is, the series converges when |z| is less
than this radius and diverges when |z| is greater than this radius.
(4) Show that in the previous exercise, just like in the archimedean case, a
power series over a nonarchimedean field can either converge or diverge at
a value of z for which |z| equals the radius of convergence.
(5) Check that (as claimed in Example 0.4.1), under the normalization
| p| = p 1 , the exponential series exp(z) over Q p has radius of convergence p 1/( p1) , while the logarithm series log(1 z) has radius of
convergence 1.
(6) Show that, over Q p , while a power series in z which converges for |z| 1
may have an antiderivative which only converges for |z| < 1, its derivative
still converges for |z| 1. This is the reverse of what happens over an
archimedean field.

Part I
Tools of p-adic Analysis

1
Norms on algebraic structures

In this chapter, we recall some basic facts about norms (absolute values),
primarily of the nonarchimedean sort, on groups, rings, fields, and modules.
We also briefly discuss the phenomenon of spherical completeness, which is
peculiar to the nonarchimedean setting. Our discussion is not particularly comprehensive; the reader new to nonarchimedean analysis is directed to [191] for
a fuller treatment.
Several proofs in this chapter make forward references to Chapter 2.
There should be no difficulty in verifying the absence of circular
references.
Convention 1.0.1. In this book, a ring means a commutative ring unless commutativity is suppressed explicitly by describing the ring as not necessarily
commutative or implicitly by its usage in certain phrases, e.g., a ring of
twisted polynomials (Definition 5.5.1).
Notation 1.0.2. For R a ring, we denote by R the multiplicative group of
units of R.

1.1 Norms on abelian groups

Let us start by recalling some basic definitions from analysis, before specializing to the nonarchimedean case.
Definition 1.1.1. Let G be an abelian group. A seminorm (or semiabsolute
value) on G is a function | | : G [0, +) satisfying the following
conditions.
13

14

Norms on algebraic structures

(a) We have |0| = 0.

(b) For f, g G, | f g| | f | + |g|. (Equivalently, |g| = |g| and | f +
g| | f | + |g|. This condition is usually called the triangle inequality.)
We say that the seminorm | | is a norm (or absolute value) if the following
additional condition holds.
(a ) For g G, |g| = 0 if and only if g = 0.
We also express this by saying that G is separated under | |. A seminorm on
an abelian group G induces a metric topology on G, in which the basic open
subsets are the open balls, i.e., sets of the form {g G : | f g| < r } for some
f G and some r > 0.
Definition 1.1.2. Let G, G  be abelian groups equipped with seminorms | |,
| | , respectively, and let : G G  be a homomorphism. Note that
is continuous for the metric topologies on G, G  if and only if there exists a
function h : (0, +) (0, +) such that for all r > 0,
{g G : |g| < h(r )} {g G : |(g)| < r }.
We say that is bounded if there exists c 0 such that |(g)| c|g| for
all g G. We say that is isometric if |(g)| = |g| for all g G. We say
two seminorms | |1 , | |2 on G are topologically equivalent if they induce
the same metric topology, i.e., the identity morphism on G is continuous in
both directions. We say that | |1 , | |2 are metrically equivalent if there exist
c1 , c2 > 0 such that, for all g G,
|g|1 c1 |g|2 ,

|g|2 c2 |g|1 ;

this implies topological equivalence but the reverse implication does not
necessarily hold.
Definition 1.1.3. Let G be an abelian group equipped with a seminorm.
A Cauchy sequence in G under | | is a sequence {xn }
n=0 in G such that,
for any  > 0, there exists an integer N such that, for all integers m, n N ,
|xm xn | < . We say the sequence {xn }
n=0 is convergent if there exists x G
such that, for any  > 0 there exists an integer N such that, for all integers
n N , |x xn | < ; in this case, the sequence is automatically Cauchy, and
we say that x is a limit of the sequence. If G is separated under | |, then limits
are unique when they exist. We say G is complete under | | if every Cauchy
sequence has a unique limit.
Theorem 1.1.4. Let G be an abelian group equipped with a norm | |. Then
there exists an abelian group G  , equipped with a norm | | under which it is
complete, and an isometric homomorphism : G G  with dense image.
This is standard, so we only sketch the proof.

1.1 Norms on abelian groups

15

Proof. We take the set of Cauchy sequences in G and declare two sequences

{xn }
n=0 , {yn }n=0 to be equivalent if the sequence x 0 , y0 , x 1 , y1 , . . . is also
Cauchy. This is easily shown to be an equivalence relation; let G  be the set
of equivalence classes. It is then straightforward to construct the group operation (termwise addition) and the norm on G  (the limit of the norms of the
terms of the sequence). The map takes g G to the constant sequence
g, g, . . .
Definition 1.1.5. With the notation of Theorem 1.1.4, we call G  the completion of G; the group G  , equipped with the norm | | and the homomorphism
, is functorial in G. That is, any continuous homomorphism G H extends
uniquely to a continuous homomorphism G  H  between the completions;
in particular, G  is unique up to unique isomorphism. Note that one can also
define the completion even if G is only equipped with a seminorm, but only by
first quotienting by the kernel of the seminorm; in that case, the map from G
to its completion need not be injective.
Definition 1.1.6. If R is a not necessarily commutative ring and | | is a seminorm on its additive group, we say that | | is submultiplicative if the following
additional condition holds.
(c) For f, g R, | f g| | f ||g|.
We say that | | is multiplicative if the following additional condition holds.
(c ) For f, g R, | f g| = | f ||g|.
The completion of a ring R equipped with a submultiplicative seminorm
admits a natural ring structure, because the termwise product of two Cauchy
sequences is again Cauchy.
Lemma 1.1.7. Let F be a field equipped with a multiplicative norm. Then the
completion of F is also a field.

Proof. Note that if { f n }

n=0 is a Cauchy sequence in F then {| f n |}n=0 is a
Cauchy sequence in R by the triangle inequality, and so has a limit since R
is complete. Since F is equipped with a true norm, if { f n }
n=0 does not conalso
must
not
converge
to
0.
In
particular,
| f n |
verge to 0 then {| f n |}
n=0
n=0 is

1
bounded below away from 0, from which it follows easily that { f n }n=0 is also
a Cauchy sequence. This proves that every nonzero element of the completion
of F has a multiplicative inverse, as desired.

Proposition 1.1.8. Two multiplicative norms | |, | | on a field F are topologically equivalent if and only if there exists c > 0 such that |x| = |x|c for
all x F.
Proof. Exercise, or see [80, Lemma I.1.2].

16

Norms on algebraic structures

Definition 1.1.9. Let G be an abelian group equipped with a seminorm | |G ,

and let G  be a subgroup of G. The quotient seminorm on the quotient G/G 
is defined by the formula
|g + G  |G/G  = inf
{|g + g  |G }.


g G

(1.1.9.1)

If | |G is a norm, then | |G/G  is a norm if and only if G  is closed in G.

1.2 Valuations and nonarchimedean norms

We now restrict our attention to nonarchimedean absolute values, which can
be described additively (using valuations) as well as multiplicatively (using
norms). It will be convenient to switch back and forth between these points of
view throughout the book.
Definition 1.2.1. A real semivaluation on an abelian group G is a function
v : G R {+} with the following properties.
(a) We have v(0) = +.
(b) For f, g G, v( f g) min{v( f ), v(g)}.
We say v is a real valuation if the following additional condition holds.
(a ) For g G, v(g) = + if and only if g = 0.
If v is a real (semi)valuation on G, then the function | | = ev() is a
(semi)norm on G which is nonarchimedean (or ultrametric), i.e., it satisfies
the strong triangle inequality, which is given as follows.
(b ) For f, g G, | f g| max{| f |, |g|}.
Conversely, for any nonarchimedean (semi)norm | |, v() = log | | is a
real (semi)valuation. We will apply various definitions made for seminorms
to semivaluations in this manner; for instance, if R is a ring and v is a real
(semi)valuation on its additive group, we say that v is (sub)multiplicative if the
corresponding nonarchimedean (semi)norm is.
Definition 1.2.2. We say that a group is nonarchimedean if it is equipped
with a nonarchimedean norm; we say that a ring or field is nonarchimedean
if it is equipped with a multiplicative nonarchimedean norm. Note that any
nonarchimedean ring is an integral domain.
Definition 1.2.3. Let F be a nonarchimedean field. The multiplicative value
group of a nonarchimedean field F is the image of F under | |, viewed as
a subgroup of R+ ; we will often denote it simply as |F |. The additive value
group of F is the set of negative logarithms of the multiplicative value group.
If these groups are discrete and nonzero (i.e., isomorphic to Z), we say F is
discretely valued. Define also

1.3 Norms on modules

17

o F = { f F : v( f ) 0},
m F = { f F : v( f ) > 0},
F = o F /m F .
Note that o F is a local ring (the valuation ring of F), m F is the maximal ideal
of o F , and F is a field (the residue field of F).
It is worth noting that there are comparatively few archimedean (i.e., not
nonarchimedean) absolute values on fields.
Theorem 1.2.4 (Ostrowski). Let F be a field equipped with a norm | |. Then
| | fails to be nonarchimedean if and only if the sequence |1|, |2|, |3|, . . . is
unbounded. In that case, F is isomorphic to a subfield of C equipped with the
restriction of the usual absolute value.
Proof. Exercise, or see [191, 2.1.6] and [191, 2.2.4], respectively.

1.3 Norms on modules

When considering norms on modules, we usually require compatibility with
the underlying ring.
Definition 1.3.1. Let R be a ring equipped with a multiplicative seminorm | |,
and let M be an R-module equipped with a seminorm | | M . We say that | | M
is compatible with | | (or with R) if the following conditions hold.
(a) For f R, x M, | f x| M = | f ||x| M .
(b) If | | is nonarchimedean, then so is | | M .
Note that (b) is not superfluous; see the end-of-chapter exercises. If R is a field,
then two norms | | M , | |M on M compatible with R are metrically equivalent
if and only if they are topologically equivalent (exercise).
One thing to be aware of is that if M  is a quotient of M and | | M  is the
quotient norm on M  induced by | | M , then in general we cannot say that | | M 
is compatible with R. Rather, we only have the inequality
| f x  | M  | f ||x  | M 

( f R, x  M  );

this implies compatibility only if R is a field.

We can generate a rich supply of norms on modules via the following
construction.
Definition 1.3.2. Let R be a ring equipped with a multiplicative (semi)norm
| |, and let M be a finite free R-module. For B a basis of M, define the
supremum (semi)norm of M with respect to B by setting

18

Norms on algebraic structures







cb b = sup{|cb |}

 bB


(cb R).

bB

This (semi)norm extends canonically to M R S for any isometric inclusion

R S. (The situation is more complicated for arbitrary (semi)norms; see
Definition 1.3.10 below.)
We say that a seminorm on M is supremum-equivalent if it is metrically
equivalent to the supremum seminorm with respect to some basis; the same is
then true of any basis, by Lemma 1.3.3 below. In particular, if | | is a norm
then any supremum-equivalent seminorm is a norm.
Lemma 1.3.3. Let R be a ring equipped with a multiplicative seminorm | |,
and let M be a finite free R-module. Then for any two bases B1 , B2 of M, the
supremum seminorms of M defined by B1 and B2 are metrically equivalent.
Proof. Put B1 = {m 1,1 , . . . , m 1,n } and B2 = {m 2,1 , . . . , m 2,n }. Define the
n n matrix A over R by the formula
m 2, j =

n


Ai j m 1,i ;

i=1

then A is invertible. In particular, we cannot have | Ai j | = 0 for all i, j.

For x M, we can uniquely write x = a1,1 m 1,1 + + a1,n m 1,n =
a2,1 m 2,1 + + a2,n m 2,n with ai, j R. We then have
a1,i =

n


Ai j a2, j

(i = 1, . . . , n)

j=1

and so

n
n 

max{|a1,i |}
|Ai j | max{|a2, j |}.
i

i=1 j=1

This inequality, together with the corresponding one with the bases reversed
(involving the matrix A1 ), implies the claim.
Corollary 1.3.4. Let R S be an isometric inclusion of rings equipped
with multiplicative seminorms. Let M be a finite free R-module. Let | | M
be a seminorm on M that is compatible with R, which is the restriction of a
supremum-equivalent seminorm on M R S that is compatible with S. Then
| | M is supremum-equivalent.
Proof. Put N = M R S, and let | | N be a supremum-equivalent norm on
N , compatible with S, whose restriction to M equals | | M . Pick any basis B

1.3 Norms on modules

19

of M; then B is also a basis of N . Since | | N is equivalent to the supremum

norm on N defined by some basis, by Lemma 1.3.3 it is also equivalent to the
supremum norm defined by B. By restriction, we see that | | M is equivalent to
the supremum norm on M defined by B.
The notion of supremum-equivalence is well-behaved under quotients.
Lemma 1.3.5. Let R be a ring equipped with a multiplicative seminorm | |,
let M be a finite free R-module, and let M1 be a finite free R-submodule of M
such that M/M1 is also free. Let | | M be a supremum-equivalent norm on M
compatible with R. Then the quotient norm | | M1 on M1 induced by | | M is
also supremum-equivalent.
Proof. Let m 1 , . . . , m k be a basis of M1 , and choose m k+1 , . . . , m n M lifting a basis of M/M1 . Then m 1 , . . . , m n is a basis of M; by Lemma 1.3.3, | | M
is equivalent to the supremum norm defined by m 1 , . . . , m n . That is, there exist
c1 , c2 > 0 such that, for any x = a1 m 1 + + an m n M,
c1 max{|a1 |, . . . , |an |} |x| M c2 max{|a1 |, . . . , |an |}.
Then for any ak+1 , . . . , an R, we have
c1 max{|ak+1 |, . . . , |an |} = c1

inf

a1 ,...,ak R

inf

a1 ,...,ak R

{max{|a1 |, . . . , |an |}}

{|a1 m 1 + + an m n | M }

= |ak+1 m k+1 + + an m n | M1
|ak+1 m k+1 + + an m n | M
c2 max{|ak+1 |, . . . , |an |}.
Thus | | M1 is equivalent to the supremum norm defined by the images of
m k+1 , . . . , m n in M1 , proving the desired result.
In general, even over a field, not every compatible norm on a vector space
need be supremum-equivalent; see the exercises. However, such supremumequivalence is true for complete fields.
Theorem 1.3.6. Let F be a field complete for a norm | |, and let V be a
finite-dimensional vector space over F. Then any two norms on V compatible
with F are metrically equivalent.
Proof. In the archimedean case, apply Theorem 1.2.4 to deduce that F = R
or F = C, then use compactness of the unit ball. In the nonarchimedean case,
we proceed as follows. (See [80, Theorem I.3.2] for a different proof.)

20

Norms on algebraic structures

We proceed by induction on n, the case n = 1 being trivial. Let m 1 , . . . , m n

be any basis of V . It suffices to show that any given norm | | on V compatible with F is equivalent to the supremum norm defined by m 1 , . . . , m n . One
inequality is evident: for any a1 , . . . , an F, we have
|a1 m 1 + + an m n |V max{|m i |} max{|ai |}.
i

Put V  = V /Fm 1 . Let | |V  denote the quotient seminorm on V  induced

by | |V . This seminorm is compatible with F, but we must check that it is
indeed a norm. Suppose on the contrary that a2 , . . . , an F are such that
|a2 m 2 + + an m n |V  = 0. Then we can choose a sequence a1,1 , a1,2 , . . .
of elements of F such that |a1,i m 1 + a2 m 2 + + an m n |V 0 as i .
But then |a1,i a1, j ||m 1 |V = |(a1,i a1, j )m 1 |V 0 as i, j , so the
a1,i form a Cauchy sequence. Since F is complete, this Cauchy sequence has
a limit a1 , and |a1 m 1 + + an m n |V = 0 contrary to the hypothesis that | |V
is a norm.
Hence | |V  is indeed a norm. By the induction hypothesis, there exists
c2 > 0 such that
|a2 m 2 + + an m n |V  c2 max{|ai |}.
i

We then have
|a1 m 1 + + an m n |V max{|a1 m 1 |V , |a2 m 2 + + an m n |V }
max{|a1 m 1 |V , |a2 m 2 + + an m n |V  }
min{|m 1 |, c2 } max{|ai |},
i

proving that | |V is equivalent to the supremum norm defined by m 1 , . . . , m n .

Even if a norm is supremum-equivalent, it need not be equal to the supremum norm defined by any basis. However, one can approximate supremumequivalent norms using supremum norms as follows. For a stronger result in
the spherically complete case, see Lemma 1.5.5.
Lemma 1.3.7 (Approximation lemma). Let F be a nonarchimedean field, let
V be a finite-dimensional vector space over F, and let | |V be a supremumequivalent norm on V compatible with F. Assume that either:
(a) c > 1 and the value group of F is not discrete; or
(b) c 1 and the value groups of F and V coincide and are discrete.
Then there exists a basis of V defining a supremum norm | |V for which
c1 |x|V |x|V c|x|V

(x V ).

1.3 Norms on modules

21

Proof. We induct on n, with trivial base case n = 0. For n > 0, pick any
nonzero m 1 V , and put V1 = V /Fm 1 . Using (a) or (b), we can rescale m 1
by an element of F to force 1 |m 1 |V c2/3 .
Equip V1 with the quotient seminorm ||V1 induced by ||V . By Lemma 1.3.5,
| |V1 is again supremum-equivalent. Moreover, in case (b) the infimum in
(1.1.9.1) is always achieved, i.e., for every x1 V1 , there exists x V lifting
x1 with |x|V = |x1 |V1 . Hence V1 again satisfies (b).
We may now apply the induction hypothesis to V1 to produce a basis
m 2,1 , . . . , m n,1 of V1 defining a supremum norm | |V1 for which
c1/3 |x1 |V1 |x1 |V1 c1/3 |x1 |V1

(x1 V1 ).

For i = 2, . . . , n, choose m i V lifting m i,1 such that |m i |V c1/3 |m i,1 |V1 ;

then
|m i |V c1/3 |m i,1 |V1 c2/3 |m i,1 |V1 = c2/3 .
Let | |V be the supremum norm defined by m 1 , . . . , m n . For x V , write
x = a1 m 1 + + an m n with ai F. On the one hand,
|x|V max {|ai ||m i |V } c2/3 |x|V .
1in

On the other hand, if x1 is the image of x in V1 then

|x1 |V1 c1/3 |x1 |V1 c1/3 |x|V ,
so |a2 |, . . . , |an | c1/3 |x|V . Moreover,
|a1 m 1 |V max{|x|V , |x a1 m 1 |V }
max{|x|V , c2/3 |x a1 m 1 |V }
= max{|x|V , c2/3 max{|a2 |, . . . , |an |}}
c|x|V .
Since |m 1 |V 1, we deduce |a1 | c|x|V and so |x|V c|x|V . This proves
the desired inequalities.
We need the following infinite-dimensional analogue of Theorem 1.3.6,
taken from [195, Proposition 10.4]. Be aware that the situation in the archimedean case is much subtler; see the notes at the end of the chapter.
Lemma 1.3.8. Let F be a complete nonarchimedean field. Let V be an
F-vector space equipped with a norm | |V compatible with F. Suppose
that V contains a dense F-subspace of countable infinite dimension over F.
Then there exists a sequence m 1 , m 2 , . . . of elements of V with the following
properties.

22

Norms on algebraic structures

(a) For each m V , there is a unique sequence a1 , a2 , . . . of elements of

ai m i converges to m.
F such that the series i=1
(b) With notation as in (a), the function | |V defined by
|m|V = sup{|ai m i |V }
i

is a norm on V compatible with F and metrically equivalent to | |V .

Proof. Choose an ascending sequence of F-subspaces 0 = V0 V1 ,
with dim F Vn = n, whose union is dense in V . For each n > 0, pick some
m n,0 Vn \ Vn1 .
Let | |n be the quotient seminorm on V /Vn induced by | |V . As in the
proof of Theorem 1.3.6, we may show by induction on n that | |n is a norm,
as follows. The claim for n = 0 is given. Supposing that | |n1 is a norm,
let m V be an element with |m|n = 0. There must exist a sequence ai of
elements of F with |ai m n,0 + m|n1 0 as i . Since |m n,0 |n1  = 0
by the induction hypothesis, the ai must form a Cauchy sequence in F whose
limit a satisfies |am n,0 + m|n1 = 0. Again by the induction hypothesis, m
and am n,0 represent the same class in V /Vn1 , so m represents the zero
class in V /Vn . Hence | |n is a norm.
Choose an increasing sequence of real numbers 0 < r1 < r2 < < 1.
Since | |n1 is a norm, we have |m n,0 |n1  = 0. We can thus choose m n
m n,0 + Vn1 with m 1 = m 1,0 and
|m n |n1 = |m n,0 |n1

rn
|m n |V
rn+1

(n > 1).

For m Vn1 and a F, we have |am n + m|V |am n |n1

(rn /rn+1 )|am n |V . If |am n |V = |m|V , this yields
|am n + m|V

rn
rn+1

max{|am n |V , |m|V };

the same holds if |am n |V  = |m|V since in that case |am n + m|V = max
{|am n |V , |m|V }.
By induction on n, we deduce that, for a1 , . . . , an F,
|a1 m 1 + + an m n |V

r1
rn+1

max{|a1 m 1 |V , . . . , |an m n |V }

r1 max{|a1 m 1 |V , . . . , |an m n |V }.
Combining this with the evident inequality
|a1 m 1 + + an m n |V max{|a1 m 1 |V , . . . , |an m n |V },

1.3 Norms on modules

23

we conclude that on n Vn the seminorms | |V and | |V are metrically equivalent. Consequently, | |V extends by continuity to a function on V which is
metrically equivalent to | |V and hence is a norm (which is evidently compatible with F). This assertion will imply both (a) and (b) as soon as we have
established the existence aspect of (a), which we will do now.
Given m V , by hypothesis there exists a sequence x1 , x2 , . . . of elements

ai, j m i , where
of n Vn converging to m. For j = 1, 2, . . . write x j = i=1
is
a
Cauchy
sequence, for
only finitely many ai, j are nonzero. Since {x j }
j=1
each  > 0 there exists N such that, for j, j  N , |x j x j  |V . Since | |V
is metrically equivalent to | |V , for each  > 0 there also exists N such that,
for j, j  N , |x j x j  |V .
On the one hand this implies that, for each fixed i, the sequence {ai, j }
j=1
is Cauchy. Since F is complete, this sequence has a limit ai . On the other
hand, for j = N there exists some i 0 such that ai, j = 0 for all i i 0 . If we


write xi, j =
h=i+1 ah, j m j , for all j N and all i i 0 we have |xi, j |V
 and hence |xi, j |V . For fixed i, xi, j converges to m a1 m 1
ai m i as j ; hence, for all i i 0 , |m a1 m 1 ai m i |V  so

ai m i converges to m. As noted earlier, both (a) and (b) now

the series i=1
follow.
Definition 1.3.9. For F a field complete for a norm | |, a Banach space over
F is a vector space V over F equipped with a norm compatible with | |, under
which it is complete. For V a Banach space and W a closed subspace, the
quotient V /W is again complete. See [214] or [195] for a full development of
the theory of Banach spaces and other topological vector spaces over complete
nonarchimedean fields.
Definition 1.3.10. Let R be a nonarchimedean ring, and let M, N be modules
over R equipped with seminorms | | M , | | N compatible with R. The product
seminorm on M R N is defined by the formula


s

|x| M R N = inf max {|m i | M |n i | N } : x =
m i ni .
1is

i=1

As in the case of the quotient seminorm, it is clear that the product seminorm is
a seminorm, but it is not clear whether it is compatible with R unless R happens
to be a field. Moreover for norms | | M and | | N , it is not clear whether the
product seminorm is a norm. However, if M and N are finite free R-modules
equipped with supremum-equivalent norms then the product seminorm will
be supremum-equivalent, which forces it to be a norm. See also the following
lemma.

24

Norms on algebraic structures

Lemma 1.3.11. Let F be a complete nonarchimedean field. Let V and W be

(possibly infinite-dimensional) vector spaces over F equipped with norms | |V
and ||W compatible with F. Then the product seminorm on V F W is a norm.
Proof. Suppose first that V admits a dense F-subspace of at most countably
infinite dimension. Then by Theorem 1.3.6 and/or Lemma 1.3.8, we can find
a (finite or infinite) sequence m 1 , m 2 , . . . of elements of V such that every

ai m i , with ai
element can be uniquely written as a convergent series i=1

F, and | |V is equivalent to the norm | |V defined by






ai m i  = sup{|ai m i |V }
(ai F).



i
i=1

More precisely, we have c||V ||V ||V for some c > 0. Let j : V F

ai m i to a j . By tensoring with W , we obtain a

be the projection carrying i=1
projection j,W : V F W W . For x V F W , we define
|x|V F W = sup{|m j |V | j,W (x)|W }.
j

s
This gives a norm by the following argument. Let
k=1 yk z k be any
representation of x V F W with yk V and z k W , so that
j,W (x) =

s


j (yk )z k .

k=1

Suppose that |x|V F W = 0; then choose the representation of x to minimize

s. If s > 0 then yk  = 0 for all k and the z k must be linearly independent over
F. We can then choose j and k such that j (yk )  = 0, but then
0 = j,W (x) =

s


j (yk )z k ,

k=1

a contradiction. Hence s = 0 and so x = 0.

s
For x =
k=1 yk z k V F W and any positive integer N , we can
express x also as
m 1 1,W (x)+ +m N N ,W (x)+

s


(yk 1 (yk )m 1 N (yk )m N )z k ;

k=1

as N , the product seminorm of the sum over k tends to zero. We thus

conclude that, on the one hand,
|x|V F W |x|V F W .

1.4 Examples of nonarchimedean norms

25

On the other hand,

max{|yk |V |z k |W } c sup max{| j (yk )m j |V |z k |W }
k

 s
 




c sup |m j |V 
j (yk )z k 


j


k=1

= c sup{|m j |V | j,W (x)|W },

j

so that
|x|V F W c|x|V F W .
That is, the product seminorm is equivalent to | |V F W and so is a norm.

s
In the general case, suppose on the contrary that x =
j=1 m j n j
V F W had product seminorm 0. This would mean that we can find a sequence

i
m i, j n i, j , and so
xi V F W in which each xi can be represented as sj=1
lim max{|m i, j |V |n i, j |W } = 0.

Then the same data would be available if we replaced V by the closure of the
F-subspace spanned by the m j and the m i, j , and similarly for W . We may thus
apply the previous case to obtain a contradiction.

1.4 Examples of nonarchimedean norms

Example 1.4.1. For any field F, there is a trivial norm of F defined by

1 f  = 0,
| f |triv =
0 f = 0.
This norm is nonarchimedean, and F is complete under it. The trivial case will
always be allowed unless explicitly excluded; it is often a useful input into a
highly nontrivial construction, as in the next few examples.
Example 1.4.2. Let F be any field, and let F((t)) denote the field of formal
Laurent series. The t-adic valuation vt on F is defined as follows: for f =

i
i ci t F((t)), vt ( f ) is the least i for which ci  = 0. This exponentiates to
give a t-adic norm under which F((t)) is complete and discretely valued. (See
Example 1.5.8 for a variation on this construction.)
Before introducing our next example, we make a more general definition for
later use.

26

Norms on algebraic structures

Definition 1.4.3. Let R be a ring equipped with a nonarchimedean submultiplicative (semi)norm | |. For 0, define the -Gauss (semi)norm | | on
the polynomial ring R[T ] by





i
Pi T  = max{|Pi | i };



i
i

it is clearly submultiplicative. Moreover, it is also multiplicative if | | is;

however, we will postpone the verification of this to the next chapter (see
Proposition 2.1.2). For r R, we define the r -Gauss (semi)valuation vr as
the (semi)valuation associated with the er -Gauss (semi)norm.
Remark 1.4.4. The definition of the -Gauss norm depends on the choice of
the indeterminate T ; that is, it is not equivariant for arbitrary endomorphisms
of the ring R[T ]. For clarity, we will sometimes need to specify that the Gauss
norm is being defined with respect to a particular indeterminate.
Example 1.4.5. For F a nonarchimedean field and > 0, the -Gauss norm
on F[t] (with respect to t) is a multiplicative norm, so it extends to the rational
function field F(t). Note that F(t) is discretely valued under the -Gauss norm
if and only if either:
(a) F carries the trivial norm and  = 1; or
(b) F is discretely valued and belongs to the divisible closure of the
multiplicative value group of F.
In case (a) the -Gauss norm is equivalent to the t-adic norm if 0 < < 1, to
the trivial norm if = 1, and to the t 1 -adic norm if > 1.
So far we have not mentioned the principal examples from number theory;
let us do so now.
Example 1.4.6. For p a prime number, the p-adic norm | | p on Q is defined
as follows. Given f = r/s with r, s Z, write r = pa m and s = p b n with
m, n not divisible by p and then put
| f | p = p a+b .
In particular, we have normalized in such a way that | p| p = p 1 ; this convention is usually taken in order to make the product formula hold. Namely, for
any f Q, if | | denotes the usual archimedean absolute value then

| f |
| f | p = 1.
p

Completing Q under | | p gives the field of p-adic numbers Q p ; it is discretely

valued. Its valuation ring is denoted Z p and called the ring of p-adic integers.

1.4 Examples of nonarchimedean norms

27

Remark 1.4.7. When converting the p-adic norm into a valuation, it is common practice to use base- p logarithms. We have instead opted to keep the
factor of log p visible when we take logarithms. One may liken this practice
to using metric units rather than normalizing some dimensioned constants to 1
(e.g., the speed of light).
Just as the only archimedean norm on Q is the usual one, every nontrivial
nonarchimedean norm on Q is essentially a p-adic norm, again by a theorem
of Ostrowski.
Theorem 1.4.8 (Ostrowski). Any nontrivial nonarchimedean norm on Q is
equivalent to the p-adic norm for some prime p.
Proof. See [191, 2.2.4].
To equip extensions of Q p with norms, we use the following result. As usual,
when E is a finite extension of a field F we write [E : F] for the degree of the
field extension, i.e., the dimension of E as an F-vector space.
Theorem 1.4.9. Let F be a complete nonarchimedean field. Then any finite
extension E of F admits a unique extension of | | to a norm on E (under
which E is also complete).
Proof. We only prove uniqueness now; existence will be established in
Section 2.3. Let ||1 and ||2 be two extensions of || to norms on E. Then these
in particular give norms on E viewed as an F-vector space; by Theorem 1.3.6,
these norms are metrically equivalent; that is, there exist c1 , c2 > 0 such that
|x|1 c1 |x|2 ,

|x|2 c2 |x|1

(x E).

They are also both metrically equivalent to the supremum norm for some basis
of E over F, under which E is evidently complete.
We now use the extra information that | |1 and | |2 are multiplicative
(because they are norms on E as a field in its own right). For any positive
integer n, we may substitute x n in place of x in the previous inequalities and
then take nth roots, to obtain
1/n

|x|1 c1 |x|2 ,

1/n

|x|2 c2 |x|1

(x E).

Taking limits as n gives |x|1 = |x|2 , as desired.

Remark 1.4.10. The completeness of F is crucial in Theorem 1.4.9. For
instance, the 5-adic norm on Q extends in two different ways to the Gaussian
rational numbers Q(i), depending on whether |2 + i| = 51 and |2 i| = 1,
or vice versa.

28

Norms on algebraic structures

Because of the uniqueness in Theorem 1.4.9, it also follows that any algebraic extension E of F, finite or not, inherits a unique extension of | |.
However, if [E : F] = then E is not complete, so we may prefer to use its
completion instead. For instance, if F = Q p , we define C p to be the completion of an algebraic closure of Q p . One might worry that this may launch us
into an endless cycle of completion and algebraic closure, but fortunately this
does not occur.
Theorem 1.4.11. Let F be an algebraically closed nonarchimedean field.
Then the completion of F is also algebraically closed.
For the proof, see Section 2.3.

1.5 Spherical completeness

For nonarchimedean fields there is an important distinction between two different notions of completeness, which does not appear in the archimedean
case.
Definition 1.5.1. A metric space is complete if any decreasing sequence of
closed balls with radii tending to 0 has nonempty intersection. (For an abelian
group equipped with a norm this reproduces our earlier definition.) A metric space is spherically complete if any decreasing sequence of closed balls,
regardless of radii, has nonempty intersection.
Example 1.5.2. The fields R and C with their usual absolute value are
spherically complete. Any complete nonarchimedean field which is discretely
valued, e.g., Q p or C((t)), is spherically complete. Any finite-dimensional vector space over a spherically complete nonarchimedean field equipped with a
compatible norm is spherically complete (exercise); in particular, any finite
extension of a spherically complete nonarchimedean field is again spherically
complete. However, the completion of an infinite algebraic extension of Q p is
not spherically complete unless it is discretely valued; see the end-of-chapter
exercises.
Theorem 1.5.3 (KaplanskyKrull). Any nonarchimedean field embeds isometrically into a spherically complete nonarchimedean field. (However, the
construction is not functorial; see the notes.)
Proof. Since completion is functorial, we may assume we are starting with
a complete nonarchimedean field F. It was originally shown by Krull [151,
Theorem 24] that F admits an extension which is maximally complete, in the
sense of not admitting any extensions preserving both the value group and the

1.5 Spherical completeness

29

residue field. (In fact, this is not difficult to prove using Zorns lemma.) The
equivalence of this condition with spherical completeness was then proved by
Kaplansky [118, Theorem 4]. See also [214, p. 151].
One can also prove the result more directly; for instance, the case F =
Q p is explained in detail in [191, 3]. For the case F = K ((t)), see
Example 1.5.8.
One benefit of the hypothesis of spherical completeness is that it can
simplify the construction of quotient norms.
Lemma 1.5.4. Let F be a spherically complete nonarchimedean field, let V
be a finite-dimensional vector space over F, and let | |V be a norm on V compatible with F (and which must be supremum-equivalent by Theorem 1.3.6).
Let V  be a quotient of V , and let | |V  be the quotient norm on V  induced by
| |V . Then, for every x  V  , there exists x V lifting x  with |x|V = |x  |V  .
Proof. We first treat the case where dim F (V  ) = dim F (V ) 1. In this case,
we can choose m 1 V so that V  = V /Fm 1 . Given x  V  , start with any
lift x0 of x  to V . Any other lift of x  to V can be written uniquely as x0 + am 1
for some a F.
For  > 0, let B be the set of a F such that |x0 + am 1 |V |x  |V  + .
By the definition of |x  |V  , B is nonempty. Pick any a B and define
r (a, ) = sup {|b a|}.
bB

Then on the one hand B is contained in the closed ball of radius r (a, ) centered at a. On the other hand, for any r < r (a, ) there exists b B with
r |b a|, so
r |m 1 |V |b a||m 1 |V
max{|x0 + am 1 |V , |x0 + bm 1 |V }
|x  |V  + .
By taking limits, we may deduce that r (a, )|m 1 |V |x  |V  + . Hence, for
any c F with |c a| r (a, ),
|x0 + cm 1 |V max{|x0 + am 1 |V , |(c a)m 1 |V }
|x  |V  + 
and so c B .
We conclude that B must equal the closed ball of radius r (a, ) centered at
a. As  decreases, the B form a decreasing family of closed balls in F. Since

30

Norms on algebraic structures

F is spherically complete, the intersection of the B is nonempty. For any a in

this intersection, x = x0 + am 1 is a lift of x  to V satisfying |x|V = |x  |V1 .
Having completed the proof in the case where dim F (V  ) = dim F (V ) 1,
we may treat the general case by induction on dim F (V ) dim F (V  ). There is
nothing to check if the difference is 0, and the above argument applies if the
difference is 1. Otherwise, we can choose a nontrivial quotient V  of V which
in turn has V  as a nontrivial quotient. Define the quotient norm | |V  on V 
induced by | |V ; then | |V and | |V  induce the same quotient norm | |V  on
V  . To lift x  V  to V while preserving its norm, it thus suffices to apply the
induction hypothesis first to lift x  to x  V  while preserving its norm and
then to lift x  to x V while preserving its norm.
In the case of a spherically complete field, we obtain the following refinement of the approximation lemma (Lemma 1.3.7).
Lemma 1.5.5. Let F be a spherically complete nonarchimedean field, let V
be a finite-dimensional vector space over F, and let | |V be a norm on V
compatible with F and having the same value group as F. Then | |V is the
supremum norm defined by some basis of V .
Proof. Note that | |V is supremum-equivalent by Theorem 1.3.6. Define V1 as
in the proof of Lemma 1.3.7. By Lemma 1.5.4, any x  V1 lifts to some x V
with |x|V = |x  |V1 . We may thus imitate the proof of case (b) of Lemma 1.3.7
to prove the desired result.
We leave as an exercise the following alternate characterizations of spherical
completeness.
Definition 1.5.6. Let X be a nonarchimedean metric space with distance
function d. A sequence {xn }
n=0 is pseudoconvergent if, for some n, either
xn = xn+1 = or d(xn , xn+1 ) > d(xn+1 , xn+2 ) > . An element x X
is a pseudolimit if, for all sufficiently large n, d(x, xn ) d(xn , xn+1 ).
Proposition 1.5.7. Let X be a nonarchimedean metric space. Then the
following conditions are equivalent.
(a) The space X is spherically complete.
(b) For any sequence B1 , B2 , . . . of balls in X , if any two of the balls have
nonempty intersection then the intersection of all the balls is nonempty.
(c) Every pseudoconvergent sequence in X has a pseudolimit.
Proof. Exercise.
To conclude, we give one explicit example of a field which is spherically
complete without being discretely valued. This example is used in the proof of
the slope filtration theorem (Theorem 16.4.1).

Notes

31

Example 1.5.8. Let K be an arbitrary field. A generalized power series, or

MalcevNeumann series, over K is a formal sum iQ xi t i for which the set

of i Q with xi  = 0 is a well-ordered subset of Q (i.e., it contains no infinite
decreasing subsequence). Let K ((t Q )) be the set of such series; we may equip

K ((t Q )) with a t-adic valuation by sending iQ xi t i to the least index i for

which xi  = 0.
We can demonstrate that K ((t Q )) is a spherically complete abelian group as
follows. Let B1 , B2 , . . . be a decreasing sequence of balls of radii r1 , r2 , . . . .
Put r = lim j r j ; we may assume that r  = r j for all j (otherwise the balls
are all equal after some point and their intersection is clearly nonempty). For
each i < log r , choose j with i log r j ; then there is a single xi K
which occurs as the coefficient of t i for every element of B j . The formal sum

xi t i
x=
i< log r

is then an element of K ((t Q )) belonging to the intersection of all the B j .

It is somewhat less obvious that K ((t Q )) is a field. Let x = iQ xi t i and

j
Q
y =
jQ y j t be two elements of K ((t )). We would like to define their
product to be




xy =
xi y j t k .
kQ

i, jQ:i+ j=k

To make this definition sensible, we must check two assertions.

(a) The set of k Q admitting at least one representation as i + j where
xi y j  = 0 is well-ordered.
(b) For each k Q, the number of representations of k as i + j, where
xi y j  = 0, is finite.
We leave both these statements, plus the fact that the resulting ring K ((t Q )) has
multiplicative inverses, as an exercise. (Once the multiplication is known to be
well-defined, such properties as associativity and distributivity over addition
are fairly trivial to check.)

Notes
The concept of a real valuation is a special case of Krulls notion of a valuation
(sometimes called a Krull valuation for emphasis), in which the role of the real
numbers is replaced by an arbitrary totally ordered group. For instance, on the
polynomial ring k[x, y] one can define a degree function taking monomials to
elements of Z Z. If we then equip the latter with the lexicographic ordering
(i.e., we compare pairs in their first component, using the second component

32

Norms on algebraic structures

only to break a tie in the first component), we may then define a valuation
taking each polynomial to the lowest degree of any of its monomials.
While not all concepts from archimedean analysis generalize as nicely to
Krull valuations as to real valuations, Krull valuations are important in algebraic geometry; indeed, they were originally advocated by Zariski as a key
component of a rigorous foundation for the study of algebraic varieties. They
were later ousted from this role by the more flexible theory of schemes, but
they continue to play an important role in the study of birational properties of
varieties and schemes, particularly in questions about the resolution of singularities by blowups. See [187] for a full account of the theory of valuations and
[113] for discussion of some points concerning Zariskis use of valuations in
algebraic geometry.
A sequence m 1 , m 2 , . . . as in the statement of Lemma 1.3.8(a) is called a
Schauder basis for V . For more discussion, see [31, 2.7.2]. Over a complete
archimedean field (i.e., R or C), the statement of Lemma 1.3.8 becomes false,
as there are Banach spaces admitting countable dense subsets (i.e., Banach
spaces which are separable) but not admitting Schauder bases. The first example is due to Enflo [85]. For a general discussion of the problem of constructing
various sorts of Banach-space bases in the archimedean setting, see [159,
Part I] or [37].
For a direct proof of Theorem 1.4.11 in the case of the completed algebraic
closure of Q p , see [191, 3.3.3].
The fact that the completion of an infinite extension of Q p fails to be
spherically complete if it is not discretely valued (see Example 1.5.2 and the
exercises) is a special case of a more general fact, namely, that any nonarchimedean metric space which admits a countable dense subset and whose
metric takes values which are dense in R+ fails to be spherically complete
[193, Theorem 20.5].
The condition of spherical completeness is quite important in nonarchimedean functional analysis, as it is needed for the HahnBanach theorem to
hold. (By contrast, the nonarchimedean version of the open mapping theorem
requires only completeness of the field.) For an expansion of this remark we
recommend [195]; an older reference is [214].
Condition (c) of Proposition 1.5.7 is due to Ostrowski [179]; it is the
definition used by Kaplansky in [118].
Example 1.5.8 was originally due to Hahn [100]. It was later generalized
independently by Malcev and by Neumann to the case where the ordered
abelian group Q is replaced by a possibly nonabelian group. Then it was used
by Poonen [182] to describe the spherical completion of an arbitrary complete
nonarchimedean field, even in the mixed-characteristic case. For instance, one

Exercises

33

obtains a description of a spherical completion of C p in terms of a generalized

power series in p.

Exercises
(1) Prove Proposition 1.1.8. (Hint: first note that |x| < 1 if and only if the
sequence x, x 2 , . . . converges to 0.)
(2) Prove Ostrowskis theorem (Theorem 1.2.4).
(3) Give an example to show that, even for a finite-dimensional vector space
V over a complete nonarchimedean field F, the requirement that a norm
on | |V compatible with F must satisfy the strong triangle inequality is not superfluous; that is, condition (b) in Definition 1.3.1 is not a
consequence of the rest of the definition. (Hint: use a modification of a
supremum norm.)
(4) Let M be a module over a nonarchimedean field F. Prove that any two
norms on M compatible with F are topologically equivalent if and only
if they are metrically equivalent.
(5) Use Remark 1.4.10 to give an example showing that the statement of
Theorem 1.3.6 may fail if F is not complete.
(6) Prove that the valuation ring o F of a nonarchimedean field is noetherian
if and only if F is trivially or discretely valued.
(7) Use Theorem 1.4.9 to prove that, for any field F, any nonarchimedean
norm | | on F, and any extension of E, there exists at least one extension
of | | to a norm on E. (Hint: reduce to the cases where E is a finite
extension or a purely transcendental extension.)
(8) Here is a more exotic variation of the t-adic valuation. Let F be a field,
and choose 1 , . . . , n R.
(a) Prove that on the rational function field F(t1 , . . . , tn ) there is a valuation v such that v( f ) = 0 for all f F and v(ti ) = i for
i = 1, . . . , n. (Hint: you may construct it by iterating the definition
of Gauss valuations.)
(b) Prove that if 1 , . . . , n are linearly independent over Q, the valuation v is uniquely determined by (a).
(c) Prove that if 1 , . . . , n are not linearly independent over Q, the valuation v is not uniquely determined by (a). (Hint: try for a starter
the case n = 2, 1 = 2 = 1.)
(9) Let E be the completion of an infinite extension of Q p which is not discretely valued. Let 1 , 2 , . . . E be any sequence of elements such
that |1 |, |2 |, . . . form a strictly decreasing sequence with positive limit.

34

Norms on algebraic structures

(Since E is not discretely valued, such a sequence must exist.) Prove that
the sequence of discs
{z E : |z 1 i | |i |}

(10)

(11)
(12)
(13)

is decreasing, but its intersection is empty. (Hint: note that if the intersection were nonempty then it would contain an element algebraic over Q p ,
since such elements are dense in E.) Deduce as a corollary that E is not
spherically complete.
Prove that any finite-dimensional vector space over a spherically complete nonarchimedean field equipped with a compatible norm is also
spherically complete. (Hint: use Lemma 1.5.5.)
Prove Proposition 1.5.7.
Prove that a subset S of R is well-ordered (contains no infinite decreasing
sequence) if and only if every nonempty subset of S has a least element.
Check the unproved assertions at the end of Example 1.5.8. (Hint: reduce
both (a) and (b) to the fact that, given any sequence of pairs (i, j)
Q2 for which i + j is nonincreasing, one can pass to a subsequence in
which one component is nonincreasing. Reduce (c) to checking that x
K ((t Q )) has an inverse whenever x 1 has positive t-adic valuation, then
use a geometric series.)

2
Newton polygons

In this chapter, we recall the traditional theory of Newton polygons for polynomials over a nonarchimedean field. In the process, we introduce a general
framework which will allow us to consider Newton polygons in a wider range
of circumstances; it is based on a version of Hensels lemma that applies in not
necessarily commutative rings. As a first application, we fill in a few missing
proofs from Chapter 1.

2.1 Introduction to Newton polygons

We start with the possibly familiar notion of the Newton polygon associated
with a polynomial over a nonarchimedean ring.
Definition 2.1.1. Let R be a ring equipped with a nonarchimedean submul

i
tiplicative (semi)norm | |. For > 0 and P =
i Pi T R[T ], define
the width of P under the -Gauss norm | | as the difference between the
maximum and minimum values of i for which maxi {|Pi | i } is achieved.
Proposition 2.1.2. Let R be a ring equipped with a nonarchimedean multiplicative seminorm | |. For > 0 and P, Q R[T ] the following results
hold.
(a) We have |P Q| = |P| |Q| . That is, | | is multiplicative.
(b) The width of P Q under | | equals the sum of the widths of P and Q
under | | .
Proof. For {P, Q}, let j , k be the minimum and maximum values of i
for which maxi {|i | i } is achieved. Write
35

36

Newton polygons

PQ =

(P Q)i T i =

Pg Q h T i .

g+h=i

In the sum (P Q)i =

g+h=i Pg Q h , each summand has norm at most
|P| |Q| i , with equality if and only if |Pg | = |P| g and |Q h | =
|Q| h . This cannot occur for i < j P + j Q , and for i = j P + j Q it can
only occur for g = j P , h = j Q . Hence
|(P Q)i | < |P| |Q| i

(i < j P + j Q ),

|(P Q)i | = |P| |Q| i

(i = j P + j Q ),

(i > j P + j Q ).

|(P Q)i | |P| |Q| i

(i < k P + k Q ),

|(P Q)i | = |P| |Q|

(i = k P + k Q ),

|(P Q)i | < |P| |Q|

(i > k P + k Q ).

|(P Q)i | |P| |Q|

Similarly, we also have

This proves both claims.

Definition 2.1.3. Let R be a ring equipped with a nonarchimedean submultiplicative seminorm | |. Let v = log | | denote the corresponding valuation.

n
Pi T i R[T ], draw the set of points
Given a polynomial P(T ) = i=0
{(i, v(Pi )) : i = 0, . . . , n; v(Pi ) < +} R2 ,
then form the lower convex hull of these points. (That is, take the intersection of every closed half-plane which contains all the points and lies above
some nonvertical line.) The boundary of this region is called the Newton polygon of P. The slopes of P are the slopes of this open polygon, viewed as
a multiset in which each slope r counts with multiplicity equal to the horizontal width of the segment of the Newton polygon of slope r (or equal to
zero if there is no such segment); the latter can also be interpreted as the
width of P under | |er . (In the case where this multiset has cardinality
less than deg(P), we include + with sufficient multiplicity to make up the
shortfall.)
Example 2.1.4. For R = Q p equipped with the p-adic norm | | p , the
Newton polygon of the polynomial T 3 + pT 2 + pT + p 3 T 3 has vertices
(3, 0), (1, log p), (0, 3 log p). Its slopes are 12 log p with multiplicity 2 and
2 log p with multiplicity 1.

2.1 Introduction to Newton polygons

37

Proposition 2.1.5. Let R be a nonarchimedean ring, and suppose that

P(T ) = (T 1 ) (T n ). Then the slope multiset of P consists of
log |1 |, . . . , log |n |.
Proof. This is immediate from the multiplicativity of | |er .
Here is an explicit example of Proposition 2.1.5 which we will use
repeatedly.
Example 2.1.6. For p a prime and m a positive integer, let m be a primitive
mth root of unity in an algebraic closure of Q p . According to Theorem 1.4.9
(which we will finish proving later in this chapter), Q p (m ) admits a unique
extension of the p-adic norm | | p on Q p . Assuming this, let us calculate the
norm of the element 1 m .
The easiest case is when m is not divisible by p. In this case m 1 is a
root of the polynomial ((T + 1)m 1)/T , which has integer coefficients and
constant term m  0 (mod p). Hence the Newton polygon has all slopes equal
to 0, so |1 m | p = 1.
Suppose next that m = p h for some positive integer h. In this case ph is a
root of the polynomial
h

P(T ) =

Tp 1
T

p h1

p1


T ip

h1

i=0

so ph 1 is a root of Q(T ) = P(T + 1). The image Q(T ) of Q(T ) in F p [T ]

satisfies
h

Q(T ) =

(T + 1) p 1
(T

h1
+ 1) p

T p + 1p 1
T

p h1

h1
+ 1p

= T ( p1) p

h1

so all the coefficients of Q except for its leading coefficient are divisible by p.
Moreover Q(0) = p, so the Newton polygon of Q is a straight line segment
with endpoints (( p 1) p h1 , 0) and (0, log p). We conclude that
|1 ph | p = p p

h+1 /( p1)

(2.1.6.1)

(see the exercises for an alternate derivation of (2.1.6.1)).

Finally, suppose that m = p h j, where j is not divisible by p. If j = 1,
h+1 /( p1)
by (2.1.6.1). Otherwise, we can write
we have |1 m | p = p p
m = ph j and
1 m = (1 ph ) + ph (1 j ).
On the right-hand side, the first term has norm less than 1 whereas the second
term has norm equal to 1. We thus have |1 m | p = 1.

38

Newton polygons

Remark 2.1.7. The data of the Newton polygon is equivalent to the data of
the -Gauss norms for all . This amounts to the statement that the graph of
a convex continuous function is determined by the positions of its supporting
lines of all slopes; this holds because the graph is the lower boundary of the
intersection of the closed halfspaces bounded below by the supporting lines.
For an example of the use of this observation, see the proof of Theorem 11.2.1.

2.2 Slope factorizations and a master factorization theorem

A key property of the operation of completing a nonarchimedean ring, first
noticed by Hensel, is that it vastly enlarges the collection of polynomials over
the ring which can be factored. Here is a sample statement of this form, which
gives us a factorization that separates slopes in the Newton polygon.
Theorem 2.2.1. Let R be a complete nonarchimedean ring. Suppose that S
R[T ], r R, and m Z0 satisfy
vr (S T m ) > vr (T m ).
Then there exists a unique factorization S = P Q satisfying the following
conditions.
(a) The polynomial P R[T ] has degree deg(S) m, and its slopes are
all less than r .
(b) The polynomial Q R[T ] is monic of degree m, and its slopes are all
greater than r .
(c) We have vr (P 1) > 0 and vr (Q T m ) > vr (T m ).
Moreover, for this factorization,
min{vr (P 1), vr (Q T m ) vr (T m )} vr (S T m ) vr (T m ). (2.2.1.1)
(In fact, this statement turns out to be an equality; see the exercises.)
Let us translate this statement into plainer language. The hypothesis vr (S
> vr (T m ) is equivalent to the following two conditions.
(a) The coefficient Sm of T m in S satisfies |Sm 1| < 1; in particular, Sm
is a unit.
(b) The supporting line of slope r of the Newton polygon of S touches the
polygon at the point (m, 0) and nowhere else.
In particular, r does not occur as a slope of the Newton polygon of S. Note that
if (b) holds and Sm is a unit, we can apply the theorem to Sm1 S instead.
It is not difficult to prove Theorem 2.2.1 directly. However, we will be
stating a number of similar results as we go along, with similar proofs. To
T m)

2.2 Slope factorizations and a master factorization theorem

39

save some effort, we state a master factorization theorem applicable to not

necessarily commutative rings, from which we can deduce Theorem 2.2.1 and
all the variants we will use later. The theorem, and the proof given here, are
due to Christol [42, Proposition 1.5.1].
Theorem 2.2.2 (Master factorization theorem). Let R be a nonarchimedean,
not necessarily commutative, ring. Suppose that the nonzero elements a, b, c
R and the additive subgroups U, V, W R satisfy the following conditions.
(a) The spaces U, V are complete under the norm, and U V W .
(b) The map f (u, v) = av + ub is a surjection of U V onto W .
(c) There exists > 0 such that
| f (u, v)| max{|a||v|, |b||u|}

(u U, v V ).

(d) We have ab c W and

|ab c| < 2 |c|.
Then there exists a unique pair (x, y) U V such that
c = (a + x)(b + y),

|x| < |a|,

|y| < |b|.

For this x, y we also have

|x| 1 |ab c||b|1 ,

|y| 1 |ab c||a|1 .

Before proving Theorem 2.2.2, let us see how it implies Theorem 2.2.1.
Proof of Theorem 2.2.1. We apply Theorem 2.2.2 with the following parameters:
R = F[T ],
| | = | |er ,
U = {P F[T ] : deg(P) deg(S) m},
V = {P F[T ] : deg(P) m 1},
W = {P F[T ] : deg(P) deg(S)},
a = 1,
b = T m,
c = S,
=1
and then put P = a + x and Q = b + y.
To see that this works, let us verify explicitly that condition (c) is satisfied
in this setup (the other conditions are more obvious). If u U, v V are such

40

Newton polygons

that max{|a||v|, |b||u|} is achieved by some term of av then that term appears
unchanged in av + ub because ub is divisible by T m . Hence | f (u, v)| |av|
in that case. Otherwise we have |av| < |bu| and so | f (u, v)| = |bu|.
With this motivation in mind, we now prove Theorem 2.2.2.
Proof of Theorem 2.2.2. We define a norm on U V by setting
|(u, v)| = max{|a||v|, |b||u|},
so that (c) implies
|(u, v)| | f (u, v)| |(u, v)|.
In particular 1, so |ab c| < |ab| = |c|.
Since a, b are nonzero, (c) implies that f is injective. By (b), f is in fact a
bijective group homomorphism between U V and W . It follows that, for all
w W,
| f 1 (w)| 1 |w|.
By (d) we may choose (0, ) with |ab c| |c|. Define
B = {(u, v) U V : |(u, v)| |c|}.
For (u, v) B we have
|a||v| |(u, v)| |c| = |a||b|,
so |v| |b|. Similarly |u| |a|. As a result,
| f 1 (c ab uv)| 1 |c ab uv|
1 max{|c ab|, |uv|}
1 max{|c|, 2 |a||b|}
= |c|.
Consequently, the map g(u, v) = f 1 (c ab uv) carries B into itself.
We next show that g is contractive. For (u, v), (t, s) B ,
|g(u, v) g(t, s)| | f 1 (ts uv)|
1 |ts uv|
= 1 |t (s v) + (t u)v|
1 max{|a||s v|, |t u||b|}
= 1 |(u t, v s)|
= 1 |(u, v) (t, s)|,
which has the desired effect because 1 < 1.

2.3 Applications to nonarchimedean field theory

41

Since g is contractive on B and U V is complete, by the Banach contraction mapping theorem (exercise) there is a unique (x, y) U V fixed by g.
That is,
ay + xb = f (x, y) = f (g(x, y)) = c ab x y
and so
c = (a + x)(b + y).
Moreover, there is a unique such (x, y) in the union of all the B , and that
element belongs to the intersection of all the B .
Remark 2.2.3. One can also use Theorem 2.2.2 to recover other instances
of Hensels lemma. For instance, if F is a complete nonarchimedean field,
P(x) o F [x], and the reduction of P(x) into F [x] factors as Q R with Q, R
coprime, then there exists a unique factorization P = Q R in o F [x] with Q, R
lifting Q, R (exercise).

2.3 Applications to nonarchimedean field theory

We now go back and apply Theorem 2.2.1 to prove some facts about extensions of nonarchimedean fields which were omitted from Chapter 1. We first
complete the proof of Theorem 1.4.9; for this we need the following lemma.
Lemma 2.3.1. Let F be a complete nonarchimedean field. Let P(T ) F[T ]
be a polynomial whose slopes are all greater than or equal to r . Let S(T )
F[T ] be any polynomial, and write S = P Q + R with deg(R) < deg(P) using
the division algorithm. Then
vr (S) = min{vr (P) + vr (Q), vr (R)}.
Proof. Exercise.
Proof of Theorem 1.4.9 (continued). It remains to show that if F is a complete
nonarchimedean field then any finite extension E of F admits an extension of
| | to a norm on E. If E  is a field intermediate between F and E, we may
first extend the norm to E  and then to E. Consequently, it suffices to check
the case where E = F() for some E, that is, E
= F[T ]/(P(T )) for
some monic irreducible polynomial P F[T ] (the minimal polynomial of ).
Apply Theorem 2.2.1; since P(T ) cannot factor nontrivially, we deduce that
P must have a single slope r .

42

Newton polygons

We now define a norm on E as follows: for = c0 + c1 + + cn1 n1 ,

with n = deg(P) = [E : F], put
|| E = max{|ci |eri }.
i

er -Gauss

That is, take || E to be the

norm of the polynomial c0 + c1 T +
+ cn1 T n1 . The submultiplicativity of | | E is then a consequence
of Lemma 2.3.1; however, since E is a field, submultiplicativity implies
multiplicativity.
We next give the proof of Theorem 1.4.11. For this, we need a crude version
of the principle that the roots of a polynomial over a complete algebraically
closed nonarchimedean field should vary continuously in the coefficients.
Lemma 2.3.2. Let F be an algebraically closed nonarchimedean field with
completion E, and suppose that P E[T ] is monic of degree d. Then, for any
 > 0, there exists z F such that |z| |P(0)|1/d and |P(z)| < .
Proof. If P(0) = 0 we may pick z = 0, so assume that P(0)  = 0. Put P =

T d + d1 Pi T i . For any > 0, we can pick a polynomial Q = T d +

d1 i=0
i
i=0 Q i d F[T ] with |Q i Pi | < for i = 0, . . . , d 1.
Now assume that < min{|P0 |, , /|P0 |}, so that |Q 0 | = |P0 |. By
Proposition 2.1.5 there exists z F with Q(z) = 0 and |z| |Q 0 |1/d =
|P0 |1/d . We now have
|P(z)| = |(P Q)(z)| max{1, |z|}d max{1, |P(0)|} < ,
as desired.
Proof of Theorem 1.4.11. We must check that the completion E of an algebraically closed nonarchimedean field F is itself algebraically closed. Let
P(T ) E[T ] be a monic polynomial of degree d. Define a sequence of polynomials P0 , P1 , . . . as follows. Put P0 = P. Given Pi , apply Lemma 2.3.2 to
construct z i with |z i | |Pi (0)|1/d and |Pi (z i )| < 2i ; then set Pi+1 (T ) =
Pi (T + z i ), so that Pi+1 (0) = Pi (z i ). If some Pi satisfies Pi (0) = 0 then
z 0 + + z i1 is a root of P. Otherwise, we get an infinite sequence z 0 , z 1 , . . .
such that z 0 + z 1 + converges to a root of P.

Notes
There is a good reason for Newtons name to be attached to the polygons considered in this chapter. He considered them in the context of finding a series
expansion at the origin for a function y = f (x) implicitly defined by a polynomial relation P(x, y) = 0 over C. This was later reinterpreted by Puiseux,

Exercises

43

in terms of computing roots of polynomials over the subfield of C((x)) consisting of series which represent meromorphic functions in a neighborhood
of x = 0.
Remark 2.1.7 is perhaps best viewed within the broader context of duality
for convex functions, which is a central theme in the study of linear optimization problems. One reference for the rigorous mathematical theory of convex
functions is [192].
Although the formulation of Theorem 2.2.2 is due to Christol, the basic
observation that Hensels lemma does not rely on commutativity is rather old.
One early instance is due to Zassenhaus [223].
The Banach contraction mapping theorem (also known as the Banach fixed
point theorem), used in the proof of Theorem 2.2.2, is probably familiar to most
readers except possibly in name. For instance, it appears in the most common
proofs of the implicit function theorem, the inverse function theorem, and the
fundamental theorem of ordinary differential equations.
The notion of a nonarchimedean metric space can be generalized to the case
where the metric takes values in an arbitrary partially ordered set. One can then
give a number of extensions of the Banach contraction mapping theorem, at
least in the case of spherically complete metric spaces. For instance, PriessCrampe has proved that if X is a spherically complete metric space with
distance function d, and f : X X is a function such that d( f (x), f (y)) <
d(x, y) for all x, y X with x  = y, then f has a fixed point. See [197] for an
exposition of a number of results of this type.

Exercises
(1) With notation as in Example 2.1.6, rederive (2.1.6.1) as follows. First note
j
that for j coprime to p, 1 ph and 1 ph are multiples of each other in
j

the ring Z[ ph ], so that |1 ph | p |1 ph | p and vice versa. Then note

j

that the product of 1 ph over those j {0, . . . , p h 1} not divisible by

p equals Q(0) = p.
(2) Prove that equality holds in (2.2.1.1). (Hint: split S T m into (P 1)Q +
(Q T m ).)
(3) Prove the claim in Remark 2.2.3. (Hint: since Q and R are coprime, we
can choose S, T F [x] such that Q S + RT = 1. We can also ensure
that deg(S) < deg(R) and deg(T ) < deg(Q). Use lifts of these to set up
the conditions of Theorem 2.2.2.)
(4) Prove Lemma 2.3.1. (Hint: separate P as a sum P1 + P2 in which vr (P)
is achieved by the leading coefficient of P1 , while vr (P2 ) > vr (P).)

44

Newton polygons

(5) Prove the Banach contraction mapping theorem (used in the proof of
Theorem 2.2.2). That is, suppose that X is a nonempty complete metric
space (not necessarily nonarchimedean) with distance function d, and let
g : X X be a map for which there exists (0, 1) such that
d(g(x), g(y)) d(x, y) for all x, y X . Prove that g has a unique fixed
point. (Hint: show that, for any x X , the sequence x, g(x), g(g(x)), . . .
is Cauchy and that its limit is the desired fixed point.)

3
Ramification theory

Recall (Theorem 1.4.9) that any finite extension of a complete nonarchimedean field carries a unique extension of the norm and is also complete. In
this chapter, we study the relationship between a complete nonarchimedean
field and its finite extensions; this relationship involves the residue fields, value
groups, and Galois groups of the fields in question. We distinguish some important types of extensions, the unramified and tamely ramified extensions. See
[187] (especially Chapter 6) for a more thorough treatment.
We also briefly discuss the special case of a discretely valued field with
a perfect residue field, in which one can say much more. We introduce the
standard ramification filtrations on the Galois groups of extensions of local
fields; these will not reappear again until Part IV, at which point they will relate
to the study of the convergence of solutions of p-adic differential equations
made in Part III. We make no attempt to be thorough in our treatment; instead
we refer the reader to the standard text [198] for more details.
Notation 3.0.1. For E/F a Galois extension of fields, write G E/F for
Gal(E/F). If E = F sep , the separable closure of E, write G F for the absolute
Galois group G F sep /F . Of course, if F has characteristic 0 (or is perfect, e.g.,
if F is a finite field) then F sep coincides with the algebraic closure F alg . (We
avoid the usual notation F for the latter because we prefer to reserve the overbar to denote reduction modulo the maximal ideal of a local ring, e.g., from
Z p to F p .)
Remark 3.0.2. Throughout this chapter, when a nonarchimedean field F is
assumed to be complete it will be sufficient to assume it is only henselian
instead. One of many equivalent formulations of this condition (for which see
[175, 43.2]) is as follows: for any monic polynomial P(x) o F [x] and any
45

46

Ramification theory

simple root r F of P F [x] there exists a unique root r o F of P

lifting r .

3.1 Defect
Let F be a complete nonarchimedean field, and let E be a finite extension
of F. Then the value group |E | contains |F |, while the residue field E may
naturally be viewed as an extension of F . The first fundamental fact about the
extension E/F is that both these containments are finite in appropriate senses
closely related to the degree of E over F.
Lemma 3.1.1. Let F be a complete nonarchimedean field, and let E be a finite
extension of F. Then
[E : F] [ E : F ]#(|E |/|F |),

(3.1.1.1)

with equality at least when F is discretely valued.

Proof. Choose 1 , . . . , m o E lifting a basis of E over F , and choose
1 , . . . , n E so that |1 |, . . . , |n | form a set of coset representatives of
|F | in |E |. Then the i j are linearly independent over F, proving (3.1.1.1).
In the case where F is discretely valued, there exists a unique (0, 1) for
which |F | = Z and |E | = ( 1/n )Z . If we choose j to have norm j1
then it is not hard to show that the i j form a basis of o E over o F (exercise).
This proves the desired equality.
If F is not discretely valued, the situation can be more complicated. One
does however have the following refinement of (3.1.1.1), which we will not
prove here.
Theorem 3.1.2 (Ostrowski). Let F be a complete nonarchimedean field, and
let E be a finite extension of F. Then the quantity
defect(E/F) =

[E : F]
[ E : F ]#(|E |/|F |)

is a positive integer. Moreover, if F has characteristic 0 then defect(E/F) = 1;

otherwise defect(E/F) is a power of the characteristic of F.
Proof. See [187, Theorem 6.2].
Definition 3.1.3. The quantity defect(E/F) in Theorem 3.1.2 is called the
defect of E over F. (Note that some sources instead define the defect as
log p defect(E/F), where p is the characteristic of F.) An extension for
which defect(E/F) = 1 is said to be defectless. For example, any finite
extension of a spherically complete field is defectless (see Remark 3.3.11

3.2 Unramified extensions

47

below). For an example of an extension with nontrivial defect, see [187,

6.3].

3.2 Unramified extensions

The easiest finite extensions of a complete nonarchimedean field to describe
are the unramified extensions, which have a particularly simple effect on the
residue field and the value group.
Definition 3.2.1. Let F be a complete nonarchimedean field. A finite extension
E of F is unramified if E is separable over F and [E : F] = [ E : F ]. This
forces E itself to be separable over F; see Proposition 3.2.3.
Lemma 3.2.2. Let F be a complete nonarchimedean field, and let U be a finite
extension of F. Then, for any subextension E of U over F, U is unramified over
F if and only if U is unramified over E and E is unramified over F.
Proof. Since E sits between U and F , having U separable over F is
equivalent to having both U separable over E and E separable over F .
By Lemma 3.1.1
[U : F] [U : F ],

[U : E] [U : E ],

[E : F] [ E : F ];

since [U : F] = [U : E][E : F] and [U : F ] = [U : E ][ E : F ], the first

of the three inequalities is an equality if and only if the other two are.
What makes unramified extensions so simple to describe is that they are
uniquely determined by their residue field extensions. Here is a precise
statement to this effect.
Proposition 3.2.3. Let F be a complete nonarchimedean field, and let E be a
finite extension of F. Then, for any separable subextension of F over E ,
there exists a unique unramified extension U of F contained in E with U
= ;
moreover, U is separable over F.
Proof. By the primitive element theorem, one can always write
=
F [x]/(P(x)) for some monic irreducible separable polynomial P F [x].
Lift P to a monic polynomial P o F [x]. Choose t o E such that the image
of t in E corresponds to x in F [x]/(P(x)); then the reduction of P(x + t)
into E [x] is divisible by x but not by x 2 . We may thus apply the slope factorization theorem (Theorem 2.2.1) to deduce that P(x + t) has a root in o E . This
proves the existence and separability of U over F.
To prove uniqueness, let U  be another such extension. Then the previous
argument applied to U  in place of E shows that oU  contains a root of P(x +t)

48

Ramification theory

congruent to 0 modulo mU  . However, there can only be one such root in U 

because the Newton polygon of P(x + t) has only one positive slope (since P
is a separable polynomial), so in fact U U  . Again, by comparing degrees
we have U = U  .
Corollary 3.2.4. For each finite separable extension of F , there exists a
unique unramified extension E of F with E
= .
Proof. Choose P(x) as in the proof of Proposition 3.2.3. Then E =
F[x]/(P(x)) is an unramified extension of F with residue field . The proof
of Proposition 3.2.3 shows that any other unramified extension with residue
field must contain E; by once again comparing degrees we see that this
containment must be an equality.
Lemma 3.2.5. Let F be a complete nonarchimedean field, and let E be a
finite extension of F. Let U1 , U2 be unramified subextensions of E over F.
Then the compositum U = U1 U2 is also unramified over F, and U = U1 U2
inside E .
Proof. Put U3 = U1 U2 inside E; by Lemma 3.2.2, U3 is unramified over
F and U1 is unramified over U3 . By Proposition 3.2.3, inside E we have
U1 U2 = U3 . Consequently
[U : U2 ] [U : U2 ]

(by Lemma 3.1.1)

= [U1 : U3 ]
= [U1 : U3 ]

(because U1 is unramified over U3 )

= [U1 U2 : U2 ]
[U : U2 ]

(because U1 U2 U ).

We deduce first that U = U1 U2 and second that [U : U2 ] = [U : U2 ].

Hence U is unramified over U2 , hence also over F by Lemma 3.2.2 again.
Definition 3.2.6. Let F be a complete nonarchimedean field, and let E be a
finite extension of F. By Lemma 3.2.5 there is a maximal unramified subextension U of E over F; by Proposition 3.2.3, U is the maximal separable
subextension of E over F . (We will also say that oU is the maximal unramified subextension of o E over o F .) We say E is totally ramified over F if
U = F.

3.3 Tamely ramified extensions

49

3.3 Tamely ramified extensions

For Galois extensions of complete nonarchimedean fields, we can refine the
ramification theory introduced in the previous section. In the case of a discretely valued base field, this material may be familiar from [198]; we will
continue the review of that case in Section 3.4.
Hypothesis 3.3.1. In this section, let F be a complete nonarchimedean field,
and let E be a finite Galois extension of F.
Definition 3.3.2. The inertia subgroup I E/F of G E/F is the kernel of the map
G E/F Aut( E ); it may be interpreted as G E/U , where U is the maximal
unramified subextension of E over F. In particular, E is unramified over F if
and only if I E/F is trivial, whereas E is totally ramified over F if and only if
I E/F = G E/F .
Definition 3.3.3. Given g I E/F and x E , let g, x denote the image of
g(x)/x in E . For fixed g, this is a homomorphism from E E ; moreover, it is trivial on o
E because g I E/F . We thus obtain a homomorphism
I E/F Hom(|E |, E ); let W E/F denote the kernel of this map, called the
wild inertia subgroup of G E/F . Note that E has no p-torsion, so neither does
Hom(|E |, E ); hence I E/F /W E/F is abelian and of order not divisible by p.
Remark 3.3.4. Some readers may recall that if F is discretely valued then
I E/F /W E/F is cyclic of order not divisible by p. This can be seen as follows.
First, let denote the group of roots of unity in E of order dividing [E : F],
and note that the image of I E/F /W E/F in Hom(|E |, E ) includes only maps
from |E | into . Second, note that is a finite cyclic group. Finally, note
that |F |
= Z implies |E |
= Z. (This last step does not apply if F is not
discretely valued, and indeed I E/F /W E/F need not be cyclic in general.)
Definition 3.3.5. We say E is tamely ramified over F if W E/F is trivial; in
this case, the degree of E over its maximal unramified subextension is not
divisible by p, and we call this the tame degree of E over F. Otherwise, we
say E is wildly ramified over F; if W E/F = G E/F , we say E is totally wildly
ramified over F. Note that if p = 0 then every finite extension of F is tamely
ramified.
The structure of tamely ramified extensions is almost as simple as that
of unramified extensions, by an observation due at this level of generality to Abhyankar (although the special case of discrete F was known long
previously).

50

Ramification theory

Proposition 3.3.6 (Abhyankar). Suppose that E/F is tamely ramified (so

that its inertia group I E/F is abelian of order not divisible by p). Let m be
an integer not divisible by p and annihilating I E/F (e.g., its order). Suppose
that t1 , . . . , th F have images in |F | which generate |F |/|F |m . Then
1/m
1/m
1/m
1/m
E(t1 , . . . , th ) is unramified over F(t1 , . . . , th ).
Proof. Since m is not divisible by p, F(m ) is unramified over F; here m
denotes a primitive mth root of unity. It is thus harmless to assume m F (by
Lemma 3.2.2). We may also assume that E is totally ramified over F.
In this case, by Kummer theory, E is contained in an extension of F
1/m
1/m
of the form F(x1 , . . . , xn ) for some x1 , . . . , xn F . To prove the
1/m 1/m
1/m
claim, it suffices to check that F(x1 , t1 , . . . , th ) is unramified over
1/m
1/m
F(t1 , . . . , th ). Since |t1 |, . . . , |th | generate |F |/|F |m , we can choose
integers 1 , . . . , h and an element z F such that x1 t11 thh z m o
F , and
we can write
1/m

F(x1

1/m

, t1

1/m

, . . . , th

1/m

) = F(t1

1/m

, . . . , th

)((x1 t11 thh z m )1/m ).

It now suffices to check that if m is an integer not divisible by p and y

then F(y 1/m ) is unramified over F. Again, it is safe to replace F by an
unramified extension before checking this, so we may assume that y reduces
to an mth power in F . In this case, we will show that y already has an mth root
in F. Namely, we may now assume that y 1 (mod m F ); then the binomial
series



1/m
1/m
=
(y 1)i
(1 + (y 1))
i
o
F

i=0

converges (since its coefficients are p-adically integral, given that m is not
divisible by p) to an mth root of y in F.
Our next argument may be viewed as a preview of the filtration construction
of the next section.
Proposition 3.3.7. The group W E/F is a p-group.
Proof. We proceed by induction on [E : F]; we may assume that E is totally
wildly ramified over F (so W E/F = G E/F ) and that E  = F. Let v E denote
the valuation on E. Pick any x o E \ o F , and set
j = min{v E (1 g(x)/x) : g G E/F }.
Note that j < + because x
/ F, and j > 0 because E is totally wildly
ramified over F.

3.3 Tamely ramified extensions

51

The map g  g(x)/x from G E/F to o

E is not a homomorphism; however,
it does induce a nonzero homomorphism
G E/F

{y o
E : v E (y 1) j}

{y o
E : v E (y 1) > j}

Since j > 0, the group on the right is isomorphic to the additive group of E
and so is a p-torsion group. Hence G E/F surjects onto a nontrivial p-group;
let E  be the fixed field of the kernel of this surjection. It is clear that E is again
totally wildly ramified over E  , so that G E/E  is also a p-group; hence G E/F
is an extension of two p-groups and is thus a p-group itself.
The following are easily verified using Proposition 3.3.6.
Lemma 3.3.8. Let E  be a subextension of E over F. Then E is tamely ramified over F if and only if E is tamely ramified over E  and E  is tamely ramified
over F.
Lemma 3.3.9. Let F be a complete nonarchimedean field, let E be a finite
extension of F, and let T1 , T2 be tamely ramified subextensions of E over F.
Then T = T1 T2 is also tamely ramified over F.
Remark 3.3.10. Let T be the maximal tamely ramified subextension of E
over F, so that G E/T = W E/F is a p-group by Proposition 3.3.7. A fact from
elementary group theory (exercise) allows us to construct a tower E 0 = T
E 1 E m = E such that each E i is Galois over T , and each group
G Ei /Ei1 is isomorphic to Z/ pZ. This is particularly important if F is of characteristic p, because every Z/ pZ-extension of a field L of characteristic p > 0
is isomorphic to L[z]/(z p z x) for some x L. (Such an extension is called
an ArtinSchreier extension.)
Remark 3.3.11. Using the result of Kaplansky mentioned in the proof of
Theorem 1.5.3 (i.e., the fact that maximal completeness is equivalent to spherical completeness), it is possible to show that any finite extension E of a
spherically complete field F is defectless. It suffices to check this first for E
separable over F and then for E purely inseparable over F (since any E can
be obtained by first making a separable extension and then a purely inseparable extension on top of that). As we go along, keep in mind that any finite
extension of F is also spherically complete (Example 1.5.2).
The key initial observation is that if [E : F] = p then by Theorem 3.1.2
either E is defectless over F or [ E : F ] = [|E | : |F |] = 1. If F is
spherically complete, the latter case is ruled out by Kaplanskys theorem.

52

Ramification theory

We now consider the separable case. Since it suffices to check defectlessness

for an extension containing E (by Theorem 3.1.2), we may replace E by its
Galois closure. Let T be the maximal tamely ramified subextension, so that T
is defectless over F. By Remark 3.3.10 we can form E from T by a sequence
of Z/ pZ-extensions; by the previous paragraph, each of these is defectless.
Hence E is defectless over F.
We next consider the purely inseparable case (which of course only applies
if F has characteristic p). This case is even easier: we can obtain E from F by
a sequence of extensions, each of which simply involves adjoining a pth root;
as above, each such extension is defectless.

3.4 The case of local fields

We now specialize the discussion of ramification theory to local fields, following [198, Chapter IV]. This material will not be used until Chapter 19, where
we will relate ramification theory to the convergence of solutions of p-adic
differential equations.
Hypothesis 3.4.1. In this section, let F be a complete discretely valued nonarchimedean field whose residue field F is perfect. (For more on what happens
when the perfectness hypothesis is lifted, see the notes.) Let E be a finite Galois
extension of F.
Definition 3.4.2. The lower numbering filtration of G E/F is defined as
follows. For i 1 an integer, put
G E/F,i = ker(G E/F Aut(o E /mi+1
E )).
In particular,
G E/F,1 = G E/F ,
G E/F,0 = I E/F ,
G E/F,1 = W E/F .
For i 1 real, we define G E/F,i = G E/F,i . The lower numbering filtration behaves nicely with respect to subgroups of G E/F but not with respect to
quotients; thus it cannot be defined on the absolute Galois group G F .
Definition 3.4.3. The upper numbering filtration of G E/F is defined by the

E/F
relation G E/F

(i)

= G E/F,i , where
 i
E/F (i) =
0

1
dt.
[G E/F,0 : G E/F,t ]

Notes

53

Note that the indices where the filtration jumps are now rational numbers but
not necessarily integers. In any case, Proposition 3.4.4 below implies that there
is a unique filtration G iF on G F which induces the upper numbering filtration
on each G E/F ; that is, G iE/F is the image of G iF under the surjection G F
G E/F . It is this filtration which plays the more important role in, e.g., class
field theory.
Proposition 3.4.4 (Herbrand). Let E  be a Galois subextension of E over F
and put H = G E/E  , so that H is normal in G E/F and G E/F /H = G E  /F .
Then G iE  /F = (G iE/F H )/H ; that is, the upper numbering filtration is
compatible with the formation of quotients of G E/F .
Proof. The proof is elementary but slightly involved, so we will not give it
here. See [198, IV.3].

Notes
Ramification theory originally emerged in the study of algebraic number fields;
the formalism we see nowadays is due largely to Hilbert. The upper and
lower numbering filtrations were originally introduced by Hilbert as part of
class field theory (the study of abelian extensions of number fields and their
Galois groups), but not in the form we see today; the modern definitions were
introduced slightly later by Herbrand. See [198] for more discussion.
Ramification theory for a complete discrete nonarchimedean field becomes
substantially more complicated when one drops the requirement of a perfect
residue field. However, the case of an imperfect residue field is of great interest
in the study of finite covers of schemes of dimension greater than 1. A satisfactory theory for abelian extensions was introduced by Kato [119]. A generalization to nonabelian extensions was later introduced by Abbes and Saito [1, 2].
However, a number of alternate approaches exist, the relationships among
which are not fully understood. These include Borgers theory of residual perfection [29], Kedlayas differential Swan conductor [133], and methods from
higher local class field theory [224, 225]. See also the notes for Chapter 19.
A henselian-valued field is called stable if every finite extension of it is
defectless. A deep theorem of Kuhlmann [152, Theorem 1] (generalizing earlier theorems of Grauert-Remmert and Gruson) states that if F is a stable
henselian field, and E is a henselian extension of F of finite transcendence
degree with transcendence defect equal to 0, then E is also stable. In this
statement, the transcendence defect of E/F is defined as
trdeg(E/F) dimQ (|E |Q /|F |Q ) trdeg( E / F ),

54

Ramification theory

where trdeg denotes transcendence degree. A theorem of Abhyankar [215,

Thorme 9.2] implies that the transcendence defect is always nonnegative;
moreover, when the transcendence defect is 0, the group |E |/|F | and the
extension E / F are both finite. By contrast, both of these can fail when the
transcendence degree is positive.

Exercises
(1) Complete the proof of Lemma 3.1.1. (Hint: construct a sequence of elements, in the span of the purported basis, converging to a given element
of o E .)
(2) Let G be a group of order p n . Prove that there exists a chain of subgroups
G 0 = {e} G 1 G n = G such that each inclusion is proper
and each G i is normal in G. Deduce that, for any finite Galois extension
L/K of fields of characteristic p of pth degree, there exists a sequence
K = K 0 K 1 K n = L of subextensions such that each inclusion
is an extension of degree p and each K i is Galois over K . (A somewhat
easier exercise is to produce such a sequence in which each K i is only
Galois over K i1 .)

4
Matrix analysis

We come now to the subject of metric properties of matrices over a field

complete for a given norm. While this topic is central to our study of differential modules over nonarchimedean fields, it is based on ideas which have their
origins largely outside number theory. We have thus opted to present the main
points first in the archimedean setting and then to repeat the presentation for
nonarchimedean fields.
The main theme is the relationship between the norms of the eigenvalues
of a matrix, which are core invariants but depend on the entries of the matrix
in a somewhat complicated fashion, and some less structured but more readily
visible invariants. The latter are the singular values of a matrix, which play a
key role in numerical linear algebra in controlling the numerical stability of
certain matrix operations (including the extraction of eigenvalues). Their role
in our work is similar.
Before proceeding, we set some basic notation and terminology for matrices.
Notation 4.0.1. Let Diag(1 , . . . , n ) denote the n n diagonal matrix D with
Dii = i for i = 1, . . . , n.
Notation 4.0.2. For A a matrix, let A T denote the transpose of A. For A an
invertible square matrix, let AT denote the inverse transpose of A.
Definition 4.0.3. An n n elementary matrix over a ring R is an n n
matrix obtained from the identity matrix by performing one of the following
operations:
(a) exchanging two rows;
(b) adding c times one row to another row, for some c R;
(c) multiplying one row by some c R .
55

56

Matrix analysis

If B is an elementary matrix, multiplying another n n matrix C on the left

by B effects the corresponding operation on C; such an operation is called
an elementary row operation. Multiplying on the right by B instead effects
an elementary column operation. (It is possible to omit type (a); see the
exercises.)

4.1 Singular values and eigenvalues (archimedean case)

Hypothesis 4.1.1. In this section and the next, let A be an n n matrix over C.
Identify Cn with the space of column vectors equipped with the L 2 norm, i.e.,
|(z 1 , . . . , z n )| = (|z 1 |2 + + |z n |2 )1/2 .
View A as a linear transformation on Cn , and write
| A| =

sup {|Av|/|v|};

vCn \{0}

that is, |A| is the operator norm of A, as will be defined in Definition 6.1.2.
As mentioned above, we are interested in two sets of numerical invariants of
A. One of these is the familiar set of eigenvalues.
Definition 4.1.2. Let 1 , . . . , n be the list of eigenvalues of A, arranged so
that |1 | |n |.
A second set of numerical invariants of A, which is better behaved from the
point of view of numerical analysis, is the set of singular values.
Definition 4.1.3. Let A denote the conjugate transpose (or Hermitian transpose) of A. The matrix A A is Hermitian and nonnegative definite and so has
nonnegative real eigenvalues. The (nonnegative) square roots of these eigenvalues comprise the singular values of A; we denote them 1 , . . . , n with
1 n . These are not invariant under conjugation, but they are invariant
under the multiplication of A on either side by a unitary matrix.
Theorem 4.1.4 (Singular value decomposition). There exist unitary n n
matrices U, V such that U AV = Diag(1 , . . . , n ).
Proof. This is equivalent to showing that there is an orthonormal basis of Cn
which remains orthogonal upon applying A. To construct it, start with a vector
v Cn maximizing | Av|/|v| and then show that, for any w Cn orthogonal
to v, Aw is also orthogonal to Av. For further details, see the references in the
notes.

4.1 Singular values and eigenvalues (archimedean case)

57

Corollary 4.1.5. The singular values of A1 are n1 , . . . , 11 .

From the singular value decomposition, we may infer a convenient interpretation of i .
Corollary 4.1.6. The number i is the smallest value of for which the following holds: for any i-dimensional subspace V of Cn , there exists v V which
is nonzero and such that |Av| |v|.
Proof. Theorem 4.1.4 provides an orthonormal basis v1 , . . . , vn of V such that
Av1 , . . . , Avn is again orthogonal and |Avi | = i |vi | for i = 1, . . . , n. Let
W be the span of vi , . . . , vn ; then, on the one hand, for any i-dimensional
subspace V of Cn , V W is nonzero, and any nonzero v V W satisfies
|Av| i |v|. On the other hand, if we take V to be the span of v1 , . . . , vi , then
we have |Av| i |v| for all v V . This proves the claim.
The relationship between the singular values and the eigenvalues is controlled by the following inequality of Weyl [218].
Theorem 4.1.7 (Weyl). We have
1 i |1 i |

(i = 1, . . . , n),

with equality for i = n.

For the proof, we need a construction that will recur frequently in what
follows.
Definition 4.1.8. Let M be a module over a ring R. The ith exterior power
or wedge power iR M (or i M if there is no ambiguity about R) of M is
the R-module generated by the symbols m 1 m i for m 1 , . . . , m i M,
modulo those relations that force the map (m 1 , . . . , m i )  m 1 m i to
be R-linear in each variable (while the others are held fixed) and alternating.
The latter means that m 1 m i = 0 if m 1 , . . . , m i are not all distinct.
Any element of i M of the form m 1 m i is said to be decomposable;
it is also called the exterior product of m 1 , . . . , m i . The set of decomposable
elements is not closed under addition and consequently will not necessarily
form a basis for i M (see the exercises). However, if M is freely generated by e1 , . . . , en then it is true that the decomposable elements of the form
e j1 e ji with 1 j1 < < ji n do form a basis of i M.
Note also that the exterior power is a functor on the category of R-modules;
in particular, any linear transformation T : M N induces a linear transformation i T : i M i N . In the special case where M = N is free and of

58

Matrix analysis

rank i, i M is a one-dimensional space and i T turns out to be multiplication

by det(T ).
One easily checks that the formula
(m 1 m i , m 1 m j )  m 1 m i m 1 m j
gives a well-defined bilinear map i M j M i+ j M. This map is usually
denoted by , since it corresponds to adding an extra wedge between an i-fold
and a j-fold wedge product.
Lemma 4.1.9. The singular values (resp. eigenvalues) of i A are the i-fold
products of the singular values (resp. eigenvalues) of A.
Proof. Remember that, starting with any basis of Cn , we can obtain a basis of
i A by taking i-fold exterior products of basis elements. In particular, if we
choose U GLn (C), so that U 1 AU is upper triangular with its eigenvalues
on the diagonal, then
i (U 1 AU ) = (i U )1 (i A)(i U )
will be upper triangular with the i-fold products of the eigenvalues on the diagonal. Similarly, if we apply Theorem 4.1.4 to construct unitary matrices U, V
such that U AV is diagonal with its singular values on the diagonal then
i (U AV ) = (i U )(i A)(i V )
will be diagonal with the i-fold products of its singular values on the diagonal.

We now return to Theorem 4.1.7.

Proof of Theorem 4.1.7. The equality for i = n holds because det(A A) =
|det(A)|2 . We check the inequality first for i = 1. Note that 1 = | A| is the
operator norm. Since there exists v Cn \ {0} with Av = 1 v, we deduce that
1 |1 |.
To handle the general case, we consider an exterior power i Cn having the
action of i A. By Lemma 4.1.9, the largest singular value (resp. eigenvalue) of
i A is equal to the product of the i largest singular values (resp. eigenvalues)
of A. Consequently, the previous inequality applied to i A gives exactly the
desired result.
We mention in passing the following converse of Theorem 4.1.7, due to
Horn [114, Theorem 4].

4.1 Singular values and eigenvalues (archimedean case)

59

Theorem 4.1.10. For 1 , . . . , n C and 1 , . . . , n R0 satisfying

1 i |1 i |

(i = 1, . . . , n),

with equality for i = n, there exists an n n matrix A over C with singular

values 1 , . . . , n and eigenvalues 1 , . . . , n .
Equality in Weyls theorem at an intermediate stage has a structural
meaning.
Lemma 4.1.11. Let F be any field. If v1 , . . . , vi , w1 , . . . , wi F n are such
that v1 vi and w1 wi are nonzero and equal, then the F-span
of v1 , . . . , vi equals the F-span of w1 , . . . , wi .
Proof. If v1 vi is nonzero then v1 , . . . , vi must be linearly independent.
We may thus extend v1 , . . . , vi to a basis v1 , . . . , vn of F n .
Consider the bilinear map
: i F n F n i+1 F n .
Suppose that w F n pairs to zero with v1 vi . If we write w = c1 v1 +
+cn vn , we must then have ci+1 = 0 or else the coefficient of v1 vi+1
in v1 vi w will be nonzero. Similarly ci+2 = = cn = 0, so w
belongs to the F-span of v1 , . . . , vi .
Since (v1 vi ) w1 = (w1 wi ) w1 = 0, w1 must
belong to the F-span of v1 , . . . , vi and likewise for w2 , . . . , wi . Consequently,
one of the two spans is contained in the other, and vice versa by the same
argument.
Theorem 4.1.12. Suppose that for some i {1, . . . , n 1} we have
i > i+1 ,

|i | > |i+1 |,

1 i = |1 i |.
Then there exists a unitary matrix U such that U 1 AU is block diagonal, the
first block accounting for the first i singular values and eigenvalues and the
second block accounting for the others.
Proof. Let v1 , . . . , vn be a basis of Cn such that v1 , . . . , vi span the generalized eigenspaces having eigenvalues 1 , . . . , i and vi+1 , . . . , vn span the
generalized eigenspaces having eigenvalues i+1 , . . . , n . Apply the singular
value decomposition (Theorem 4.1.4) to construct an orthonormal basis
w1 , . . . , wn such that Aw1 , . . . , Awn are also orthogonal and | Awi | = i |wi |.
Since i > i+1 , the only nonzero vectors v i Cn for which | Av|/|v|
achieves its maximum value 1 i are the nonzero multiples of w1 wi .

60

Matrix analysis

However, this is also true for v1 vi . By Lemma 4.1.11, w1 , . . . , wi

span V ; this implies that the orthogonal complement of V is spanned by
wi+1 , . . . , wn , and so it is also preserved by A. This yields the desired
result.
Theorem 4.1.13. The following are equivalent.
(a) There exists a unitary matrix U such that U 1 AU is diagonal.
(b) The matrix A is normal, i.e., A A = A A .
(c) The eigenvalues 1 , . . . , n and singular values 1 , . . . , n of A satisfy
|i | = i for i = 1, . . . , n.
Proof. It is clear that (a) implies both (b) and (c), because U , being unitary, is
equivalent to U = U 1 . Given (b), we can perform a joint eigenspace decomposition for the commuting matrices A and A . On any common generalized
eigenspace, A has some eigenvalue , A has eigenvalue , and so A A has
eigenvalue ||2 . This implies (c).
Given (c), Theorem 4.1.12 implies that A can be conjugated by a unitary
matrix into a block diagonal matrix in which each block has a single eigenvalue and a single singular value , such that || = . Let B be such
a block, corresponding to a subspace V of Cn . If = 0 then B = 0.
Otherwise,  = 0 and 1 B is unitary. Hence given orthogonal eigenvectors
v1 , . . . , vi V of B, the orthogonal complement in V of their span is preserved by B, so is either zero or contains another eigenvector vi+1 . This shows
that B is diagonalizable with a single eigenvalue and thus is itself a scalar
matrix. (One can also argue this last step using the compactness of the unitary
group.)
In general, we can conjugate any matrix into an almost normal matrix; the
almost only applies when the matrix is not semisimple.
Lemma 4.1.14. For any > 1, we can choose U GLn (C) such that,
for i = 1, . . . , n, the ith singular value of U 1 AU is at most |i |. If A is
semisimple (i.e., diagonalizable), we can also take = 1.
Proof. Put A in Jordan normal form, then rescale so that, for each eigenvalue
, the superdiagonal terms have absolute value at most (2 1)1/2 || and all
other terms are zero.

4.2 Perturbations (archimedean case)

Another inequality of Weyl [217] shows that the singular values do not change
much under a small (additive) perturbation.

4.2 Perturbations (archimedean case)

61

Theorem 4.2.1 (Weyl). Let B be an n n matrix, and let 1 , . . . , n be the

singular values of A + B. Then
|i i | |B|

(i = 1, . . . , n).

It is more complicated to describe what happens to the eigenvalues under a

small additive perturbation. The best we can do here is to quantify the effect of
an additive perturbation on the characteristic polynomial.

n1
Pi T i and
Theorem 4.2.2. Let B be an n n matrix. Let P(T ) = T n + i=0

n1
Q(T ) = T n + i=0 Q i T i be the characteristic polynomials of A and A + B.
Then
  
i1

 i n 


max{ j , |B|}
(i = 1, . . . , n).
|Pni Q ni | 2
|B|
i 
j=1

 
The superfluous enclosure of the integer 2i ni in absolute value signs is quite
deliberate; it will become relevant in the nonarchimedean setting.
 
Proof. Note that Q ni is the sum of the ni principal i i minors of A + B.
(A minor is the determinant of the i i submatrix obtained by choosing a
set of i rows and i columns. A principal minor is a minor in which the rows
and columns correspond; for instance, if the first row is included, then the
first column must also be included.) By multilinearity of the determinant,
each principal minor can be written as a sum of 2i terms, each of which
is the product of a sign, a k k minor of A, and an (i k) (i k)
minor of B. The terms with k = i sum to Pni itself; the others all have
k < i and so the norm of each is bounded by 1 k |B|ik . This proves the
claim.
We also need to consider multiplicative perturbations. For a considerable
generalization of the following inequality, see Theorem 4.5.2.
Proposition 4.2.3. Let B GLn (C) satisfy |B| . Let 1 , . . . , n be the
singular values of B A. Then
i i

(i = 1, . . . , n).

(The analogous result with B A replaced by AB follows from this, since

transposal does not change singular values.)
Proof. We use the interpretation of singular values given by Corollary 4.1.6.
Choose an i-dimensional subspace V of Cn such that |B Av| i |v| for all
v V . Then choose a nonzero v V such that | Av| i |v|. We have

62

Matrix analysis
i |v| |B Av| |B||Av| i |B||v|,

proving the claim.

This implies that the norms of the eigenvalues can be recovered from the
singular values, provided that we consider not just the matrix A but also its
powers.
Proposition 4.2.4. Let k,1 , . . . , k,n be the singular values of Ak . Then
1/k
lim
k k,i

= |i |

(i = 1, . . . , n).

Proof. Pick > 1, and choose U as in Lemma 4.1.14; that is, U 1 AU is

upper triangular and each block of eigenvalue differs from the scalar matrix
Diag(, . . . , ) by a matrix of norm at most (2 1)1/2 ||. In a block with
eigenvalue , the singular values of the kth power are bounded below by
||k and above by k ||k . Consequently, we may apply Proposition 4.2.3 to
deduce that
|i |k |U |1 |U 1 |1 k,i k |i |k |U ||U 1 |.
Taking kth roots and then taking k , we deduce
1/k

|i | lim inf k,i ,

k

1/k

lim sup k,i |i |.

k

Since > 1 was arbitrary, we obtain the desired result.

Remark 4.2.5. The case i = 1 of Proposition 4.2.4, and the corresponding
case for its nonarchimedean analogue (Proposition 4.4.10), are instances of
a general fact about the spectra of bounded operators on Banach spaces. See
Remark 6.1.7.

4.3 Singular values and eigenvalues (nonarchimedean case)

We now pass to nonarchimedean analogues.
Hypothesis 4.3.1. Throughout this section and the next, let F be a complete
nonarchimedean field, and let A be an n n matrix over F. View A as a linear
transformation on F n , equip F n with the supremum norm
|(z 1 , . . . , z n )| = max{|z 1 |, . . . , |z n |},
and again define the operator norm
|A| =

sup {|Av|/|v|}.

vF n \{0}

4.3 Singular values and eigenvalues (nonarchimedean case)

63

Note, however, that we also have the simpler expression

|A| = max{|Ai j |}.
i, j

Definition 4.3.2. Given a sequence s1 , . . . , sn , we define the associated

polygon for this sequence to be the polygonal line joining the points
(n + i, s1 + + si )

(i = 0, . . . , n).

This polygon is the graph of a convex function on [n, 0] if and only if

s1 sn .
Definition 4.3.3. Let s1 , . . . , sn be the sequence with the property that, for
i = 1, . . . , n, s1 + + si is the minimum valuation of an i i minor
of A; that is, the si are the elementary divisors (or invariant factors) of A.
The associated polygon is called the Hodge polygon of A (see the notes
for an explanation of the terminology). Define the singular values of A as
1 , . . . , n = es1 , . . . , esn ; these are invariant under multiplication on either
side by a matrix in GLn (o F ). Note that, as in the nonarchimedean case,
1 = | A|.
We also have an analogue of the singular value decomposition.
Theorem 4.3.4 (Smith normal form). There exist U, V GLn (o F ) such that
U AV is a diagonal matrix whose entries have norms 1 , . . . , n .
Proof. It is equivalent to prove that, starting with A, one can perform elementary row and column operations defined over o F so as to produce a diagonal
matrix (this amounts to a limited GaussJordan elimination over A). To do
this, find the largest entry of A, permute rows and columns to put this entry at
the top left, and then use it to clear the remainder of the first row and column.
Repeat with the matrix obtained by removing the first row and column, and
so on.
Corollary 4.3.5. The slopes s1 , . . . , sn of the Hodge polygon satisfy
s1 sn .
Proof. The ith slope si is evidently the ith smallest valuation of a diagonal
entry of the Smith normal form.
We again have a characterization as in Corollary 4.1.6.
Corollary 4.3.6. The number i is the smallest value of for which the following holds: for any i-dimensional subspace V of F n , there exists v V
nonzero and such that |Av| |v|.

64

Matrix analysis

Definition 4.3.7. Let 1 , . . . , n be the eigenvalues of A in some algebraic

extension of F equipped with the unique extension of | |, arranged so that
|1 | |n |. The associated polygon is the Newton polygon of A; this is
invariant under conjugation by any element of GLn (F).
The nonarchimedean analogue of Weyls inequality is the following.
Theorem 4.3.8 (Newton polygon above Hodge polygon). We have
1 i |1 i |

(i = 1, . . . , n),

with equality for i = n. In other words, the Hodge and Newton polygons have
the same endpoints and the Newton polygon is everywhere on or above the
Hodge polygon.
Proof. Again, the case i = 1 is clear because 1 is the operator norm of A,
and the general case follows by considering exterior powers (using the obvious
analogue of Lemma 4.1.9).
Like its archimedean analogue, Theorem 4.3.8 also has a converse, but in
this case we can write the construction down quite explicitly.

n1
Pi T i a monic polynomial of degree n
Definition 4.3.9. For P = T n + i=0
over a ring R, the companion matrix of P is defined as the matrix

0 0
P0
1 0
P1

. .
..
..
..
.
0 1 Pn1

with 1s on the subdiagonal, the negated coefficients of P in the far right-hand

column, and 0s elsewhere. The companion matrix is constructed to have its
characteristic polynomial equal to P (exercise).
Proposition 4.3.10. Choose 1 , . . . , n F alg such that |1 | |n |,

n1
Pi T i
and such that the polynomial P(T ) = (T 1 ) (T n ) = T n + i=0
has coefficients in F. Choose c1 , . . . , cn F with i = |ci | such that 1
n and
1 i |1 i |

(i = 1, . . . , n),

4.3 Singular values and eigenvalues (nonarchimedean case)

65

with equality for i = n. Then the matrix

1
0
0 c11 cn1
P0
cn1 0 c1 c1 P1
1
n2

..
..
..

.
.
0

Pn1

c1

has singular values 1 , . . . , n and eigenvalues 1 , . . . , n .

Proof. The given matrix is conjugate to the companion matrix of P, so its
eigenvalues are also 1 , . . . , n . To compute the singular values, we note that,
for i = 1, . . . , n 1,



 1
1
1
Pi  = 11 ni1
|Pi |
c1 cni1
1
11 ni1
|1 ni |

ni .
Thus we can perform column operations over o F to clear everything
1
P0 , which has norm
in the far right-hand column except c11 cn1
1
1
1 n1 |1 n | = n . By permuting the rows and columns we obtain
a diagonal matrix with entries that are the norms 1 , . . . , n . This proves the
claim.
Again equality has a structural meaning, but the proof requires a bit
more work than in the archimedean case since we no longer have access to
orthogonality. However, this extra work is rewarded by a slightly stronger
result.
Theorem 4.3.11 (HodgeNewton decomposition). Suppose that for some i
{1, . . . , n 1} we have
|i | > |i+1 |,

1 i = |1 i |.

(That is, the Newton polygon has a vertex with x-coordinate n + i and this
vertex also lies on the Hodge polygon.) Then there exists U GLn (o F ) such
that U 1 AU is block upper triangular, with the top left block accounting for
the first i singular values and eigenvalues and the bottom right block accounting for the others. Moreover, if i > i+1 then we can ensure that U 1 AU is
block diagonal.
Proof. We first note that, by Theorem 2.2.1 applied to the characteristic polynomial of A, P(T ) = (T 1 ) (T i ) and Q(T ) = (T i+1 ) (T n )
have coefficients in F. Since P and Q have no common roots, we can write

66

Matrix analysis

1 = P B + QC for some B, C F[T ], and then the products P(A)B(A) and

Q(A)C(A) give the projectors for a direct sum decomposition separating the
first i generalized eigenspaces from the others.
In other words, we can find a basis v1 , . . . , vn of F n such that v1 , . . . , vi
span the generalized eigenspaces with eigenvalues 1 , . . . , i and vi+1 , . . . , vn
span the generalized eigenspaces with eigenvalues i+1 , . . . , n . Choose a
basis w1 , . . . , wn of onF such that w1 , . . . , wi is a basis of onF (Fv1 + +
Fvi ). Let e1 , . . . , en be the standard basis of F n , and define U GLn (o F ) by

w j = i Ui j ei . Then


B C
1
U AU =
0 D
is block upper triangular. By Cramers rule, each entry of B 1 C is an i i
minor of A divided by the determinant of B. Since |det(B)| = 1 i , B 1 C
must thus have entries in o F . Writing



B 0
Ii B 1 C
1
U AU =
,
0 D
0
Ini
we see that the singular values of B and D together must comprise 1 , . . . , n .
The only way for this to happen, given the constraint that the product of the
singular values of B equals 1 i , is for B to account for 1 , . . . , i and for
D to account for i+1 , . . . , n .
This proves the first claim; we may thus assume now that i > i+1 . In that
case, conjugating by the matrix


Ii B 1 C
0
Ini
gives a new matrix,

B
0

C1
,
D

with C1 = B 1 C D. Since
|C1 | |B 1 ||C||D| = i1 |C|i+1 < |C|,
this process converges. More explicitly, we obtain a sequence of matrices Ui
GLn (o F ) converging to the identity, such that the convergent product U =
U1 U2 satisfies


B 0
,
U 1 AU =
0 D
as desired.

4.3 Singular values and eigenvalues (nonarchimedean case)

67

Note that the slopes of the Hodge polygon are forced to be in the additive
value group of F, whereas the slopes of the Newton polygon need only lie in
the divisible closure of the additive value group. Consequently, it is possible for
a matrix to have no conjugates over GLn (F) for which the Hodge and Newton
polygons coincide. However, the following is true; see also Corollary 4.4.8
below.
Lemma 4.3.12. Suppose that one of the following holds.
(a) The value group of |F | is dense in R>0 , and > 1.
(b) We have |i | |F | for i = 1, . . . , n (so in particular i  = 0), and
1.
Then there exists U GLn (F) such that the ith singular value of U 1 AU is
at most |i |.
Proof. Case (a) will follow from Corollary 4.4.8 below. Case (b) is directly
analogous to Lemma 4.1.14.
One also has the following variant.
Lemma 4.3.13. Suppose that |F | is discrete. Then there exists U GLn (F)
such that, for each positive integer m, |U 1 Am U | is the least element of |F |
greater than or equal to |m
1 |.
Proof. We may normalize the valuation on F for convenience so that
log |F | = Z. As in the proof of Theorem 4.3.11, we may also reduce to
the case where all the eigenvalues of A have the same norm.
Let E be a finite extension of F containing an element with || = |1 |.
(For instance, we could take E = F(1 ) and = 1 , but any other choice
would also work.) By Lemma 4.3.12, there exists U0 GLn (E) such that
1 U01 AU0 GLn (o E ); in other words, there exists a supremum norm | |0
on E n such that |1 Av|0 = |v|0 for all v E n .
Put V = {v F n : |v|0 1}. Let v1 , . . . , vn be a basis of V over o F , and
let | |1 be the supremum norm on F n defined by v1 , . . . , vn . Given a nonzero
v F n , choose F with log |v|0 [0, 1); then v is an element of
V which is not divisible by m F , so |v|1 = 1. We conclude that
e1 |v|1 < |v|0 |v|1

(v F n ).

(4.3.13.1)

Let e1 , . . . , en be the standard basis of F n , and define U GLn (F) by

v j = i Ui j ei . Then, for each positive integer m, log |U 1 Am U | Z is at

most m log |1 | by Theorem 4.3.8 but is strictly greater than m log |1 | 1
by (4.3.13.1). We thus obtain the desired equality.

68

Matrix analysis

4.4 Perturbations (nonarchimedean case)

Again, we can ask about the effect of perturbations. The analogue of Weyls
second inequality is more or less trivial.
Proposition 4.4.1. If B is a matrix with |B| < i then the first i singular
values of A + B are 1 , . . . , i .
Proof. Exercise.
We next consider the effect on the characteristic polynomial.

n1
Theorem 4.4.2. Let B be an n n matrix. Let P(T ) = T n + i=0
Pi T i and

n1
Q(T ) = T n + i=0 Q i T i be the characteristic polynomials of A and A + B.
Then
|Pni Q ni | |B|

i1


max{ j , |B|}

(i = 1, . . . , n).

j=1

Proof. The proof is as for Theorem 4.2.2, except that now the factor |2i
dominated by 1.

n 
i

| is

Question 4.4.3. Is the inequality in Theorem 4.4.2 the best possible?

We may also consider multiplicative perturbations.
Proposition 4.4.4. Let B GLn (F) satisfy |B| . Let 1 , . . . , n be the
singular values of AB. Then
i i

(i = 1, . . . , n).

Proof. As for Proposition 4.2.3 but using the Smith normal form
(Theorem 4.3.4) instead of the singular value decomposition.
Corollary 4.4.5. Suppose that the Newton and Hodge slopes of A coincide
and that U GLn (F) satisfies |U | |U 1 | . Then each Newton slope of
U 1 AU differs by at most log from the corresponding Hodge slope.
Here is a weak converse to Corollary 4.4.5. (We leave the archimedean
analogue to the readers imagination.)
Proposition 4.4.6. Suppose that the Newton slopes of A are nonnegative and
that 1 1. Then there exists U GLn (F) such that
|U 1 AU | 1,

|U 1 | 1,

|U | 1n1 .

Proof. Let e1 , . . . , en denote the standard basis vectors of F n . Let M be the

smallest o F -submodule of F n containing e1 , . . . , en and stable under A. For

4.4 Perturbations (nonarchimedean case)

69

each i, if j = j (i) is the least integer such that ei , Aei , . . . , A j ei are lin
j1
early dependent then we have A j ei = h=0 ch Ah ei for some ch F. The

j1
polynomial T j h=0 ch T h has roots which are eigenvalues of A, so the nonnegativity of the Newton slopes forces |ch | 1. Hence M is finitely generated,
and thus free, over o F .
Let v1 , . . . , vn be a basis of M, and let U be the change-of-basis matrix

v j = i Ui j ei ; then |U 1 AU | 1 because M is stable under A and |U 1 |

1 because M contains e1 , . . . , en . The desired bound on |U | follows from the
fact that, for any x = c1 e1 + + cn en M, we have
max{|ci |} 1n1 .
i

(4.4.6.1)

It suffices to check (4.4.6.1) for x = Ah ei for i = 1, . . . , n and h =

0, . . . , j (i) 1, as these generate M over o F . But it is evident that | Ah ei |
1h |ei | = 1h ; since j (i) n, we are done.
Example 4.4.7. The example

1
A = 0
0

c 0
1 c
0 1

with |c| > 1 shows that the bound of Proposition 4.4.6 is sharp; in particular,
the bound |U | 1n1 cannot be improved to |U | 1 , as one might initially have expected. However, one should be able to get a more precise bound
(which agrees with the bound given in this example) by accounting for the
other singular values; see the exercises.
Corollary 4.4.8. There exists a continuous function
f n (1 , . . . , n , 1 , . . . , n , ) : (0, +)2n [0, +) (0, +)
(independent of F) with the following properties.
(a) Suppose that, for each i = 1, . . . , n, either i = i or max{i , i }.
Then
f n (1 , . . . , n , 1 , . . . , n , ) = 1.
(b) If A has singular values 1 , . . . , n and eigenvalues 1 , . . . , n , if i =
|i | for i = 1, . . . , n, and if i |F | whenever i > then there
exists U GLn (F) such that
|U 1 | 1,

|U | f n (1 , . . . , n , 1 , . . . , n , ),

for which the multiset of singular values of U 1 AU matches

1 , . . . , n in its values greater than .

70

Matrix analysis

Proof. This follows by induction on n as follows. If 1 , we can deduce

the whole claim by Proposition 4.4.6 (after rescaling in the case  = 1). Otherwise, again by Proposition 4.4.6 (and again after rescaling), we can find U1
GLn (F) such that
|U11 A1 U1 | 1 ,

|U11 | 1,

|U1 | (1 /1 )n1 .

Let i be the largest index such that i = 1 . Then the first i singular values of U11 AU1 are all at most 1 but at least i . Hence they are all equal,
and A1 = U11 AU1 satisfies the hypothesis of Theorem 4.3.11. We may thus
choose U2 GLn (o F ) such that A2 = U21 A1 U2 is block upper triangular, the
top left block accounting for the first i singular values and eigenvalues and the
bottom right block accounting for the others.
If i = n then we may take U = U1 U2 and be done. Otherwise, note that by
applying Proposition 4.4.4 we may bound the singular values of A2 by a continuous function of 1 , . . . , n , 1 , . . . , n , . We may then apply the induction
hypothesis to construct a block diagonal matrix U3 , where the top left block
of U3 is the identity, |U31 | 1, |U3 | is bounded by a continuous function
of 1 , . . . , n , 1 , . . . , n , , and the multiset of singular values of the bottom
 , . . . ,  in its values greater
right block of A3 = U31 A2 U3 agrees with i+1
n
than .
We may bound the norm of the top right block of A3 by a continuous function of 1 , . . . , n , 1 , . . . , n , . We can then conjugate by a suitable block
diagonal matrix U4 , with scalar matrices in the diagonal blocks, to ensure that
the multiset of singular values of A4 = U41 A3 U4 agrees with 1 , . . . , n in
its values greater than . We then take U = U1 U4 .
For the purposes of this book, it is immaterial what the function f n is, as
long as it is continuous. However, for numerical applications it may be quite
helpful to identify a good function f n ; here is a conjectural best possible result
in the case = 0, phrased in a somewhat stronger form. (One can formulate
an archimedean analogue. It should also be possible to prove, using the Horn
inequalities, that this conjecture cannot be improved; see the next section.)
Conjecture 4.4.9. If A has singular values 1 , . . . , n and eigenvalues
1 , . . . , n , none of which is equal to 0, and if 1 , . . . , n |F | satisfy
1 i 1 i |1 i |

(i = 1, . . . , n),

then there exists U GLn (F) such that

|U 1 | 1,

|U | max{(1 i )/(1 i )},

i

for which U 1 AU has singular values 1 , . . . , n .

4.5 Horns inequalities

71

By imitating the proof of Proposition 4.2.4 (after enlarging F to contain the

eigenvalues of A), we obtain the following.
Proposition 4.4.10. Let k,1 , . . . , k,n be the singular values of Ak . Then
1/k
lim
k k,i

= |i |

(i = 1, . . . , n).

4.5 Horns inequalities

Although they will not be needed in this book, it is quite natural to mention
here some stronger versions of the perturbation inequalities in the archimedean and nonarchimedean cases, introduced conjecturally by Horn [115] in
the archimedean case and resolved in the work of Klyachko, Knutson, Speyer,
Tao, Woodward, and others. See the beautiful survey article of Fulton [95] for
more information.
Definition 4.5.1. To introduce these stronger inequalities, we must set up some
notation. Put

Urn = (I, J, K ) : I, J, K {1, . . . , n}, #I = #J = #K = r,




r (r + 1)
i+
j=
k+
.

2
iI

jJ

kK

For (I, J, K ) Urn , write I = {i 1 < < ir } and similarly for J, K . For
r = 1, put T1n = U1n . For r > 1, put

Trn = (I, J, K ) Urn : for all p < r and (F, G, H ) T pr ,




p( p + 1)
if +
jg
kh +
.

2
f F

gG

hH

For multiplicative perturbations, we obtain the following results, which

include Propositions 4.2.3 and 4.4.4. It is important for the proofs that one can
rephrase the Horn inequalities in terms of LittlewoodRichardson numbers;
see [95, 3].

72

Matrix analysis

Theorem 4.5.2. For {A, B, C}, let ,1 , . . . , ,n be a nonincreasing

sequence of nonnegative real numbers. Then the following are equivalent.
(a) There exist n n matrices A, B, C over C with AB = C such that, for
{A, B, C}, has singular values ,1 , . . . , ,n .

n

A,i nj=1 B, j = nk=1 C,k and, for all r < n and
(b) We have i=1
(I, J, K ) Trn ,



C,k
A,i
B, j .
kK

iI

jJ

Proof. See [95, Theorem 16]. Note that the first condition in (b) is omitted in
the statement given in [95], but this was a typographical error.
Theorem 4.5.3. Let F be a complete nonarchimedean field. For
{ A, B, C}, let ,1 , . . . , ,n be a nonincreasing sequence of elements of |F|.
Then the following are equivalent.
(a) There exist n n matrices A, B, C over F with AB = C such that, for
{A, B, C}, has singular values ,1 , . . . , ,n .

n

A,i nj=1 B, j = nk=1 C,k and, for all r < n and
(b) We have i=1
(I, J, K ) Trn ,



C,k
A,i
B, j .
kK

iI

jJ

Proof. See [95, Theorem 7].

Example 4.5.4. Let us see explicitly how Theorem 4.5.2 implies
Proposition 4.2.3. Since T1n = U1n , condition (b) of Theorem 4.5.2 includes
all cases with (I, J, K ) U1n . In particular, we may take
I = {i},

J = {1},

K = {i}

to obtain the inequality C,i A,i B,1 ; this is precisely Proposition 4.2.3.
Remark 4.5.5. For additive perturbations, one has an analogous result in the
archimedean case; see [95, Theorem 15]. We are not aware of an additive result
in the nonarchimedean case. Also, in the archimedean case one has analogous
results (with slightly different statements) in which one restricts to Hermitian
matrices.

Notes
The subject of archimedean matrix inequalities is an old one, with many important applications. A good reference for this is [28]; for instance, see [28,
I.2] for the singular value decomposition, [28, Theorem II.3.6] for the Weyl

Notes

73

inequalities, in a much stronger form known as Weyls majorant theorem, [28,

Theorem III.4.5] for a strong form of Proposition 4.2.3 (also a consequence of
the Horn inequalities), and so on. (A variant of our Theorem 4.2.2 appears as
[28, Problem I.6.11].)
The strong analogy between archimedean and p-adic matrix inequalities
appears to be a little-known piece of folklore. As a result, we have been unable
to locate a suitable reference.
It should be pointed out that most of what we have done here is the special
case for GLn of a more general theory encompassing the other reductive algebraic groups. This point of view can be seen in [95], where GLn makes some
explicit appearances for which other groups can be substituted.
In Theorem 4.1.13, the equivalence of (a) and (b) is standard. We do not
have a reference for the equivalence with (c), although it is implicit in most
proofs of the equivalence of (a) and (b).
The reader familiar with the notions of elementary divisors or invariant factors may be wondering why the terminology Hodge polygon is necessary
or reasonable. The answer is that the Hodge numbers of a variety over a padic field are reflected by the elementary divisors of the action of Frobenius
groups on crystalline cohomology. The fact that the Newton polygon lies
above the Hodge polygon then implies a relation between the characteristic
polynomial of Frobenius and the Hodge numbers of the original variety; this
relationship was originally conjectured by Katz and proved by Mazur. See
[27] for further discussion of this point and of crystalline cohomology as a
whole.
Much of the work in this chapter can be carried over to the case of a transformation which is only semilinear for some isometric endomorphism of F. We
will adopt that point of view in Chapter 14; for instance, this will lead to a generalization (in Theorem 14.5.5) of the HodgeNewton decomposition theorem
(Theorem 4.3.11). In this case, the carrying over is really in the other direction: it is the latter result (due to Katz; see the notes for Chapter 14) which
inspired our presentation of Theorem 4.3.11 and its archimedean analogue
(Theorem 4.1.12). Similarly, our treatment of Proposition 4.4.10 and its archimedean analogue (Proposition 4.2.4) are modeled on [120, Corollary 1.4.4].
The question of quantifying the sensitivity to perturbation of the characteristic polynomial of a square matrix arises in numerical applications. The
question is familiar in the archimedean case but perhaps less so in the nonarchimedean case; numerical applications of the latter include using p-adic
cohomology to compute zeta functions of varieties over finite fields. See for
instance [3, 1.6] and [96, 3].

74

Matrix analysis

Exercises
(1) Prove that any elementary matrix of type (a) (swapping two rows) can
be factored as a product of elementary matrices of types (b) and (c); see
Definition 4.0.3.
(2) Let e1 , . . . , e4 be a basis of C4 . Prove that in 2 C4 the element e1 e2 +
e3 e4 is not decomposable.
(3) Check that the characteristic polynomial of the companion matrix of a
polynomial P (Definition 4.3.9) is equal to P.
(4) Prove Proposition 4.4.1. (Hint: use Corollary 4.3.6.)
(5) With notation as in Theorem 4.3.11, suppose that U, V GLn (o F ) are
congruent to the identity matrix modulo m F . Prove that the product of
the i largest eigenvalues of U AV again has norm |1 i |. (Hint: use
exterior powers to reduce to the case i = 1.) This yields as a corollary
[34, Lemma 5]: if D GLn (F) is diagonal and U, V GLn (o F ) are
congruent to the identity matrix modulo m F then the Newton polygons of
D and U DV coincide.
(6) State and prove an archimedean analogue of the previous problem.
(7) Prove the following improved version of Proposition 4.4.6. Suppose that
the Newton slopes of A are nonnegative. Then there exists U GLn (F)
such that
|U 1 AU | 1,

|U 1 | 1,

|U |

n1


max{1, i }.

i=1

We do not know of an appropriate archimedean analogue.

Part II
Differential Algebra

5
Formalism of differential algebra

In this chapter, we introduce some basic formalism of differential algebra.

This may be viewed as a mild perturbation of commutative algebra in which
we consider commutative rings equipped with the additional noncommutative
structure of a derivation. This allows us to manipulate differential equations
and differential systems in a manner that keeps the more useful structure visible, though we will need to convert back and forth from this point of view. (One
thing we will not do is generalize further, to the realm of differential schemes
and sheaves; we leave this as a thought exercise for the curious reader.)
A particularly important result that we introduce is the cyclic vector theorem,
which gives a compact but highly noncanonical way to represent a finite differential module over a field. While the cyclic vector theorem will prove
indispensable at a few key points in our treatment of p-adic differential
equations, we will ultimately make more progress by limiting its use. See
Remark 5.7.1 for further discussion.

5.1 Differential rings and differential modules

Definition 5.1.1. A differential ring is a commutative ring R equipped with a
derivation d : R R; the latter is an additive map satisfying the Leibniz rule
d(ab) = ad(b) + bd(a)

(a, b R).

We expressly allow d = 0 unless otherwise specified; this will be useful in

some situations. A differential ring which is also a domain or field, etc., will
be called a differential domain, field, etc. Note that there is a unique extension
of d to any localization of R, using the quotient rule; in particular, if R is a
domain then there is a unique extension of d to Frac(R).
77

78

Formalism of differential algebra

Definition 5.1.2. A differential module over a differential ring (R, d) is a

module M equipped with an additive map D : M M satisfying
D(am) = a D(m) + d(a)m;
such a D will also be called a differential operator on M relative to d. For
example, (R, d) is a differential module over itself; any differential module
isomorphic to a direct sum of copies of (R, d) is said to be trivial. (If we
refer to the trivial differential module, though, we mean (R, d) itself.) A
differential module which is a successive extension of trivial modules is said
to be unipotent (see Proposition 7.2.5 for the reason why). A differential ideal
of R is a differential submodule of R itself, i.e., an ideal stable under d.
Definition 5.1.3. For (M, D) a differential module, define
H 0 (M) = ker(D),

H 1 (M) = coker(D) = M/D(M).

The latter computes Yoneda extensions; see Lemma 5.3.3 below. Elements of
H 0 (M) are said to be horizontal (see the notes). Note that H 0 (R) = ker(d)
is a subring of R; if R is a field then ker(d) is a subfield. We call them the
constant subring and constant subfield of R.
We now make an observation about base changes of the constant subring.
For the definition of a more general base change, see Definition 5.3.2. (See
also Proposition 6.9.1.)
Lemma 5.1.4. Let R0 be the constant subring of the differential ring (R, d),
and let R0 be an R0 -algebra. Let (M, D) be a differential module over (R, d).
View R  = R R0 R0 as a differential ring by defining the derivation d  by





ai ri =
d(ai ) ri
(ai R, ri R0 ).
d
i

Similarly, view M  = M R0 R0 as a differential module over (R  , d  ) by

defining the differential operator





D
m i ri =
D(m i ) ri
(m i M, ri R0 ).
i

(a) There are natural maps H i (M) R0 R0 H i (M  ) for i = 0, 1.

(b) The map in (a) is always an isomorphism for i = 1.
(c) If R0 is flat over R0 then the map in (a) is an isomorphism for i = 0.

5.1 Differential rings and differential modules

79

Proof
(a) Tensoring the structure morphism R0 R0 with M induces a map
M M  . This in turn induces maps H i (M) H i (M  ) of R0 modules; using the R  -module structure on H i (M  ), we also obtain
maps H i (M) R0 R0 H i (M  ).
(b) Tensoring with R0 is always a right exact functor on R0 -modules.
Since
D

M M H 1 (M) 0
is an exact sequence in the category of R0 -modules,
D

M  M  H 1 (M) R0 R0 0
is also exact. Hence the induced map H 1 (M) R0 R0 H 1 (M  ) is an
isomorphism.
(c) Recall that, by definition, R0 is flat over R0 if and only if tensoring with
R0 is an exact functor on R0 -modules. Since
D

0 H 0 (M) M M
is an exact sequence in the category of R0 -modules,
D

0 H 0 (M) R0 R0 M  M 
is also exact. Hence the induced map H 0 (M) R0 R0 H 0 (M  ) is an
isomorphism.
Another frequently used observation is the following.
Lemma 5.1.5. Let (R, d) be a differential field with constant subfield R0 .
Then, for any differential module (M, D) over (R, d), the natural map
H 0 (M) R0 R M is injective. In particular, dim R0 H 0 (M) dim R M.
Proof. An equivalent statement is that if m 1 , . . . , m n H 0 (M) are linearly
dependent over R then they are also linearly dependent over R0 . Suppose on
the contrary that, for some positive integer n, there exist m 1 , . . . , m n H 0 (M)
which are linearly dependent over R but linearly independent over R0 . Choose
n as small as possible with this property; then there exist c1 , . . . , cn R all
nonzero such that c1 m 1 + +cn m n = 0. We may rescale the ci so that c1 = 1.
Since m 1 , . . . , m n H 0 (M), we also have
d(c1 )m 1 + + d(cn )m n = 0.

80

Formalism of differential algebra

That is, d(c1 ), . . . , d(cn ) form another linear dependence relation between
m 1 , . . . , m n . But d(c1 ) = d(1) = 0, so to avoid contradicting the choice of
n we must have d(c2 ) = = d(cn ) = 0. That is, c1 , . . . , cn R0 , contradicting the hypothesis that m 1 , . . . , m n are linearly independent over R0 . This
contradiction proves the claim.
Definition 5.1.6. Let M be a differential module over a differential ring R
admitting a finite exhaustive filtration with irreducible successive quotients.
(For instance, R could be a differential field and M could be a finitely generated differential module over R.) Then the multiset of these quotients is
independent of the choice of the filtration; we call them the (JordanHlder)
constituents of M.

5.2 Differential modules and differential systems

We now describe the link between differential modules and linear differential
systems.
Definition 5.2.1. Let R be a differential ring, and let M be a finite free differential module of rank n over R. Let e1 , . . . , en be a basis of M. Then, for any
v M, we can write v = v1 e1 + + vn en for some v1 , . . . , vn R and
compute
D(v) = v1 D(e1 ) + + vn D(en ) + d(v1 )e1 + + d(vn )en .
Define the matrix of action of D (i.e., the matrix representing the action of D)
on the basis e1 , . . . , en to be the n n matrix N over R given by the formula
D(e j ) =

n


Ni j ei .

i=1

We then have
D(v) =

n

i=1

d(vi ) +

n


Ni j v j ei .

j=1

That is, if we identify v with the column vector [v1

vn ] then

D(v) = N v + d(v).
Conversely, it is clear that, given the underlying finite free R-module, any
differential module structure is given by such an equation.

5.3 Operations on differential modules

81

Definition 5.2.2. With notation as in Definition 5.2.1, let v1 , . . . , vn be a second basis of M. The change-of-basis matrix from e1 , . . . , en to v1 , . . . , vn is
the n n matrix U defined by

Ui j ei .
vj =
i

The effect of changing basis is that the matrix of action of D on v1 , . . . , vn is

U 1 N U + U 1 d(U ).
Remark 5.2.3. In other words, differential modules are a coordinate-free version of differential systems. If you are a geometer, you may wish to go further
and think of differential bundles, i.e., vector bundles equipped with a differential operator. A differential operator on a vector bundle is usually called a
connection.

5.3 Operations on differential modules

We now describe the basic operations in the category of differential modules
over a differential ring.
Definition 5.3.1. For R a differential ring, we regard the differential modules
over R as a category in which the morphisms (or homomorphisms) from M1
to M2 are R-module homomorphisms f : M1 M2 satisfying D( f (m)) =
f (D(m)) (we sometimes say these maps are horizontal).
Definition 5.3.2. The category of differential modules over a differential ring
admits certain functors corresponding to familiar functors on the category of
modules over an ordinary ring, such as the following. (Be aware that in the
following notation the subscripted R on such symbols as the tensor product
will often be suppressed when it is unambiguous. Our habit tends to be to
drop the subscript when tensoring modules over a single ring but not when
performing a base change.)
Given two differential modules M1 , M2 , the tensor product M1 R M2
in the category of rings may be viewed as a differential module via the
formula
D(m 1 m 2 ) = D(m 1 ) m 2 + m 1 D(m 2 ).
This in particular gives meaning to the base change M R R  of a differential
R-module M to a differential R-algebra R  ; we also denote this by f M if f

82

Formalism of differential algebra

is the map from R to R  . Similarly, the exterior power nR M may be viewed as

a differential module via the formula
D(m 1 m n ) =

n


m 1 m i1 D(m i ) m i+1 m n .

i=1

(A similar fact is true for the symmetric power SymnR M, but we will have
no need of this.) The module of R-homomorphisms Hom R (M1 , M2 ) may be
viewed as a differential module via the formula
D( f )(m) = D( f (m)) f (D(m));
the homomorphisms from M1 to M2 as differential modules are precisely the
horizontal elements of Hom R (M1 , M2 ). (In this case, the subscript is quite
crucial: we have Hom(M1 , M2 ) = H 0 (Hom R (M1 , M2 )).)
We write M1 for Hom R (M1 , R) and call it the dual of M1 ; if M1 is finite
free then Hom R (M1 , M2 )
= M1 M2 and the natural map M1 (M1 )
is an isomorphism. In particular, M1 M1 contains a horizontal element corresponding to the identity map M1 M1 ; we call this the trace (element) of
M1 M1 , and we call the trivial submodule generated by the trace element the
trace component of M1 M1 . If R is a Q-algebra then M1 M1 splits as the
direct sum of the trace component with the set of elements of Hom R (M1 , M2 )
of trace zero; we call the latter the trace-zero component of M1 M1 . Even
if R is not a Q-algebra, we can still view the trace component as a quotient of
M1 M1 by duality, but the map given by embedding the trace component
into M1 M1 and then projecting onto the trace component need not be an
isomorphism.
Lemma 5.3.3. Let M, N be differential modules with M finite free. Then the
group H 1 (M N ) is canonically isomorphic to the Yoneda extension group
Ext(M, N ).
Proof. The group Ext(M, N ) consists of equivalence classes of exact sequences 0 N P M 0 under the relation that this sequence
is equivalent to a second sequence 0 N P  M 0 if there
is an isomorphism P
= P  that induces the identity maps on M and N .
Addition consists of taking two such sequences and returning the Baer sum
0 N Q/ M 0, where Q is the set of elements in P P  whose
images in M coincide and  = {(n, n) Q : n N }. The identity element
is the split sequence 0 N M N M 0. The inverse of a sequence
0 N P M 0 is the same sequence with the map N P negated.
(See [216, 3.4] for the proof that this indeed gives a group.)

5.3 Operations on differential modules

83

We first construct a canonical isomorphism Ext(M, N ) Ext(R, M N ).

Given an extension 0 N P M 0, tensor with M to get
0 M N M P M M 0.
Let Q be the inverse image of the trace component of M M; we then get
an extension
0 M N Q R 0,
yielding a map Ext(M, N ) Ext(R, M N ), which is easily shown to be
a homomorphism. In the other direction, given an extension
0 M N Q R 0,
tensor with M to get
0 M M N M Q M 0,
then quotient M Q by the kernel of the projection M M R tensored
with N . These are seen to be inverses by a diagram-chasing argument, which
we omit.
By the previous paragraph, we may reduce the statement of the lemma
to the case M = R. (One can also describe the construction without first
making this reduction, but it is a bit harder to follow.) Given an extension
0 N P R 0, compute H 0 and H 1 and apply the snake lemma to
obtain a connecting homomorphism H 0 (R) H 1 (N ). The image of 1 R
under this homomorphism determines an element of H 1 (N ), thus giving a
map Ext(R, N ) H 1 (N ). This map is easily shown to be a homomorphism
(exercise).
It remains to construct an inverse map. Given an element of H 1 (N ) represented by x N , we equip N R with the structure of a differential module
by setting
D(n, r ) = (D(n) + r x, d(r )).
This module is indeed an extension of the desired form, and the image of 1 R
under the resulting connecting homomorphism is precisely the class of x in
H 1 (N ). In the other direction, any extension splits at the level of modules and
so must have this form for some x N . This yields the claim.
Remark 5.3.4. Although we will not need this, we note that the isomorphisms Hom R (M1 , M2 )
= M1 M2 and M1 (M1 ) , and the assertion
of Lemma 5.3.3, also carry over to the case where M1 is a finite projective
R-module, i.e., a direct summand of a finite free R-module. Such a module is

84

Formalism of differential algebra

always flat. If R is noetherian then a finitely generated R-module M is projective if and only if it is locally free, i.e., if there exists a finite subset f 1 , . . . , f m
of R, generating the unit ideal, such that M[ f i1 ] is free over R[ f i1 ] for each
i [84, Exercise 4.12].

5.4 Cyclic vectors

Definition 5.4.1. Let R be a differential ring, and let M be a finite free differential module of rank n over R. A cyclic vector for M is an element m M
such that m, D(m), . . . , D n1 (m) form a basis of M.
Theorem 5.4.2 (Cyclic vector theorem). Let R be a differential field of characteristic 0 with nonzero derivation. Then every finite differential module over
R has a cyclic vector.
Many proofs are possible; we give here the proof from [80, Theorem
III.4.2]. For another proof that applies over some rings other than fields, see
Theorem 5.7.3 below. See also the notes for further discussion. (For a comment
on characteristic p, see the exercises.)
Proof. We start by normalizing the derivation. For u R , given one differential module (M, D) over (R, d), we get another differential module (M, u D)
over (R, ud), and m is a cyclic vector for one of these modules if and only if it
is a cyclic vector for the other (because the image of m under (u D) j is in the
span of m, D(m), . . . , D j (m)). We may thus assume (thanks to the assumption that the derivation is nontrivial) that there exists a nonzero element x R
such that d(x) = x.
Let M be a differential module of dimension n, and choose m M such
that the dimension of the span of m, D(m), . . . is as large as possible. We
derive a contradiction under the hypothesis < n.
For z M and Q, we have
(m + z) D(m + z) D (m + z) = 0
in the exterior power +1 M. If we write this expression as a polynomial in , it vanishes for infinitely many values so must be identically zero.
Hence each coefficient must vanish separately, including the coefficient of 1 ,
which is


i=0

m D i1 (m) D i (z) D i+1 (m) D (m).

(5.4.2.1)

5.5 Differential polynomials

85

Pick s Z, substitute x s z for z in (5.4.2.1), divide by x s , and set the result

equal to zero. We get

s i i (m, z) = 0

(s Z)

(5.4.2.2)

i=0

for

i 

i+ j
i (m, z) =
m D i+ j1 (m)D j (z)D i+ j+1 (m) D (m).
i
j=0

Again because we are in characteristic zero, we may conclude that (5.4.2.2),

viewed as a polynomial in s, has all coefficients equal to zero; that is,
i (m, z) = 0 for all m, z M.
We now take i = to obtain
(m D 1 (m)) z = 0

(m, z M);

since < n, we may use this to deduce that

m D 1 (m) = 0

(m M).

But this means that the dimension of the span of m, D(m), . . . is at most 1,
contradicting the definition of .

5.5 Differential polynomials

We now give the interpretation of differential modules as modules over a
mildly noncommutative ring.
Definition 5.5.1. Let (R, d) be a differential ring. The ring of twisted
polynomials R{T } over R in the variable T is the additive group
R (R T ) (R T 2 ) ,
with noncommuting multiplication given by the formula



j 




j
ai T i
bjT j =
ai d h (b j )T i+ jh .
h
i=0

j=0

i, j=0 h=0

In other words, we impose the relation

T a = aT + d(a)

(a R)

and then check that the result is a not necessarily commutative ring (see the
exercises). We define the degree of a twisted polynomial, in the usual way,

86

Formalism of differential algebra

as the exponent of the largest power of T with a nonzero coefficient; the

degree of the zero polynomial may be taken to be any particular negative
value.
Proposition 5.5.2 (Ore). For R a differential field, the ring R{T } admits a left
division algorithm. That is, if f, g R{T } and g  = 0 then there exist unique
q, r R{T } with deg(r ) < deg(g) and f = gq + r . (There is also a right
division algorithm.)
Proof. Exercise.
Using the Euclidean algorithm, this yields the following consequence as in
the untwisted case.
Theorem 5.5.3 (Ore). Let R be a differential field. Then R{T } is both left
principal and right principal; that is, any left ideal (resp. right ideal) has the
form R{T } f (resp. f R{T }) for some f R{T }.
Definition 5.5.4. Note that the ring opposite to R{T }, i.e., the ring in which
multiplication is performed by first switching the order of the factors, is again
a twisted polynomial ring except that the derivation is d instead of d. Given
f R{T }, we define the formal adjoint of f as the element f in the opposite
ring. This operation looks a bit less formal if we also take the coefficients over
to the other side, giving what we will call the adjoint form of f . For instance,
the adjoint form of T 3 + aT 2 + bT + c is
T 3 + T 2 a + T (b 2d(a)) + d 2 (a) d(b) + c.
Remark 5.5.5. The twisted polynomial ring is engineered precisely so that, for
any differential module M over R, we obtain an action of R{T } on M under
which T acts like D. In particular, R{T } acts on R itself with T acting like d.
In fact, the category of differential modules over R is equivalent to the category
of left R{T }-modules. Moreover, if M is a finite differential module over R,
any cyclic vector m M corresponds to an isomorphism M
= R{T }/R{T }P
for some monic twisted polynomial P, where the isomorphism carries m to
the class of 1. (We can think of f as a sort of characteristic polynomial for
M, except that it depends strongly on the choice of the cyclic vector.) Under
such an isomorphism, a factorization P = P1 P2 corresponds to a short exact
sequence 0 M1 M M2 0 with
M1
= R{T }/R{T }P1 ,
= R{T }P2 /R{T }P

M2
= R{T }/R{T }P2 .

5.7 Cyclic vectors: a mixed blessing

87

5.6 Differential equations

You may have been wondering when differential equations will appear, as these
purport to be the objects of study of this book. If so, your wait is over.
Definition 5.6.1. A differential equation of order n over the differential ring
(R, d) is an equation of the form
(an d n + + a1 d + a0 )y = b,
with a0 , . . . , an , b R, and y indeterminate. We say the equation is homogeneous if b = 0 and inhomogeneous otherwise.
Remark 5.6.2. Using our setup, we may write this equation as f (d)y = b
for some f R{T }. Similarly, we may view systems of differential equations
as being equations of the form f (D)y = b, where b lives in some differential
module (M, D). By the usual method (i.e. introducing extra variables corresponding to derivatives of y), we can convert any differential system into a
first-order system Dy = b. We can also convert an inhomogeneous system
into a homogeneous one by adding an extra variable, with the understanding
that we would like the value of that last variable to be 1 in order to get back a
solution of the original equation.
Remark 5.6.3. Here is a more explicit relationship between adjoint polynomials and the process of solving differential equations. Suppose that you start
with the cyclic differential module M
= R{T }/R{T } f and you want to find
a horizontal element. That means that you want to find some g R{T } such
that T g R{T } f ; we may as well assume that deg(g) < deg( f ). Then, by
comparing degrees, we see that in fact T g = r f for some r R. Write f in
adjoint form as f 0 + T f 1 + + T n ; then
r f r f 0 d(r ) f 1 + d 2 (r ) f 2 + (1)n d n (r ) (mod T R{T }).
In this manner, finding a horizontal element becomes equivalent to solving a
differential equation.

5.7 Cyclic vectors: a mixed blessing

The reader may at this point be wondering why so many points of view are
necessary, since the cyclic vector theorem can be used to transform any differential module into a differential equation, and ultimately differential equations
are the things one writes down and wants to solve. Permit me to interject here
a countervailing opinion.

88

Formalism of differential algebra

Remark 5.7.1. In ordinary linear algebra (or in other words, when considering differential modules for the trivial derivation), one can pass freely between
linear transformations on a vector space and square matrices if one is willing to
choose a basis. The merits of doing this depend on the situation, so it is valuable to have both the matricial and coordinate-free viewpoints well in hand.
One can then pass to the characteristic polynomial, but not all information is
retained (one loses information about nilpotency) and even information that
in principle is retained is sometimes not so conveniently accessed. In short,
no one would seriously argue that one can dispense with studying matrices
because of the existence of the characteristic polynomial.
The situation is not so different in the differential case. The difference
between a differential module and a differential system is merely the choice of
a basis, and again it is valuable to have both points of view in mind. However,
the cyclic vector theorem may seduce one into thinking that collapsing a differential system into a differential polynomial is an operation without drawbacks,
but this is far from the case. For instance, determining whether two differential
polynomials correspond to the same differential system is not straightforward.
More seriously for our purposes, the cyclic vector theorem only applies
over a differential field. Many differential modules are more naturally defined
over some ring which is not a field. For instance, differential modules arising from geometry (such as PicardFuchs modules) are usually defined over
a ring of functions on some geometric space. While there are forms of the
cyclic vector theorem available over nonfields (see for instance Theorem 5.7.3
below), these do not suffice for our purposes. We also find that working with
differential modules instead of differential polynomials has a tremendously
clarifying effect, partly because it improves the parallelism with difference
algebra, where there is no good analogue of the cyclic vector theorem even
over a field. (See Part IV.)
We find it unfortunate that much literature on complex ordinary differential equations, and nearly all the literature on p-adic ordinary differential
equations, is mired in the language of differential polynomials. By switching
instead between differential modules and differential polynomials, as appropriate, we will be able to demonstrate strategies that lead to a more systematic
development of p-adic theory.
As promised, we offer some results concerning cyclic vectors over rings,
due to Katz [122].
Theorem 5.7.2. Let R be a differential local Q-algebra. Suppose that the maximal ideal of R contains an element t such that d(t) = 1. Let M be a finite free
differential module over R, and let e1 , . . . , en be a basis of M. Then

5.7 Cyclic vectors: a mixed blessing

v=

89


j
n1 j 

t
j
(1)k
D k (e j+1k )
j!
k
j=0

k=0

is a cyclic vector of M.
Proof. For i, j 0, put

j

k j
c(i, j) =
(1)
D k (ei+ j+1k )
k
k=0

with the convention that eh = 0 for h > n. It is easily checked that

c(i + 1, j) = D(c(i, j)) + c(i, j + 1).
By induction on i, we find that
D i (v) =

n1 j

t
j=0

j!

c(i, j) ei+1

(mod t).

Hence v, D(v), . . . , D n1 (v) freely generate M modulo the maximal ideal

of R. They thus freely generate M itself by Nakayamas lemma [84,
Corollary 4.8].
Theorem 5.7.3. Let R be a differential Q-algebra containing an element t for
which d(t) = 1. Let M be a finite free differential module over R, and let I
be a prime ideal of R. Then there exists f R \ I such that M R R[ f 1 ]
contains a cyclic vector.
Proof. Choose a basis e1 , . . . , en of M, and define c(i, j) as in the proof of
Theorem 5.7.2. For x R, put
ci (x) =

n1 j

x
c(i, j).
j!
j=0

Then there exists a polynomial P(T ) R[T ] such that

c0 (x) cn1 (x) = P(x)(e1 en )

(x R).

Since c(i, 0) = ei+1 , we have P(0) = 1. In particular, the image Q(T )

(R/I )[T ] of P(T ) is not identically zero.
Since R contains Q, the images of t a for a Z are all distinct, so only
finitely many of them are roots of Q(T ). We can thus choose a Z such that
P(t a)
/ I . Since
D i (c0 (t a)) = ci (t a)

90

Formalism of differential algebra

as in the previous proof, c0 (t a) is a cyclic vector of M R R[ f 1 ] for

f = P(t a), as desired.

5.8 Taylor series

Definition 5.8.1. Let R be a topological differential ring, i.e., a ring equipped
with a topology and a derivation such that all operations are continuous.
Assume also that R is a Q-algebra. Let M be a topological differential module
over R, i.e., a differential module such that all operations are continuous. For
r R and m M, we define the Taylor series T (r, m) as the infinite sum
i

r
i=0

i!

D i (m)

whenever the sum converges absolutely (i.e., all rearrangements converge to

the same value).
Remark 5.8.2. The expression T (r, m) is de facto additive in m: if
m 1 , m 2 M then
T (r, m 1 ) + T (r, m 2 ) = T (r, m 1 + m 2 )
whenever all three terms are well-defined. For s R, T (r, s) is also de facto
multiplicative: if s1 , s2 R, then (by the Leibniz rule)
T (r, s1 )T (r, s2 ) = T (r, s1 s2 )
whenever all three terms are well-defined. More generally, for m M, T (r, m)
is de facto semilinear: if s R, m M then
T (r, s)T (r, m) = T (r, sm)
whenever all three terms are well-defined.
Example 5.8.3. A key instance of the previous remark is the case where R is
a completion of a rational function field F(t) and d is the differential operator
d/dt. In this case, the ring homomorphism T (r, ) is the substitution t  t +r ;
note that this makes sense only if |r | 1.
Remark 5.8.4. Another use for Taylor series is to construct horizontal
sections. Note that
D(T (r, m)) =


i=1

d(r )

 ri
r i1
D i (m) +
D i+1 (m)
(i 1)!
i!

= (1 + d(r ))T (r, m)

i=0

Exercises

91

if everything converges absolutely. In particular, if d(r ) = 1 then T (r, m) is

horizontal.

Notes
The subject of differential algebra is rather well developed; a classic treatment,
though possibly too dry to be useful to the casual reader, is the book of Ritt
[188]. As in abstract algebra in general, the development of differential algebra
was partly driven by differential Galois theory, i.e., the study of the expression
of solutions of differential equations in terms of solutions to ostensibly simpler
differential equations. A relatively lively introduction to the latter is [202].
Calling an element of a differential module horizontal when it is killed by
the derivation makes sense if you consider connections in differential geometry. In that setting, the differential operator is measuring the extent to which a
section of a vector bundle deviates from some prescribed horizontal direction
identifying points on one fibre with points on nearby fibres.
The history of the cyclic vector theorem is rather complicated. It appears
to have been first proved by Loewy [162] in the case of meromorphic functions, and (independently) by Cope [60] in the case of rational functions. For a
detailed historical discussion, see [53].
Twisted polynomials were introduced by Ore [177]. They are actually somewhat more general than we have discussed; for instance, one can also twist
by an endomorphism : R R by imposing the relation T a = (a)T .
(This enters the realm of the analogue of differential algebra called difference
algebra, which we will treat in Part IV.) Moreover, one can twist by both an
endomorphism and a derivation if they are compatible in an appropriate way,
and one can even study differential or difference Galois theory in this setting. A unifying framework for doing so, which is also suitable for considering
multiple derivations and automorphisms, was given by Andr [4].
Differential algebra in positive characteristic has a rather different flavor
than in characteristic 0; for instance, the pth power of the derivation d/dt on
F p (t) is the zero map. A brief discussion of the characteristic- p situation is
given in [80, III.1].

Exercises
(1) Prove that if M is a locally free differential module over R of rank 1 then
M M is trivial (as a differential module).
(2) Check that the bijection Ext(R, N ) H 1 (N ) constructed in the proof of
Lemma 5.3.3 is indeed a homomorphism.

92

Formalism of differential algebra

(3) Check that, in characteristic p > 0, the cyclic vector theorem holds for
modules of rank at most p but may fail for modules of rank p + 1.
(4) Give a counterexample to the cyclic vector theorem for a differential field
of characteristic 0 with trivial derivation.
(5) Verify that R{T } is indeed a not necessarily commutative ring; the content
in this exercise is to check the associativity of multiplication.
(6) Prove the division algorithm (Proposition 5.5.2).

6
Metric properties of differential modules

In this chapter, we study the metric properties of differential modules over nonarchimedean differential rings. The principal invariant that we will identify is
a familiar quantity from functional analysis known as the spectral radius of a
bounded endomorphism. When applied to the derivation acting on a differential module, we obtain a quantity which can be related to the least slope of the
Newton polygon of the corresponding twisted polynomial.
We can give meaning to the other slopes as well, by proving that over a
complete nonarchimedean differential field any differential module decomposes into components whose spectral radii are computed by the various
slopes of the Newton polygon. However, this theorem will provide somewhat incomplete results when we apply it to p-adic differential modules in
Part III; we will have to remedy the situation using Frobenius descendants and
antecedents.
This chapter provides important foundational material for much of what
follows, but on its own it may prove indigestably abstract at first. The reader
arriving at this opinion is advised to read Chapter 7 in conjunction with this
one, to see how the constructions of this chapter become explicit in a simple
but important class of examples.

6.1 Spectral radii of bounded endomorphisms

Before considering differential operators, let us recall the difference between
the operator norm and the spectral radius of a bounded endomorphism of an
abelian group.
Hypothesis 6.1.1. Throughout this section, let G be a nonzero abelian group
equipped with a norm | |, and let T : G G be a bounded endomorphism
93

94

Metric properties of differential modules

of G. Recall that this means that there exists c 0 such that |T (g)| c|g| for
all g G.
Definition 6.1.2. The operator norm |T |G of T is defined to be the least c 0
for which |T (g)| c|g| for all g G, i.e.,
|T |G =

sup {|T (g)|/|g|}.

gG,g=0

Recall that, if M is a finite free module over a nonarchimedean ring R, | | M is

the supremum norm for some basis, and T is an R-linear transformation, then
|T | = | A|, where A is the matrix of action of T on the chosen basis of M and
|A| = supi j {|Ai j |} as in Chapter 4.
One can similarly define the operator norm for a map between two different
abelian groups, each equipped with a norm. We may even allow seminorms as
long as we take the supremum over elements of the source group which are not
in the kernel of the seminorm.
Although the condition that T is bounded is preserved on replacing the norm
by a metrically equivalent norm, the operator norm is not preserved. To obtain
a less fragile numerical invariant, we introduce the spectral radius.
Definition 6.1.3. The spectral radius of T is defined as
1/s

|T |sp,G = lim |T s |G ;
s

the existence of the limit follows from the fact |T m+n |G |T m |G |T n |G and
from the following lemma.
Lemma 6.1.4 (Fekete). Let {an }
n=1 be a sequence of real numbers such that
am+n am + an for all m, n. Then the sequence {an /n}
n=1 either converges
to its supremum or diverges to +.
Proof. Exercise.
Proposition 6.1.5. The spectral radius of T depends on the norm | | only up
to metric equivalence.
Proof. Suppose that | | is a norm metrically equivalent to | |. We can then
choose c > 0 such that c1 |g| |g| c|g| for all g G. We then have
|T (g)|/|g| c2 |T (g)| /|g| for all g G \ {e}. Applying this with T replaced
by T s gives |T s |G c2 |T s |G , so
|T s |sp,G lim c2/s (|T s |sp,G )1/s .
s

6.2 Spectral radii of differential operators

95

Since c2/s 1 as s , this gives |T s |sp,G |T s |sp,G . The reverse

inequality is found to hold on reversing the roles of the norms.
We have the following relationship to the concepts studied in Chapter 4.
Remark 6.1.6. Let F be a field equipped with a norm, and let V be a finitedimensional vector space over F. Pick a basis for V , and equip V with either
the L 2 norm or the supremum norm defined by this basis, according to whether
F is archimedean or nonarchimedean. Let A be the matrix of action of T on
this basis. Then |T |V equals the largest singular value of A, whereas |T |sp,V
equals the largest norm of an eigenvalue of A (by Proposition 4.2.4 in the
archimedean case, and Proposition 4.4.10 in the nonarchimedean case).
Remark 6.1.7. The spectral radius derives its name from the fact that, by a
celebrated theorem of Gelfand, for a bounded endomorphism of a commutative
Banach algebra over R or C the spectral radius computes the radius of the
smallest disc containing the entire spectrum of the operator. (This includes the
archimedean case of the previous remark.)
When dealing with the spectral radius we will frequently use the following
observation.
Lemma 6.1.8. For any  > 0 there exists c = c() such that for all s 0,
|T s |G c(|T |sp,G + )s .
Proof. For c = 1, the claim already holds for all but finitely many s. It thus
suffices to increase c in order to cover the finite set of exceptions.
Remark 6.1.9. Some authors writing about p-adic differential equations
(including the present author) tend to refer to the spectral radius as the spectral norm. On the one hand, this is somewhat dangerous, because the spectral
radius is not in general a norm or even a seminorm, even for matrices over
a complete field (exercise). On the other hand, we will be using the spectral
radius as a measure of size and so the normlike notation | |sp,V is useful. We
will compromise by keeping this notation but referring to spectral radii rather
than spectral norms.

6.2 Spectral radii of differential operators

We now specialize the previous discussion to the case of differential modules.
Definition 6.2.1. By a nonarchimedean differential ring or field, we mean
a nonarchimedean ring equipped with a bounded derivation. For F a

96

Metric properties of differential modules

nonarchimedean differential field, we can define the operator norm |d| F and
the spectral radius |d|sp,F ; by hypothesis the former is finite, so the latter is too.
Definition 6.2.2. Let F be a nonarchimedean differential field. By a normed
differential module over F we mean a vector space V over F equipped with a
compatible norm | |V and a derivation D with respect to d which is bounded
as a endomorphism of the additive group of V . For V nonzero, we may then
consider the operator norm |D|V and the spectral radius |D|sp,V .
Remark 6.2.3. If V is finite-dimensional over F and F is complete then the
spectral radius does not depend on the norm on V , since by Theorem 1.3.6 any
two norms on V compatible with the norm on F are metrically equivalent.
In general, one cannot have differential modules with arbitrarily small
spectral radius.
Lemma 6.2.4. Let F be a nonarchimedean differential field, and let V be a
nonzero normed differential module over F. Then
|D|sp,V |d|sp,F .
Proof. (This proof was suggested by Liang Xiao.) For a F and v V
nonzero, the Leibniz rule gives

si 

si si j
D si (a D i (v)) = d si (a)D i (v)+
(a)D i+ j (v) (0 i s).
d
j
j=1

Inverting this system of equations gives an identity of the form

d s (a)v =

s


cs,i D si (a D i (v))

i=0

for certain universal constants cs,i Z. Consequently,

|d s (a)v|V max {|D si (a D i (v))|V }.
0is

(6.2.4.1)

By Lemma 6.1.8, given  > 0 we can choose c = c() such that, for all s 0,
|D s |V c(|D|sp,V + )s .
Using (6.2.4.1), we deduce that
|d s (a)v|V c2 (|D|sp,V + )s |a||v|V .
Dividing by |a||v|V and taking the supremum over a F, we obtain
|d s | F c2 (|D|sp,V + )s .

6.2 Spectral radii of differential operators

97

Taking sth roots and then letting s , we get

|d|sp,F |D|sp,V + .
Since  > 0 was arbitrary, this yields the claim.
It is sometimes useful to compute in terms of a basis of V over F.
Lemma 6.2.5. Let F be a complete nonarchimedean differential field, and
let V be a nonzero finite differential module over F. Fix a basis e1 , . . . , en
of V , and let Ds be the matrix of action of D s on this basis (i.e., D s (e j ) =

i (Ds )i j ei ). Then
|D|sp,V = max{|d|sp,F , lim sup |Ds |1/s }.

(6.2.5.1)

Proof. (Compare [49, Proposition 1.3].) Equip V with the supremum norm
defined by e1 , . . . , en ; then |D s |V maxi, j |(Ds )i, j |. This plus Lemma 6.2.4
imply that the left-hand side of (6.2.5.1) is greater than or equal to the the
right-hand side.
Conversely, for any x V , if we write x = x1 e1 + + xn en with
x1 , . . . , xn F then
s 
n 

s j
D s (x) =
d (xi )D s j (ei ),
j
i=1 j=0

so



1/s
1/s
|D s |V max |d j | F |Ds j |1/s .
0 js

(6.2.5.2)

Given  > 0, apply Lemma 6.1.8 to choose c = c() such that, for all s 0,

s
|d s | F c |d|sp,F +  ,

s
|Ds |V c lim sup |Ds |1/s +  .
s

Then (6.2.5.2) implies that


1/s
|D s |V c2/s max |d|sp,F + , lim sup |Ds |1/s +  .
s

As in the previous proof, the factor c2/s tends to 1 as s . From this it

follows that the right-hand side of (6.2.5.1) is greater than or equal to the lefthand side minus ; since  > 0 was arbitrary, we get the same inequality with
 = 0.
Remark 6.2.6. In Lemma 6.2.5, if |D|sp,V > |d|sp,F then the limit superior
can be replaced by a limit; see the exercises.

98

Metric properties of differential modules

Using Lemma 6.2.5, we may infer the following base-change property for
the spectral radius. We will need to refine this result later; see Corollary 6.5.5
and Proposition 10.6.6 below.
Corollary 6.2.7. Let F F  be an isometric embedding of complete nonarchimedean differential fields. (In particular, the differential on F  must restrict
to the differential on F.) Then, for any nonzero finite differential module V
over F,
|D|sp,V  = max{|d|sp,F  , |D|sp,V }.
Consequently, if |d|sp,F  = |d|sp,F then |D|sp,V  = |D|sp,V by Lemma 6.2.4.
Proof. Choose a basis of V , and use Lemma 6.2.5 to compute both |D|sp,V
and |D|sp,V  in terms of this basis. This yields the desired formula.
Here is how the spectral radius behaves with respect to basic operations on
the category of differential modules.
Lemma 6.2.8. Let F be a complete nonarchimedean differential field.
(a) For a short exact sequence 0 V1 V V2 0 of nonzero finite
differential modules over F,
|D|sp,V = max{|D|sp,V1 , |D|sp,V2 }.
(b) For V a nonzero finite differential module over F,
|D|sp,V = |D|sp,V .
(c) For V1 , V2 nonzero finite differential modules over F,
|D|sp,V1 V2 max{|D|sp,V1 , |D|sp,V2 },
with equality when |D|sp,V1  = |D|sp,V2 .
Proof. To prove (a), choose a splitting V = V1 V2 of the short exact sequence
in the category of vector spaces over F. The action of D on V is then given
by D(v1 , v2 ) = (D(v1 ) + f (v2 ), D(v2 )) for some F-linear map f : V2 V1
(as in the proof of Lemma 5.3.3). From this, the rest of the proof of (a) is an
exercise in the spirit of the proofs of Lemmas 6.2.4 and 6.2.5.
We next recall from Definition 5.3.2 that for any finite differential modules
V1 , V2 over F, the action of D on V = Hom F (V1 , V2 ) is given by
D( f )(v1 ) = D( f (v1 )) f (D(v1 )).

6.2 Spectral radii of differential operators

99

The Leibniz rule in this setting is that, for any nonnegative integer s,
 !
s
"

s
i i
(1)
D i f (D si (v1 )) .
D ( f )(v1 ) =
s
i=0

We may again (as an exercise) deduce that |D|sp,V max{|D|sp,V1 , |D|sp,V2 }.

This implies first (b) (since we get |D|sp,V |D|sp,V by Lemma 6.2.4 and
similarly for the opposite inequality) and then the first assertion of (c) (since
V1 F V2 = Hom F (V1 , V2 )).
It remains to prove the second assertion of (c). Suppose that |D|sp,V1 >
|D|sp,V2 . Then by (b) and the first assertion of (c),
|D|sp,V1 = max{|D|sp,V1 , |D|sp,V2 }
max{|D|sp,V1 V2 , |D|sp,V2 }
|D|sp,V1 V2 V2 .
Moreover, V2 V2 contains a trivial submodule (the trace), so V1 V2
V2 contains a copy of V1 . Hence, by (a), |D|sp,V1 V2 V2 |D|sp,V1 .
We thus obtain a chain of inequalities leading to |D|sp,V1 |D|sp,V1 ; this
forces the intermediate equality |D|sp,V1 = max{|D|sp,V1 V2 , |D|sp,V2 }. Since
|D|sp,V1  = |D|sp,V2 = |D|sp,V2 we have |D|sp,V1 = |D|sp,V1 V2 , as
desired.
Corollary 6.2.9. If V1 , V2 are irreducible finite differential modules over
a complete nonarchimedean differential field and if |D|sp,V1  = |D|sp,V2
then every irreducible submodule W of V1 V2 satisfies |D|sp,W =
max{|D|sp,V1 , |D|sp,V2 }.
Proof. Suppose the contrary; we may assume that |D|sp,V1 > |D|sp,V2 . The
inclusion W V1 V2 corresponds to a nonzero horizontal section of W
V1 V2
= (W V2 ) V1 , which in turn corresponds to a nonzero map

W V2 V1 . Since V1 is irreducible, the map has image V1 ; that is, W V2

has a quotient isomorphic to V1 .
However, we can contradict this using Lemma 6.2.8. Namely,
|D|sp,W V2 max{|D|sp,W , |D|sp,V2 } < |D|sp,V1 ,
so each nonzero subquotient of W V2 has spectral radius strictly less than
|D|sp,V1 .
Remark 6.2.10. By contrast, when |D|sp,V1 = |D|sp,V2 it is entirely possible for an irreducible submodule W of V1 V2 to satisfy |D|sp,W  =
max{|D|sp,V1 , |D|sp,V2 }. For instance, take V1 with |D|sp,V1 > |d|sp,F , put
V2 = V1 , and let W be the trace component of V1 V2 .

100

Metric properties of differential modules

Remark 6.2.11. It would be convenient to have the analogue of

Corollary 6.2.9 also for irreducible subquotients of V1 V2 . We will prove
something slightly weaker later (Corollary 6.6.3).
We now refine the notion of the spectral radius to give a working notion of
the spectrum of a differential operator.
Definition 6.2.12. For V a finite differential module over a nonarchimedean
differential field F, let V1 , . . . , Vl be the JordanHlder constituents of V
(listed with multiplicity). Define the full spectrum of V to be the multiset consisting of |D|sp,Vi with multiplicity dim F Vi , for i = 1, . . . , l. We say V is pure
if its full spectrum consists of a single element (with multiplicity). We say V
is refined if V is pure and |D|sp,V V < |D|sp,V . If V and W are two refined
finite differential modules over F, we write V W if
|D|sp,V = |D|sp,W > |D|sp,V W .
We will see that this is an equivalence relation (Lemma 6.2.14).
Remark 6.2.13. It may be helpful to keep in mind how the above notions
behave when the derivation on F is zero. In this case, a finite differential module over F is simply a finite-dimensional vector space V equipped with a linear
transformation. The full spectrum consists of the norms of the eigenvalues in
F alg of this linear transformation. The module V is pure if these eigenvalues
all have the same norm. The module V is refined if it is pure and, moreover,
the ratio of any two eigenvalues is congruent to 1 modulo m F . In particular,
we can always decompose a finite differential module into pure summands,
but not necessarily into refined summands; for instance, one cannot separate a
linear transformation over Q p with characteristic polynomial T 2 p.
However, we can achieve a refined decomposition after a finite tamely ramified extension. We may see this from Definition 3.3.3 as follows. Let E be
a finite Galois extension of F containing the eigenvalues of the linear transformation. Then, for any g in the wild inertia subgroup W E/F of G E/F and
any eigenvalue , we have |g()/ 1| < 1. We thus obtain a refined
decomposition over the fixed field of W E/F .
We can also describe a suitable tamely refined extension F  of F explicitly, by ensuring that it satisfies the following conditions in terms of the
characteristic polynomial Q of the linear transformation.
(a) For each root of Q we have

p = 0,
|(F  ) |
|| #
h


1/
p
p > 0,
h0 |(F ) |
where p is the characteristic of F .

6.2 Spectral radii of differential operators

101

(b) For any two roots , of Q, we have

/ F  .
Lemma 6.2.14. Let F be a nonarchimedean differential field. Then the
relation on refined finite differential modules over F is an equivalence
relation.
Proof. The reflexivity of holds because we are considering only refined
modules. The symmetry of holds because |D|sp,V W = |D|sp,W V by
Lemma 6.2.8(b). To check the transitivity, suppose that V W and W X .
Since V X occurs as a direct summand of (V W ) (W X ), by
Lemma 6.2.8(a) and (c) we have
|D|sp,V X |D|sp,V W W X
max{|D|sp,V W , |D|sp,W X }
< |D|sp,V = |D|sp,W = |D|sp,X .
Hence V X .
Lemma 6.2.15. Let F be a complete nonarchimedean differential field, and
let F  be a finite tamely ramified extension of F. Then d extends uniquely to
F  , and |d| F  = |d| F .
Proof. Suppose first that F  is unramified over F. For F  , let P(T ) =

i
i Pi T F[T ] be the minimal polynomial of . Then the unique extension
of d to F  is characterized by

0 = d()P  () +
d(Pi ) i .
(6.2.15.1)
i

If

o
F

then the unramified condition implies that |P  ()| = 1, so that

|d()| max{|d(Pi )|||i } |d| F .
i

In general, we can write any F  as with o

F  and F. We then
have d() = d() + d( ), so that
|d()| max{| ||d()|, |||d( )|} |d| F ||| | = |d| F ||.
This proves the claim.
We next verify the claim in the case F  = F(t 1/m ) for some t F and some
/ |F |
integer m, not divisible by the characteristic p of F , such that |t|1/e

for any divisor e > 1 of m. Again, the unique extension of d to F is given

102

Metric properties of differential modules

by (6.2.15.1). Also, each element of F  has a unique expression of the form

m1 i/m
with ci F, and the norm on F  is given by
i=0 ci t
m1





ci t i/m  = max{|ci ||t|i/m }.



i
i=0

It thus suffices to check that |d(ci t i/m )| |d| F |ci t i/m | for i = 0, . . . , m 1.
By the Leibniz rule, this reduces to checking |d(t i/m )| |d| F |t i/m |. But
|d(t i/m )| = |t i/m1 d(t)| |t i/m ||t|1 |d| F |t| = |d| F |t i/m |,
so the claim follows.
To treat the general case, we may choose an integer m not divisible by p and
annihilating I E/F and some t1 , . . . , th F such that |t1 |, . . . , |th | generate
1/m
1/m
|F |/|F |m . Put F  = F(t1 , . . . , th ). By the two previous paragraphs,
1/m
we have |d| F  = |d| F . (More precisely, when adjoining ti
we split the
reasoning into two cases: first, for e the largest divisor of m such that |ti |1/e is
1/e
already present, adjoining ti gives an unramified extension, as in the proof
of Proposition 3.3.6, so the operator norm of d does not change. After that,
1/m
proceeds as in the second paragraph.) By Proposition 3.3.6, F 
adjoining ti
is contained in an unramified extension of F  so the claim follows by applying
the first paragraph again.

6.3 A coordinate-free approach

We note in passing that our definition of the spectral radius is rather sensitive
to the choice of the derivation d. (See the exercises for Chapter 9 for an explicit
example.) A coordinate-free approach is suggested by the work of Baldassarri
and Di Vizio.
Proposition 6.3.1. Let F be a nonarchimedean differential field. Let F{T }(s)
be the set of twisted polynomials, of degree at most s, equipped with the
seminorm |P| = |P(d)| F compatible with F (that is, consider P(d) as an
operator on F). Let V be a nonzero finite differential module over F, and fix
a norm on V compatible with F. Let L(V ) be the space of bounded endomorphisms of the additive group of V , equipped with the operator norm. Let
Ds : F{T }(s) L(V ) be the map P  P(D). Then
|D|sp,V |d|sp,F lim inf |Ds |1/s .
s

(6.3.1.1)

Conversely, suppose that, for any nonnegative integer s and any c0 , . . . , cs

F, we have

6.3 A coordinate-free approach

|c0 + c1 d + + cs d s | F = max{|c0 | F , |c1 d| F , . . . , |cs d s | F }.

103
(6.3.1.2)

Then equality holds in (6.3.1.1), and if |d|sp,F > 0 then the limit inferior on
the right is also a limit.
The condition (6.3.1.2) is satisfied in the situation of greatest interest to us;
see Proposition 9.10.2.
Proof. By taking T s F{T }(s) we obtain the inequality |D s |V |d s | F |Ds |.
Taking sth roots of both sides and then taking limits as s yields (6.3.1.1).
Conversely, suppose that (6.3.1.2) holds. Given  > 0, apply Lemma 6.1.8
to choose c > 0 such that, for all s 0,
|D s |V c(|D|sp,V + )s .

s
Given a nonnegative integer s such that |Ds | > , choose P = i=0
Pi T i
(s)
F{T } nonzero and such that |P(D)|V |P(d)| F (|Ds | ). (This could
only fail to hold if we had |P(d)| F = 0 for all P F{T }(s) , but the presence
of constant polynomials eliminates this possibility.) Then
max{|Pi ||d|isp,F (|Ds | )} max{|Pi d i | F (|Ds | )}
is

is

= |P(d)| F (|Ds | )
|P(D)|V
max{|Pi D i |V }
is

max{|Pi |c(|D|sp,V + )i }.
is

For the index i which maximizes the right-hand side, we have

|d|isp,F (|Ds | ) c(|D|sp,V + )i .
Since |d|sp,F < |D|sp,V +  by Lemma 6.2.4, we can increase the exponent
from i to s on both sides while maintaining the inequality; we may then also
include values of s for which |Ds | < . Taking sth roots and then taking the
limit as s yields
|d|sp,F lim sup(|Ds | )1/s |D|sp,V + .
s

Since this holds for any  > 0, we obtain the inequality

|D|sp,V |d|sp,F lim sup |Ds |1/s .
s

This forces equality in (6.3.1.1) and forces the limit inferior therein to be a
limit provided that |d|sp,F > 0.

104

Metric properties of differential modules

6.4 Newton polygons for twisted polynomials

Twisted polynomials admit a partial analogue of the theory of Newton polygons; we will use these polygons in the next section to compute spectral radii
of differential operators.
Definition 6.4.1. Let R be a nonarchimedean differential field domain. For
|d| R , define the -Gauss norm on the twisted polynomial ring R{T } by





i
Pi T  = max{|Pi | i };



i
i

thus it is the same as the -Gauss norm of the untwisted polynomial i Pi T i

R[T ]. For r log |d| R we obtain a corresponding r -Gauss valuation
vr (P) = log |P|er .
Lemma 6.4.2. For |d| R the -Gauss norm is multiplicative. Moreover,
any polynomial and its formal adjoint have the same -Gauss norm.
Proof. It suffices to check the statement for > |d| R , as the boundary case
may be inferred from the continuity of the map  |P| for fixed P. The key
observation (and the source of the restriction on ) is that, for P, Q R{T }
and > |d| R ,
|P Q Q P| 1 |d| R |P| |Q| < |P| |Q| .
(The first inequality is evident in the case P = T and Q R; the reduction
of the general case to this one is an exercise.) This allows us to deduce multiplicativity on R{T } from multiplicativity on R[T ] (Proposition 2.1.2). The
claim about the adjoint follows by a similar argument.

Definition 6.4.3. We define the Newton polygon of P = i Pi T i R{T }

by taking the Newton polygon of the corresponding untwisted polynomial

Pi T i R[T ] and then omitting all slopes greater than or equal to

log |d| R . We define similarly the multiplicity for slopes less than log |d| R .
By Lemma 6.4.2 and an argument as in Remark 2.1.7, the multiplicity of a
slope r < log |d| R in a product P Q is equal to the sum of the multiplicities
of r as a slope of P and as a slope of Q. However, this fails for r = log |d| R
because we are only controlling the left endpoint of the segment of the Newton
polygon with this slope (see the exercises); hence we will omit log |d| R and
all larger slopes from the Newton polygon.
By another application of the master factorization theorem (Theorem 2.2.2),
we obtain the following.

6.5 Twisted polynomials and spectral radii

105

Theorem 6.4.4. Let R be a complete nonarchimedean differential domain.

Suppose that S R{T }, r < log |d| R , and m Z0 satisfy
vr (S T m ) > vr (T m ).
Then there exists a unique factorization S = P Q satisfying the following
conditions.
(a) The polynomial P R{T } has degree deg(S) m, and its slopes are
all less than r .
(b) The polynomial Q R{T } is monic of degree m, and its slopes are all
greater than r .
(c) We have vr (P 1) > 0 and vr (Q T m ) > vr (T m ).
Moreover, for this factorization,
min{vr (P 1), vr (Q T m ) vr (T m )} vr (S T m ) vr (T m ).
Similarly, we can make a factorization S = Q P having the same properties
(but the factors themselves may differ).
Proof. Use the same procedure as in Theorem 2.2.1.
Corollary 6.4.5. Let F be a complete nonarchimedean differential field. If
P F{T } is irreducible, then either it has no slopes or all its slopes are equal
to some value less than log |d| F .
Remark 6.4.6. Instead of defining the Newton polygon as above and then
truncating, one may prefer to declare all the missing slopes to be equal to
log |d| R . One can achieve this by modifying the set whose lower convex hull
is used to define the Newton polygon, in order to build in the right truncation
behavior. The correct modified set is
n 

$
(x, y) R2 : x i, y v(Pi ) (x + i) log |d| F ,
i=0

provided that d is nonzero. (If d = 0 then (x +i) log |d| F must be interpreted
to mean 0 if x + i = 0 and + if x + i > 0.)

6.5 Twisted polynomials and spectral radii

One can use twisted polynomials over nonarchimedean differential fields to
detect part of the full spectrum of a normed differential module.
Definition 6.5.1. For V a finite differential module over a nonarchimedean
differential field F, define the visible spectrum of V to be the submultiset of
the full spectrum of V consisting of those values greater than |d| F .

106

Metric properties of differential modules

Remark 6.5.2. In the application to regular singularities (Chapter 7), we will

consider the case where |d| F = |d|sp,F . In such a case, there is no real loss
in restricting to the visible spectrum: the only missing norm is |d| F itself,
and one can infer its multiplicity from the dimension of the module. However, in the applications to p-adic differential equations in Part III we will
have |d| F > |d|sp,F , so restriction to the visible spectrum will cause real problems; these will have to be remedied using pullback and pushforward along a
Frobenius map.
The key theorem relating spectral radii to Newton polygons is the following.
Theorem 6.5.3 (ChristolDwork). Let F be a complete nonarchimedean differential field. For nonconstant P F{T }, put V = F{T }/F{T }P. Let r
be the least slope of the Newton polygon of P or log |d| F if no such slope
exists. Then
max{|d| F , |D|sp,V } = er .
Proof. Let r1 rk be the slopes of P counted with multiplicity, and
define rk+1 = = rn = log |d| F . Equip V with the norm

n1



i
ai T  = max{|ai |ern1 rni }.



i
i=0

As in the proof of Proposition 4.3.10, we then have |D|V = er1 and so

|D|sp,V er1 .
To finish, we must check that if r1 < log |d| F then |D|sp,V = er1 . Let
be the operation defined by
 n1
n1


i
ai T =
d(ai )T i ;

i=0

i=0

then ||V = |d| F , D is F-linear, and |D |V = er1 . Then, for all positive
integers s,
|(D )s |V = er1 s ,

|D s (D )s |V er1 (s1) |d| F < er1 s ,

so |D s |V = er1 s and |D|sp,V = er1 as desired.

Corollary 6.5.4. Let F be a complete nonarchimedean differential field. For
any P F{T }, the visible spectrum of the differential module F{T }/F{T }P
consists of er , where r runs over the slope multiset of the Newton polygon
of P.

6.6 The visible decomposition theorem

107

Proof. Write down a maximal factorization of P; it corresponds to a maximal

filtration of F{T }/F{T }P. By Corollary 6.4.5, each factor in the factorization
has only a single slope, so Theorem 6.5.3 gives us what we want.
We now obtain a partial refinement of Corollary 6.5.4. For a further
refinement, see Proposition 10.6.6.
Corollary 6.5.5. Let F F  be an isometric embedding of complete nonarchimedean differential fields. For any finite differential module V over F, the
visible spectrum of V F F  is the submultiset of the visible spectrum of V
consisting of those values greater than |d| F  .
Proof. We may reduce to the case where V is irreducible (but V F F  need
not be). In this case any nonzero element of V is a cyclic vector, so we may
apply Corollary 6.5.4 to deduce the claim.
Using Theorem 6.5.3, we can give a differential version of Proposition 4.4.6.
Proposition 6.5.6. Let F be a complete nonarchimedean differential field
with |d| F 1. Let V be a finite differential module of rank n > 0 over
F with |D|sp,V 1. Let | |V be the supremum norm on V defined by a
basis e1 , . . . , en , and suppose that |D|V = c 1. Then there exists a basis
of V defining a second supremum norm | |V , for which |D|V 1 and
|x|V |x|V cn1 |x|V for all x V .
Proof. We proceed as in Proposition 4.4.6. Let M be the smallest
o F -submodule of V containing e1 , . . . , en and stable under D. For each i,
if j = j (i) is the least integer such that ei , D(ei ), . . . , D j (ei ) are linearly

j1
dependent then we have D j (ei ) = h=0 ch D h (ei ) for some ch F. Since
|D|sp,V 1 and |d| F 1, Theorem 6.5.3 implies that |ch | 1 for each h.
Hence M is finitely generated, and thus free, over o F .
Let | |V be the supremum norm on V defined by a basis of M. Then
|x|V |x|V because e1 , . . . , en M. Conversely, for i = 1, . . . , n and
h = 0, . . . , j (i) 1, we have |D h (ei )|V |D|hV ch cn1 . Since the
D h (ei ) generate M, this implies that |x|V cn1 |x|V for all x V .

6.6 The visible decomposition theorem

Using twisted polynomials, we can split V into components corresponding to
the elements of the visible spectrum (see Definition 6.5.1).
Theorem 6.6.1 (Visible decomposition theorem). Let F be a complete
nonarchimedean differential field of characteristic 0, and let V be a finite
differential module over F. Then there exists a unique decomposition

108

Metric properties of differential modules

V = V0

Vs

s>|d| F

of differential modules, such that every subquotient of Vs has spectral radius s

and every subquotient of V0 has spectral radius at most |d| F .
Proof. If the derivation on F is zero then we are just considering a vector
space equipped with an endomorphism, and the claim becomes an elementary
exercise in linear algebra (as noted in Remark 6.2.13). Thus we may assume
hereafter that the derivation is nonzero.
We will induct on dim(V ). By Theorem 5.4.2 and our hypothesis that
d  = 0, there exists a cyclic vector for V . We thus obtain an isomorphism
V
= F{T }/F{T }P for some P F{T }. If the Newton polygon of P is empty,
we may put V = V0 and be done. Otherwise, let r < log |d| F be the least
slope. By applying Theorem 6.4.4 once to P, we obtain a short exact sequence
0 V1 V V2 0 in which (by Theorem 6.5.3) every subquotient
of V1 has spectral radius er , and every subquotient of V2 has spectral radius
less than er . Applying Theorem 6.4.4 again to P but with the factors in the
opposite order, we obtain a short exact sequence 0 V2 V V1 0 in
which every subquotient of V1 has spectral radius er and every subquotient
of V2 has spectral radius less than er . This yields V1 V2 = 0, so V1 V2
injects into V . Moreover, dim V1 = dim V1 and dim V2 = dim V2 because P
and its formal adjoint have the same Newton polygon (Lemma 6.4.2). By a
dimension count V1 V2 must equal V . Applying the induction hypothesis to
V2 gives the claim.
Corollary 6.6.2. Let F be a complete nonarchimedean differential field of
characteristic 0. Let V be a finite differential module over F such that every
subquotient of V has spectral radius greater than |d| F . Then H 0 (V ) =
H 1 (V ) = 0.
Proof. The claim about H 0 is clear: a nonzero element of H 0 (V ) would generate a differential submodule of V which would be trivial and thus would have
spectral radius |d|sp,F |d| F . In the case of H 1 , let 0 V W F 0
be a short exact sequence of differential modules. Decompose W = W0 W1
according to Theorem 6.6.1, so that every subquotient of W0 has spectral radius
at most |d| F and every subquotient of W1 has spectral radius greater than |d| F .
The map V W0 must vanish (its image is a subquotient of both V and W0 ),
so V W1 . But W1  = W , as otherwise W could not surject onto a trivial module, so for dimensional reasons we must have V = W1 . Hence the sequence
splits, proving that H 1 (V ) = 0 by Lemma 5.3.3.

6.7 Matrices and the visible spectrum

109

In this setting we can obtain a refinement of Corollary 6.2.9 in which we

allow subquotients rather than just submodules.
Corollary 6.6.3. Let F be a complete nonarchimedean differential field of
characteristic 0. If V1 , V2 are finite irreducible differential modules over F,
|D|sp,V1 > |d| F , and |D|sp,V1 > |D|sp,V2 , then every irreducible subquotient
W of V1 V2 satisfies |D|sp,W = |D|sp,V1 .
&
Proof. Decompose V1 V2 as V0 s>|d| F Vs according to Theorem 6.6.1; we
have Vs = 0 whenever s > |D|sp,V1 . If either V0 were nonzero or some Vs with
s < |D|sp,V1 were nonzero, then V1 V2 would have an irreducible submodule
of spectral radius less than |D|sp,V1 , in violation of Corollary 6.2.9.
These results are quite sufficient for applications to the study of the singularities of complex meromorphic differential equations, at which we hint in
Chapter 7. However, in the p-adic situation we have to decompose V0 further;
we will do this using Frobenius antecedents in Chapter 10.

6.7 Matrices and the visible spectrum

The proof of Theorem 6.5.3 relied on the fact that one can detect the spectral
radius of a differential module admitting a cyclic vector by using the characteristic polynomial of the matrix of action of D (see Definition 5.2.1) on
the cyclic basis. For some applications we need to extend this to certain bases
not necessarily generated by cyclic vectors; for this, the relationship between
singular values and eigenvalues considered in Chapter 4 will be crucial.
Lemma 6.7.1. Let R be a complete nonarchimedean differential domain. Let
N be a 2 2 block matrix over R with the following properties.
(a) The matrix N11 has an inverse A over R.
(b) We have | A| max{|d| F , |N12 |, |N21 |, |N22 |} < 1.
Then there exists a block upper triangular unipotent matrix U over R, such
that |U12 | | A| max{|N12 |, |N21 |, |N22 |} and U 1 N U + U 1 d(U ) is block
lower triangular.
Proof. Put
= | A| max{|N12 |, |N21 |, |N22 |} < 1,

 = | A||N12 | .

Let X be the block upper triangular nilpotent matrix with X 12 = AN12 , put
U = I X , and put
N  = U 1 N U + U 1 d(U ).

110

Metric properties of differential modules

Since U 1 = I + X , we have N  = N + X N N X X N X d(X ). In block

form,


N11 + X 12 N21 N12 N11 X 12 + X 12 N22 X 12 N21 X 12 d(X 12 )
.
N =
N22 N21 X 12
N21
We now claim that

|N12
|  max{, |d| F |A|}|A|1 ,

|N21
| | A|1 ,

|N22
| | A|1 .

The second and third lines hold because

|(U 1 N U N )2 j | = |(N X )2 j | | A|1

( j = 1, 2).

The first line holds because N12 N11 X 12 = 0, so we can write


= X 12 N22 X 12 N21 X 12 d(X 12 ),
N12

in which the first two terms have norm at most | A|1 and the third has norm
at most |d| F .
 , we write it as (I + X N A)N . Because |X N A|
To analyze N11
12 21
11
12 21
|X 12 |(|N21 ||A|) <  < 1 the first factor is invertible, and it and its inverse both
 is invertible, |N  | = |N |, and |(N  )1 | = |A|.
have norm 1. Hence N11
11
11
11
Since max{, |d| F |A|} for some fixed < 1, iterating the construction
of N  from N yields a convergent sequence of conjugations whose limit has
the desired property.
We need a version of the argument used in the proof of Theorem 6.5.3 that
is no longer restricted to cyclic vectors.
Lemma 6.7.2. Let F be a complete nonarchimedean differential field. Let V
be a nonzero finite differential module over F. Let e1 , . . . , en be a basis of
V , and let N be the matrix of the action of D on e1 , . . . , en . Suppose that
|N | = > |d| F and |N 1 | = 1 . Then the full spectrum of V consists
entirely of .
Proof. As in the proof of Theorem 6.5.3, we find that for the supremum norm
for e1 , . . . , en we have |D s (v)|V = s |v|V for all nonnegative integers s and
all v V . Consequently, for any nonzero differential submodule W of V ,
we have |D|sp,W = . By Theorem 6.6.1, it follows that every irreducible
subquotient of V also has spectral radius , as desired.

6.7 Matrices and the visible spectrum

111

By combining what we have so far, we can obtain a further refinement of

Lemma 6.7.2 in which we can detect elements of the visible spectrum, which
need not all be equal.
Lemma 6.7.3. Let F be a complete nonarchimedean differential field. Let V
be a finite differential module over F. Let e1 , . . . , en be a basis of V , and
let N be the matrix of action of D on e1 , . . . , en . Let 1 n be the
singular values of N , and let 1 , . . . , n be the eigenvalues of N , arranged so
that |1 | |n |. Suppose that the following conditions hold for some
i = 1, . . . , n and some |d| F .
(a) We have i > .
(b) Either i = n or i+1 .
(c) We have j = | j | for j = 1, . . . , i.
Then the elements of the full spectrum that are greater than are precisely
1 , . . . , i .
Proof. We first check that conditions (a)(c) are invariant under the change of
basis
N  U 1 N U + U 1 d(U ) = U 1 (N + d(U )U 1 )U
for U GLn (o F ). Note that, by Proposition 4.4.1, N and N + d(U )U 1 have
the same singular values greater than . Thanks to conditions (a)(c) we may
apply Theorem 4.4.2 to deduce that N and N + d(U )U 1 also have the same
eigenvalue norms for eigenvalues greater than . Since neither the singular
values nor the eigenvalues are altered upon conjugating by U , we may draw
the same conclusion about N and U 1 N U + U 1 d(U ). In particular, the truth
of (a)(c) is not affected by this change of basis.
If 1 |d| F then we have nothing to check. If 1 = = n > |d| F
then Lemma 6.7.2 implies the claim. If neither of these cases applies, we
may induct on n: choose i with 1 = = i > i+1 , so that necessarily 1 > |d| F . View N as a 2 2 block matrix with block sizes i, n i.
Apply Lemma 6.7.1 to obtain an upper triangular unipotent block matrix U
over o F such that N  = U 1 N U + U 1 d(U ) is lower triangular. We may
then reduce to checking the claim with N replaced by the two diagonal blocks
of N  .
To put everything together, we relax the condition on the singular values.
Theorem 6.7.4. Let F be a complete nonarchimedean differential field. Let
V be a finite differential module over F. Let e1 , . . . , en be a basis of V ,
and let N be the matrix of action of D on e1 , . . . , en . Let 1
n be the singular values of N , and let 1 , . . . , n be the eigenvalues of

112

Metric properties of differential modules

N arranged so that |1 | |n |. Define f n as in Corollary 4.4.8

and put = f n (1 , . . . , n , |1 |, . . . , |n |, |d| F ). Suppose that the following
conditions hold for some i = 1, . . . , n and some |d| F .
(a) We have |i | > .
(b) Either i = n or |i+1 | .
Then the elements of the full spectrum greater than are precisely
|1 |, . . . , |i |.
Proof. By Corollary 6.5.5 we can replace F by a larger complete nonarchimedean differential field F  , provided that |d| F  = |d| F . In particular we may take
F  to be the completion of F(t) for the -Gauss norm for any > 0, extending d so that d(t) = 0 (exercise). After enlarging F suitably in this manner, by
Corollary 4.4.8 we can choose a matrix U GLn (F) such that the following
conditions hold.
(a) We have |U 1 | 1 and |U | .
(b) The first i singular values of U 1 N U are |1 |, . . . , |i |.
(c) Either i = n or the (i + 1)th singular value of U 1 N U is at most .
Since |U 1 d(U )| |d| F , by Proposition 4.4.1 the new conditions
(b) and (c) hold when U 1 N U is replaced by U 1 N U + U 1 d(U ). We may
thus apply Lemma 6.7.3 to obtain the desired result.

6.8 A refined visible decomposition theorem

We give a refinement of the visible decomposition theorem (Theorem 6.6.1).
This can be used to obtain the TurrittinHukuharaLevelt decomposition
theorem; see Theorem 7.5.1. We first give an extension of Remark 6.2.13.
Lemma 6.8.1. Let F be a complete nonarchimedean differential field of char

acteristic 0. Suppose that P = i Pi T i F{T } and P = i Pi T i F{T }

are nonconstant twisted polynomials such that, for some s < log |d| F ,
the slopes of P and P are all equal to s. Put V = F{T }/F{T }P and

Let Q = i Q i U i F[T ] and Q = i Q i U i F[T ]

W = F{T }/F{T } P.
respecbe the untwisted polynomials having the same coefficients as P and P,
tively. Then |D|sp,V W < |D|sp,V if and only if for any root F alg of Q
and any root F alg of Q we have |/ 1| < 1.
Proof. As in the proof of Theorem 6.7.4, we may enlarge F to reduce to the
case where |D|sp,V = || for some F . Equip each of V and W with
the basis given by 1, 1 T, 2 T 2 , . . . , equip V with the dual basis, and use
these bases to define supremum norms. As in the proof of Theorem 6.5.3, we
have |D|V = |D|W = ||.

6.8 A refined visible decomposition theorem

113

The matrix of action of D on the resulting basis of V W has eigenvalues

By Theorem 6.7.4
of the form , where is a root of Q and is a root of Q.

we have |D|sp,V W < || if and only if each eigenvalue has norm strictly less
than ||, which occurs if and only if |/ 1| < 1 for all , .
Theorem 6.8.2 (Refined visible decomposition theorem). Let F be a complete nonarchimedean differential field of characteristic 0, and let V be a finite
differential module over F such that no subquotient of Vs has norm less than
or equal to |d| F . Then, for some finite tamely ramified extension F  of F, there
exists a (unique) decomposition
%
Vi
V F F  =
i

of V F F  into refined differential modules Vi , such that for i  = j we have

Vi  V j in the sense of Definition 6.2.12.
Proof. Uniqueness follows directly since is an equivalence relation (by
Lemma 6.2.14). To check existence, by Theorem 6.6.1 we may reduce to the
case where V is pure (see Definition 6.2.12). We may also assume that d  = 0,
as otherwise the result follows from Remark 6.2.13. Note that for any finite
tamely ramified extension F  of F we have |d| F  = |d| F by Lemma 6.2.15.
Apply Theorem 5.4.2 to choose an isomorphism V
= F{T }/F{T }P. By
Corollary 6.5.4, all the slopes of P are the same. Let Q F[U ] be the
untwisted polynomial with the same coefficients as P. By Remark 6.2.13, we
can choose a finite tamely ramified extension F  of F such that Q can be factorized in F  [U ] into factors Q i such that, for any two roots , of any Q i ,
we have |/ 1| < 1. By applying Theorem 2.2.2 (as in Remark 2.2.3) we
may correspondingly factorize P in F  {T } so that each factor corresponds to
a refined differential module. By performing this factorization again with the
residual roots in the opposite order and then arguing as in Theorem 6.6.1, we
obtain the desired splitting.
Remark 6.8.3. From the proof of Theorem 6.8.2 one may extract a map from
the classes of refined finite differential modules V with |D|sp,V = s > |d| F to
the quotient
{x o F alg : |x| s}
;
{x o F alg : |x| < s}
namely, given an isomorphism V
= F{T }/F{T }P, one associates with V the
class of the roots of the untwisted polynomial having the same coefficients
as P. However, one must check with some care that this map does not depend

114

Metric properties of differential modules

on the initial choice of cyclic vector. A better approach is to identify certain

test objects for which one can deduce the map more easily; this has been
done by Xiao [221].
From the proof of Theorem 6.8.2, we may also deduce the following simple but important observations. They will be used in the proofs of both
the TurrittinLeveltHukuhara decomposition (Theorem 7.5.1 below) and the
p-adic local monodromy theorem (Theorem 20.1.4).
Proposition 6.8.4. Let F be a complete nonarchimedean differential field of
characteristic 0. Let V be a finite differential module over F of rank n > 0
which is refined and has spectral radius s > |d| F .
(a) For any positive integer m which is nonzero in F , V m is again refined
with spectral radius s.
(b) The spectral radius of (n V ) V n is strictly less than s.
Proof. The result is elementary (as discussed in Remark 6.2.13) if d = 0, so
we may assume that d  = 0. By Theorem 5.4.2 we may choose an isomorphism V
= F{T }/F{T }P. Let Q be the untwisted polynomial with the same
coefficients as P, and let be a root of Q.
As in the proof of Theorem 6.7.4 we may enlarge F in such a way that
we reduce to the case where |D|sp,V = || for some F . Equip V with
the basis given by 1, 1 T, . . . , n+1 T n1 and the corresponding supremum
norm. Equip V m and (n V ) V n with the corresponding induced bases
and norms.
On V m the matrix of action of D has eigenvalues which are m-fold sums
of roots of Q. In particular, each eigenvalue satisfies |/(m) 1| < 1. We
can then deduce (a) from Theorem 6.7.4.
On (n V ) V n , the matrix of action of D has eigenvalues each of which
is an n-fold sum of certain roots of Q minus the sum of all the roots of Q. In
particular, each eigenvalue satisfies |/| < 1. Finally we deduce (b) from
Theorem 6.7.4.

6.9 Changing the constant field

We will now check that, in many cases, the operation of computing horizontal sections of a differential module commutes with the operation of forming
completed tensor products.
Proposition 6.9.1. Let R be a complete nonarchimedean differential domain
such that R and Frac(R) have the same constant ring R0 (which is necessarily
a complete nonarchimedean field). Let R0 be a complete field extension of R0 .

6.9 Changing the constant field

115

Let R  be the completed tensor product of R R0 R0 for the product norm
(which is a norm by Lemma 1.3.11). View R  as a differential ring by equipping
it with the unique continuous extension of d with R0 in its kernel. Then, for any
differential module M over R, the natural map
H 0 (M) R0 R0 H 0 (M R R  )
is an isomorphism of R0 -modules.
Proof. We first check injectivity. Since R0 is flat over R0 , the map H 0 (M) R0
R0 H 0 (M R0 R0 ) is bijective by Lemma 5.1.4. Since M is finite over
R, we may identify M R R  with the completed tensor product of M R
(R R0 R0 ) = M R0 R0 , on which the product seminorm is a norm by
Lemma 1.3.11. In particular M R0 R0 injects into M R R  , so the map
H 0 (M R0 R0 ) H 0 (M R R  ) is also injective.
We next check surjectivity in the case where R0 is the completion of a
finitely generated field extension S0 of R0 . Then S0 is of countable dimension
over R0 , so the hypothesis of Lemma 1.3.8 is satisfied by F = R0 , V = R0 .
Let m 1 , m 2 , . . . be the sequence of elements of R0 given by Lemma 1.3.8. As
in the proof of Lemma 1.3.11 these define projection maps j,R : R  R and
j,M : M R R  M; these maps are both horizontal.

Given an element x H 0 (M R R  ), choose a presentation x = sk=1 yk z k

with yk M and z k R  . Then
j,M (x) =

s


j,R (z k )yk

k=1

is an element of H 0 (M) for each j. Moreover, we can write x as a convergent sum

x=

j,M (x) m j .

j=1

Hence x lies in the closure of H 0 (M) R0 R0 under the product norm. However,
since H 0 (M) is finite-dimensional over R0 by Lemma 5.1.5, H 0 (M) R0 R0
is complete under the product norm, hence closed. Thus x H 0 (M) R0 R0
as desired.
We next check surjectivity in general. Pick any x H 0 (M R R  ). As noted
earlier, we may identify M R R  with the completion of M R0 R0 under the
product norm. We can thus choose a convergent series
x=


k=1

yk z k

(yk M, z k R0 ).

116

Metric properties of differential modules

Let R0 be the completion of the subfield of R0 generated over R0 by z 1 , z 2 , . . .

Then x is a horizontal element of the completion of M R0 R0 , so the previous
paragraph shows that x H 0 (M) R0 R0 . This proves the claim.

Notes
Lemma 6.2.5 is tacitly assumed at various places in the literature (including
by the present author) but we were unable to locate even an explicit statement, let alone a proof. We thank Liang Xiao for contributing the proof given
here.
Proposition 6.3.1 is modeled on some ongoing work of Baldassarri and Di
Vizio (a promised sequel to [11]), which gives a development of much of the
material we are discussing from the point of view of Berkovich analytic spaces.
This point of view will probably be vital for the study of differential modules
on higher-dimensional spaces. However, there are some key differences from
our approach; see the notes for Chapter 9.
Newton polygons for differential operators were considered by Dwork and
Robba [81, 6.2.3]; the first systematic treatment seems to have been given by
Robba [189]. Our treatment using Theorem 2.2.2 follows [42].
The proof of Theorem 6.5.3 given here is close to the original proof of Christol and Dwork [49, Thorme 1.5], save that we avoid a small logical gap in the
latter. The gap is in the implication 1  2; there one made a finite extension
of the differential field without accounting for the possibility that this might
increase |d| F to the point where Corollary 6.2.7 fails to show that |D|V is preserved. (It would be obvious that this does not occur if the finite extension were
being made in the constant subfield, but that was not the case.) Compare also
[80, Lemma VI.2.1].
Proposition 6.5.6 answers a conjecture of Christol and Dwork [48, Introduction, Conjecture A]. This conjecture was posed in the context of giving
effective convergence bounds, and that is one use to which we will put it here;
see Theorem 18.2.1 and its proof.
The use in Lemma 6.7.3 of a well-chosen norm (meaning a norm in which at
least some singular values match eigenvalues of the matrix of action of d) is an
extension of the notion of the canonical lattice (or DeligneMalgrange lattice)
introduced by Malgrange in the study of irregular meromorphic connections.
See [165].
The refined visible decomposition theorem is due to Liang Xiao [221]. It
was motivated by applications to refined Swan conductors of tale sheaves
and overconvergent isocrystals, as suggested by the work of Kato [119] in the
rank 1 case.

Exercises

117

We thank Andrea Pulita for the suggestion that Proposition 6.9.1 should be
included.

Exercises
(1) Prove Feketes lemma (Lemma 6.1.4).
(2) (a) Let A, B be commuting bounded endomorphisms on an abelian group
G equipped with a norm. Prove that
|A + B|sp,G |A|sp,G + |B|sp,G .
(b) Prove that if the norm on G is nonarchimedean then the inequality in
(a) can be improved to
|A + B|sp,G max{|A|sp,G , |B|sp,G }

(3)
(4)
(5)

(6)

(7)

and that equality occurs when the maximum is achieved only once.
(c) Prove that both these assertions may fail in the case where A and B
do not commute. (Hint: write an identity matrix as a sum of nilpotent
matrices.)
Prove that, in Lemma 6.2.5, if |D|sp,V > |d|sp,F then |Ds |1/s converges to
a limit as s . (Hint: again reduce to Lemma 6.1.4.)
Fill in the missing arguments in the proofs of Lemma 6.2.8(a), (b), using
Lemma 6.1.8.
Prove the claim from the proof of Lemma 6.4.2 that, for R a nonarchimedean differential domain, > |d| R , and P, Q R{T }, we have
|P Q Q P| 1 |d| R |P| |Q| . (Hint: reduce to the case where P
and Q are monomials and then to the case where P = T and Q R.)
Exhibit an example to show that, in Definition 6.4.3, if we had not omitted
slopes greater than or equal to log |d| R then it would no longer be the
case that the multiplicity of a slope r in a product P Q is the sum of the
multiplicities of r as a slope of P and as a slope of Q. (Hint: this can
already be seen using the t-adic valuation on Q(t) with d = d/dt.)
Let (F, d) be a complete nonarchimedean differential field. Let F be the
completion of F(t) for the -Gauss norm. Prove that there is a unique
continuous extension of d to F with d(t) = 0 and that it satisfies
|d| F = |d| F ,

|d|sp,F = |d|sp,F .

(Hint: first check everything on F[t].)

7
Regular singularities

In the next part of the book, which begins with Chapter 8, we will use the
results from the previous chapters to make a detailed analysis of ordinary
differential equations over nonarchimedean fields of characteristic 0, the motivating case being that of positive residual characteristic. However, before doing
so it may be helpful to demonstrate how the results apply in a somewhat
simpler setting.
In this chapter, we reconstruct some of the traditional Fuchsian theory of
regular singular points of meromorphic differential equations. (The treatment
is modeled on [80, 3].) We first introduce a quantitative measure of the
irregularity of a singular point. We then recall how, in the case of a regular singularity (i.e., a singularity with irregularity equal to zero), one has an
algebraic interpretation, using the notion of exponents, of the eigenvalues of
the monodromy operator around the singular point. We then describe how to
compute formal solutions of meromorphic differential equations and go on
to sketch the proof of Fuchss theorem, that the formal solutions of a regular
meromorphic differential equation actually converge in some disc. We finally
establish the TurrittinLeveltHukuhara decomposition theorem, which gives
a decomposition of an arbitrary formal differential module that is analogous to
the eigenspace decomposition of a complex linear transformation. The search
for an appropriate p-adic analogue of this result will lead us in Part V to the
p-adic local monodromy theorem.
Although this chapter focuses on issues different from much of the rest of
the book, it should not be considered entirely optional. It is referenced in the
discussions of p-adic exponents in Chapter 13 and of effective convergence
bounds in Chapter 18.
Hypothesis 7.0.1. Throughout this chapter, we will view C((z)) as a complete
nonarchimedean differential field with valuation given by the z-adic valuation
118

7.1 Irregularity

119

vz and derivation given by d = zd/dz; note that |d|C((z)) = 1. Let K be a field

of characteristic 0; although we do not need to think of K as a subfield of C
it is harmless to do so. (The reason is the Lefschetz principle: every statement
we make about K will refer to at most countably many elements of K , so each
individual instance of the statement can be realized over a subfield of K which
is countably generated over Q and hence embeds into C.)

7.1 Irregularity
Definition 7.1.1. Let V be a finite differential module over C((z)), and
decompose V according to Theorem 6.6.1. Define the irregularity of V as

( log s) dim(Vs ).
irr(V ) =
s>1

For F a subfield of C((z)) stable under d and V a finite differential module

over F, we define the irregularity of V to be the irregularity of V F C((z)).
We say that V is regular if irr(V ) = 0.
Theorem 7.1.2. For any isomorphism V
= F{T }/F{T }P, the irregularity of
V is equal to minus the sum of the slopes of P; consequently, it is always an

d1
integer. More explicitly, if P = T d + i=0
Pi T i then

irr(V ) = max 0, max{vz (Pi )} .
i

Proof. The second assertion follows from Corollary 6.5.4, since in this case
|d| F = |d|sp,F = 1. The first assertion follows from the second because, by
Theorem 5.4.2, V always admits a cyclic vector.
Theorem 7.1.2 gives rise to several criteria for regularity.
Corollary 7.1.3. Let F be any subfield of C((z)) containing z and stable
under d, and let V be a finite differential module over F. Then the following
conditions are equivalent.
(a) The module V is regular, i.e., irr(V ) = 0.
(b) For some isomorphism V
= F{T }/F{T }P with P monic, P has coefficients in o F .
(c) For any isomorphism V
= F{T }/F{T }P with P monic, P has coefficients in o F .
(d) There exists a basis of V on which D acts via a matrix over o F .
(e) For any basis B of V , the o F -span of the set {D i (v) : i {0, . . . ,
dim(V ) 1}, v B} is stable under D.

120

Regular singularities

Proof. By Theorem 7.1.2, (a) implies (c). It is obvious that (c) implies (b) and
that (b) implies (d). Given (d), let | |V be the supremum norm defined by the
chosen basis of V ; then |D|V 1, which implies (a).
This proves that (a), (b), (c), (d) are all equivalent. To add (e) to the circle of implications, on the one hand note that (e) implies (d). On the other
hand, given (a), pick any v B; then v, D(v), D 2 (v), . . . generate a differential submodule W of V for which v is a cyclic vector. Since V is regular,
its submodule W is also regular by Lemma 6.2.8; since (a) implies (c), the
o F -span of {D i (v) : i {0, . . . , dim(W ) 1}} is stable under D. This
implies (e).
Remark 7.1.4. One can also view C((z)) as a differential field with the derivation d/dz instead of zd/dz. The categories of differential modules for these
two choices of derivation are equivalent in an obvious fashion: given the action
zd/dz, we obtain the action d/dz by dividing by z. If V is a differential module
for zd/dz with spectral radius s then the spectral radius of V for d/dz is s|z|1
(exercise). The notion of irregularity translates naturally: for instance, if V is
a differential module for d/dz that is isomorphic to F{T }/F{T }P for some

n1
Pi T i then V is regular if and only if vz (Pi ) n + i for
P = T n + i=0
i = 1, . . . , n. For example, for a, b C, the differential system corresponding
to the hypergeometric differential equation
y  +

c (a + b + 1)z 
ab
y
y=0
z(1 z)
z(1 z)

is regular.
Example 7.1.5. For another example from the classical theory of ordinary
differential equations, consider the Bessel equation
y  +

1  z2 n2
y=0
y +
z
z2

for some parameter n. This gives rise to a regular differential system if we

expand around z = 0. However, if we expand around , i.e., substitute 1/z
for z and then expand around z = 0, we get an irregular differential system
(exercise).

7.2 Exponents in the complex analytic setting

To see why regular singularities are so important in the complex analytic
setting (by way of motivation for our p-adic studies), let us consider the
monodromy transformation. First we recall a familiar fact.

7.2 Exponents in the complex analytic setting

121

Theorem 7.2.1 (Cauchy). Fix > 0, and let R Cz be the ring of
power series convergent for |z| < . Let N be an n n matrix over R. Then
the differential system D(v) = N v + (d/dz)(v) has a basis of horizontal
sections.
Proof. This can be deduced from the fundamental theorem of ordinary differential equations; however, for future reference it will be useful to give a slightly
more detailed explanation. (Specifically, we are anticipating Definition 7.3.1.)
Note that there exists a unique n n matrix U over Cz such that U In

(mod z) and NU + (d/dz)(U ) = 0; this follows by writing N = i=0

Ni z i

i
and U = i=0 Ui z and then rewriting the equation N U + (d/dz)(U ) = 0 as
the recurrence
(i + 1)Ui+1 =

i


N j Ui j

(i = 0, 1, . . . ).

(7.2.1.1)

j=0

Following an argument of Cauchy [80, Appendix III], we may deduce that the
entries of U converge on a disc of positive radius, as follows. (In this argument,
we use the L 2 operator norm on matrices over C.) Pick any (0, ). Since
N converges in the open disc of radius , we have |Ni |i 0 as i ; in
particular, we may choose c > 1 with |Ni |i+1 c for all i. We then have by
induction on i that
|Ui |i ci

(i 0):

namely, this holds for j = 0, and (7.2.1.1) implies that

1 
|N j | j+1 |Ui j |i j ci+1 .
i +1
i

|Ui+1 |i+1

j=0

Consequently, the entries of U converge on the open disc of radius /c.

The previous argument applied to the system (N )U T +(d/dz)(U T ) = 0
shows that the entries of U 1 also converge on an open disc. Consequently,
we can find an open neighborhood of z = 0 on which the differential system admits a basis of horizontal sections. By translating we may derive the
same conclusion around any point of the original disc; since that disc is simply
connected, we obtain a basis of horizontal sections over the entire disc.
Remark 7.2.2. In the p-adic setting we will see that the first step of the proof
of Theorem 7.2.1 remains valid, but there is no analogue of the second step
(analytic continuation), and indeed the whole conclusion becomes false; see
Example 0.4.1.

122

Regular singularities

Let us now consider a punctured disc and look at monodromy.

Notation 7.2.3. In this chapter only, for K a subfield of C, let K {z} be the
subfield of K ((z)) consisting of formal Laurent series which represent meromorphic functions on some neighborhood of z = 0. The exact choice of that
neighborhood may vary with the series.
Definition 7.2.4. Let V be a finite differential module over C{z}; choose a
basis of V and let N be the matrix of action of D on this basis. On some disc
centered at z = 0, the entries of N are meromorphic with no poles away from
z = 0. On any subdisc not containing 0, by Theorem 7.2.1 we obtain a basis
of horizontal sections. If we start with a basis of horizontal sections in a neighborhood of some point away from 0 and then analytically continue around a
circle, proceeding once counterclockwise around the origin, we end up with a
new basis of local horizontal sections. The linear transformation taking the old
basis to the new one is called the monodromy transformation of V (or of its
associated differential system). The (topological) exponents of V are defined
(modulo translation by Z) to be the multiset of numbers 1 , . . . , n for which
e1 , . . . , en are the eigenvalues of the monodromy transformation.
The monodromy transformation controls our ability to construct global
horizontal sections, by the following statement whose proof is evident.
Proposition 7.2.5. In Definition 7.2.4, any fixed vector under the monodromy
transformation corresponds to a horizontal section defined on some punctured disc rather than on the universal covering space of a punctured disc.
As a result, the monodromy transformation is unipotent (i.e., the exponents are
all zero) if and only if there exists a basis on which D acts via a nilpotent
matrix.
Definition 7.2.6. In Definition 7.2.4, we say that V is quasiunipotent if its
exponents are rational; equivalently, the monodromy transformation of V
becomes unipotent after V is pulled back along z  z m for some positive
integer m. This situation arises for PicardFuchs modules; see Chapter 22.
Remark 7.2.7. The relationship between the properties of the monodromy
transformation and the existence of horizontal sections of the differential module begs the following question: is it possible to extract the monodromy
transformation for a differential module, whose definition is purely analytic, from the algebraic data that defines the differential system? The only
case in which this is straightforward is that of a regular module, which we
consider next.

7.3 Formal solutions of regular differential equations

123

7.3 Formal solutions of regular differential equations

One can formally imitate the proof of Cauchys theorem (Theorem 7.2.1) in
the regular case, as follows.

i
Definition 7.3.1. Let N =
i=0 Ni t be an n n matrix with entries in
K z. A fundamental solution matrix for N is an n n matrix U with U In
(mod z) such that U 1 N U + U 1 z(d/dz)(U ) = N0 .
Remark 7.3.2. Note that if U is a fundamental solution matrix for N then


d
T T T
T d
T
T T T
T T
T
U N U + U z (U ) = U N U U U
z (U ) U T
dz
dz


d
= U T N T U T z (U T ) U T
dz
= N0T .
That is, U T is a fundamental solution matrix for N T . Consequently, by
proving a general result about U we also obtain a corresponding result for
U T , and hence for U 1 .
To specify when a fundamental solution matrix exists, we need the following
definition.
Definition 7.3.3. We say that a square matrix N with entries in a field of
characteristic 0 has prepared eigenvalues if the eigenvalues 1 , . . . , n of N
satisfy the following conditions:
i Z i = 0,
i j Z i = j .
If only the second condition holds, we say that N has weakly prepared
eigenvalues.
We will also need the following lemma, which will come up again several
times later.
Definition 7.3.4. Let N be a nilpotent n n matrix over K . The nilpotency
index of N is the smallest positive integer e such that N e = 0.
Lemma 7.3.5. Let N1 and N2 be matrices of respective sizes m m and n n
over K . Let 1,1 , . . . , 1,m and 2,1 , . . . , 2,n be the eigenvalues of N1 and
N2 , respectively. Then the eigenvalues of the K -linear endomorphism X 
N1 X + X N2 on the space of m n matrices over K are equal to 1,i + 2, j for
i = 1, . . . , m, j = 1, . . . , n. Moreover, if N1 and N2 are themselves nilpotent,

124

Regular singularities

with respective nilpotency indices e1 , e2 , then the nilpotency index of X 

N1 X + X N2 equals e1 + e2 1.
Proof. There is no harm in enlarging K , so we may assume that it is algebraically closed. Moreover, if U1 GLm (K ) and U2 GLn (K ) then it is
equivalent to calculate the eigenvalues of the conjugated endomorphism
X  U11 (N1 (U1 XU21 )+(U1 XU21 )N2 )U2 = (U11N1 U1 )X +X (U21N2 U2 ).
Consequently, we may conjugate N1 and N2 into Jordan normal form. By separating X into blocks, we may reduce to the case where N1 and N2 consist of
single Jordan normal blocks with eigenvalues 1 and 2 . By subtracting 1 and
2 from the overall endomorphisms, we may reduce to the case 1 = 2 = 0,
for which we want to show that the map X  N1 X + X N2 is nilpotent with
nilpotency index at most e1 + e2 1. This can be done easily by hand but can
also be seen as follows: put f (X ) = N1 X +X N2 , and write the ith composition
of f as
i 

i
j
i j
f i (X ) =
N1 X N2 .
j
j=0

If i e1 + e2 1 then in each term we have either j e1 or i j e2 ,

so the whole sum vanishes.
i = e1 + e2 2 then similarly every term
 If2 2
N1e1 1 X N2e2 1 . Since we are assuming that
vanishes except possibly e1e+e
1 1


2 2
is nonzero and the
K has characteristic 0, the binomial coefficient e1e+e
1 1

matrices N1e1 1 , N2e2 1 must be nonzero, so we can make this product nonzero
by choosing a suitable matrix X .

Proposition 7.3.6. Let N = i=0

Ni t i be an n n matrix with entries in
K z such that N0 has weakly prepared eigenvalues. Then N admits a unique
fundamental solution matrix.
Proof. Let 1 , . . . , n K alg be the eigenvalues of N0 . Rewrite the defining

Ui t i and write
equation as NU +z(d/dz)(U ) = U N0 , then expand U as i=0
the new defining equation as a recurrence:
iUi = Ui N0 N0 Ui

i


N j Ui j

(i > 0).

(7.3.6.1)

j=1

Viewing the map X  X N0 N0 X as a linear transformation on the space

of n n matrices over K ((t)), we see by Lemma 7.3.5 that its eigenvalues
are the differences j k for j, k = 1, . . . , n. Likewise, the eigenvalues
of X  i X X N0 + N0 X are i j + k ; for i a positive integer, the

7.3 Formal solutions of regular differential equations

125

condition that the s are weakly prepared ensures that i j + k cannot

vanish (indeed, it cannot be an integer unless it equals i). Consequently, given
N and U0 , . . . , Ui1 there is a unique choice of Ui satisfying (7.3.6.1); this
proves the desired result.
Remark 7.3.7. Suppose that N is an n n matrix with entries in Cz whose
constant term has prepared eigenvalues. If we convert the differential system
defined by N into a differential module M, the C-span of the columns of the
fundamental solution matrix forms a C-submodule of M that is stable under
D. In fact, the action of D on M is the linear transformation defined by N0 .
Since this transformation may be nonzero, the elements of this span are not
all horizontal; however, one can show (exercise) that every horizontal element
of M does appear in the span. This justifies the name fundamental solution
matrix.
This formal argument becomes relevant in the complex analytic setting by
virtue of the following informal fact. For a proof see [80, III.8, Appendix II]
or the exercises.

Ni t i be an n n matrix with entries

Theorem 7.3.8 (Fuchs). Let N = i=0
in C{z} such that N0 has weakly prepared eigenvalues. Then the fundamental solution matrix for N over Cz also has entries in C{z} (as does its
inverse).
Corollary 7.3.9. With notation as in Theorem 7.3.8, let 1 , . . . , n be the
eigenvalues of N0 . Then the eigenvalues of the monodromy transformation of
the system D(v) = N v + d(v) are e2i1 , . . . , e2in .
Proof. In terms of a basis on which the matrix of action of D is N0 , the matrix
exp(N0 log(z)) provides a basis of horizontal elements. (The case N0 = 0 is
given by Theorem 7.2.1.)
In order to enforce the condition on prepared eigenvalues, we use what are
classically known as shearing transformations.
Proposition 7.3.10 (Shearing transformations). Let N be an n n matrix
over K z whose entries have constant term N0 . Let 1 , . . . , m K alg be
eigenvalues of N forming a single Galois orbit over K . Then there exists U
GLn (K [z, z 1 ]) such that U 1 N U + U 1 d(U ) also has entries in K z, and
its matrix of constant terms has the same eigenvalues as N0 except that every
instance of each i has been replaced by i + 1. The same conclusion holds if
i 1 replaces i + 1.
Proof. Exercise.

126

Regular singularities

Corollary 7.3.11 (Fuchs). Let V be a regular finite differential module over

C{z}. Then any horizontal element of V C{z} C((z)) belongs to V itself; that
is, any formal horizontal section is convergent. (This is false in the irregular
case; see the notes.)
Proof. Using Proposition 7.3.10 we may construct a basis of V on which D
acts via a matrix in Cz C{z} whose constant term has weakly prepared
eigenvalues. By Remark 7.3.7, any horizontal element of V C{z} C((z)) is a
C-linear combination of column vectors of the fundamental solution matrix;
by Theorem 7.3.8, any such linear combination converges in a disc.
To put everything together, we need the following result.
Proposition 7.3.12. Let V be a regular finite differential module over K ((z)).
Then there exists a basis of V on which the matrix of action of d has entries in
K and prepared eigenvalues.
Proof. By Proposition 6.5.6 we can find a K z-lattice L in V (i.e., a finitely
generated K z-submodule whose K ((z))-span is the whole of V ) which is
stable under d. By Proposition 7.3.10 we can modify L in such a way that the
constant term of the matrix of action of d on some basis of L has prepared
eigenvalues. (Because we cannot separate Galois conjugates, the previous
statement relies on the fact that no two distinct Galois conjugates , over
K may differ by a nonzero integer. This fact holds because the trace from the
Galois closure of K (, ) to K vanishes on but is injective on K because
K is of characteristic 0.) We may then apply Proposition 7.3.6 to deduce the
claim.

7.4 Index and irregularity

We mention very briefly an alternate interpretation of irregularity considered
by Malgrange.
Definition 7.4.1. Let F be any subfield of K ((z)) stable under d, and let V be
a finite differential module over F. We say V has an index if dim K H 0 (V ) and
dim K H 1 (V ) are both finite; in this case we define the index of V as (V ) =
dim K H 0 (V ) dim K H 1 (V ).
Proposition 7.4.2. For any finite differential module V over K ((z)),
dim K H 0 (V ) = dim K H 1 (V ) < and so (V ) = 0.
Proof. Exercise.

7.5 The TurrittinLeveltHukuhara decomposition theorem

127

In the convergent case, the index carries more information.

Theorem 7.4.3. Let V be a finite differential module over C{z}. Then V has
an index, and (V ) = irr(V ).
Proof. See [164, Thorme 2.1].

7.5 The TurrittinLeveltHukuhara decomposition theorem

One can classify differential modules over K ((z)) rather simply, provided that
one is willing to admit finite field extensions. Note that the results do not
descend from C((z)) to C{z}, owing to the failure of Corollary 7.3.11 in the
irregular case. Nonetheless, even a formal classification is extremely useful;
see the notes.
Note the use of the refined visible decomposition theorem (Theorem 6.8.2)
in the following proof.
Theorem 7.5.1. Let V be a finite differential module over K ((z)). Then there
exist a positive integer h and a finite extension K  of K such that V K ((z))
K  ((z 1/ h )) admits a direct sum decomposition i Vi in which each Vi Vi is
regular.
Proof. We induct on the dimension of V . Let v be a generator of n V , write
D(v) = sv, and let W be the differential module over K ((z)) having one
generator w satisfying D(v) = n 1 sw. Note that W V cannot be refined,
as otherwise Proposition 6.8.4(a) would imply that W V has the same
spectral radius as n (W V )
= (W )n n V , which is trivial.

By Theorem 6.8.2, either W V is regular or refined or, for suitable h, K  ,

(W V ) K ((z)) K  ((z 1/ h )) splits as a nontrivial direct sum. Since we have
ruled out the refined case, either W V is regular, and so V V is regular,
or we may invoke the induction hypothesis.
Remark 7.5.2. If we insist that the decomposition in Theorem 7.5.1 be minimal (i.e., that it should have as few summands as possible), then it is in fact
unique. Consequently, such a minimal decomposition must be respected by
any extra structures on V which respect the action of d. For instance, if K carries one or more derivations, and we equip V with actions of these derivations,
then the decomposition is preserved by these actions.
From the proof of Theorem 7.5.1 we can immediately read off the following
classification of the components appearing in the direct sum decomposition
therein.

128

Regular singularities

Theorem 7.5.3. Let V be a finite differential module over K ((z)) such that
V V is regular. Then there exists a differential module W of rank 1 such
that W V is regular.
Corollary 7.5.4. Suppose that K is algebraically closed. Let V be a finite
differential module over K ((z)) such that V V is regular. Then there is a
unique decomposition V = i Vi such that Vi V j has all exponents zero if
i = j and all exponents nonzero if i  = j.
Proof. If V itself is regular, the corollary follows from Proposition 7.3.12 and
the fact that a linear transformation over an algebraically closed field decomposes as a direct sum of generalized eigenspaces. In general we may replace
V with W V for some V as specified in Theorem 7.5.3 and reduce to the
regular case.
The strongest form of the TurrittinLeveltHukuhara decomposition is
the following statement, which both eliminates the base extension in
Theorem 7.5.1 and incorporates the statement of Theorem 7.5.3.
Definition 7.5.5. Let h be a positive integer, and suppose that P =
{P1 , . . . , Ph } is a Galois orbit over K ((z)). Let F be a finite Galois extension of K ((z)) containing P1 , . . . , Ph . Then the differential module of rank h
over F with generators e1 , . . . , eh satisfying
d(ei ) = Pi ei

(i = 1, . . . , h)

descends uniquely to a differential module over K ((z)), which we denote

E(P). When the orbit is a singleton, {P1 }, we also write E(P1 ) as shorthand
for E({P1 }). (See the proof of Theorem 7.5.6 for more on the Galois descent
construction.)
Theorem 7.5.6. Let V be a finite differential module over K ((z)). Then V
admits a direct sum decomposition
%
V =
E(Pi ) X i
i

for some Galois orbits Pi and some regular differential modules X i .

Proof. Suppose first that the conclusion of Theorem 7.5.1 applies with
K  ((z 1/ h )) = K ((z)). Then Theorem 7.5.3 implies that V = E(P) X for
some P K ((z)) and some regular differential module X . In fact, there
is a unique choice of P z 1 K [z 1 ] for which such an X exists, and the
decomposition is also unique.

Notes

129

In general, we get a decomposition of the desired form over K  ((z 1/ h ))

for some finite extension K  of K and some positive integer h. We can
deduce the desired assertion from this using a Galois descent argument. Put
G = Gal(K  ((z 1/ h ))/K ((z))). For G, note that we can base change along
(as in Definition 5.3.2) to define V .
As above, we choose each Pi to consist entirely of elements of
1/
z h K  [z 1/ h ]. For each i and each G, we must then have (Pi ) = P j
for some j. From the uniqueness of the decomposition, we get an isomorphism
: (E(Pi ) X i ) E(P j ) X j .
These isomorphisms satisfy a cocycle condition: for , G, we have
= ( ) .
We may canonically identify E(Pi ) with E(P j ) by matching up the generator w 1 with the generator w. We may then canonically identify
with a horizontal element of E(P j ) E(P j ) X i X j . By projecting onto the trace component of E(P j ) E(P j ) we get a horizontal element
of X i X j , which in turn we identify with a morphism  : X i X j .
These maps again satisfy the cocycle condition, so Galois descent allows us
to identify each X i with the base extension of a differential module X i over
K ((z)). Under this identification, if Pi denotes the Galois orbit of Pi , we get a
canonical identification
%
E(P j ) X j .
(E(Pi ) X i ) K ((z)) K  ((z 1/ h ))
=
j:P j Pi

This proves the claim.

Remark 7.5.7. The proof of Theorem 7.5.6 shows that both the direct sum
decomposition and the identification of the tensor factors in each summand
can be made canonical, by insisting that each Pi consist of elements of
z 1/ h K  [z 1/ h ] for some finite extension K  of K and some positive integer h. However, everything still depends implicitly on the choice of the series
parameter z in the field K ((z)); in many applications, such a choice is not at
all obvious.

Notes
The notion of a regular singularity was introduced by Fuchs in the nineteenth century as part of a classification of those differential equations
with everywhere-meromorphic singularities on the Riemann sphere which
have algebraic solutions. Regular singularities are sometimes referred to as

130

Regular singularities

Fuchsian singularities. Much of our modern understanding of the regularity condition, especially in higher dimensions, comes from the book of
Deligne [68].
As noted in the text, Corollary 7.3.11 is false for irregular modules. This was
originally noticed in numerous examples of particular differential equations
(e.g., the Bessel equation at infinity; see Example 20.2.1) and motivated the
definition of irregularity in the first place.
The relationship between the interpretations of irregularity using the Newton
polygon and using indices is due to Malgrange [164]. Our treatment, in which
the spectral radius plays a pivotal role, is based on [80, 3]; this point of view
is ultimately due to Robba.
A complex analytic interpretation of the Newton polygon, in the manner
of the relation between irregularity and index, was given by Ramis [186]. It
involves considering subrings of C{z} composed of functions having certain
extra convergence restrictions (Gevrey functions) and looking at the index of
z d/dz after tensoring the given differential module with one of these subrings.
Theorem 7.5.6 is a slight reformulation of classification results due to
Turrittin [212] (building on earlier work of Hukuhara in the rank 1 case) for the
existence aspect and to Levelt [158] for the uniqueness aspect. See [7, II.3.1]
for further discussion; we further recommend [7] for higher-dimensional
analogues of the results discussed in this chapter. See [144] for some additional higher-dimensional analogues, presented in much the same style as in
this chapter.

Exercises
(1) Let V be a differential module over K ((z)) with spectral radius s for
zd/dz. Prove that the spectral radius of V for d/dz is s|z|1 .
(2) Verify the claims of Example 7.1.5.
(3) Let N be an n n matrix over K z whose constant term has prepared
eigenvalues. Prove that any vector v K ((z))n with N v +z(d/dz)(v) = 0
is a K -linear combination of columns of the fundamental solution matrix
of N . (Hint: change basis first.)

i
(4) Prove Fuchss theorem (Theorem 7.3.8). (Hint: let U =
i=0 Ui z be
the fundamental solution matrix of N . Relate the norms of Ui and iUi
Ui N0 + N0 Ui by considering the singular values of the map X  X
X N0 + N0 X . Then use (7.3.6.1) to obtain


i



Ui Ui N0 N0 Ui  1
|N j ||Ui j | max {|N j ||Ui j |}

 i
1 ji
i
j=1

and proceed as in the proof of Theorem 7.2.1.)

Exercises

131

(5) Prove Proposition 7.3.10. (Hint: show that, for any space V of constant
vectors stable under multiplication by N0 , the set of vectors with entries
in K z whose reductions modulo z belong to V are closed under the
operation v  N v + d(v). Then take V to be a union of generalized
eigenspaces.)
(6) Prove Proposition 7.4.2. (Hint: change basis by preparing eigenvalues and
then applying Theorem 7.3.8.)
(7) Let Q K ((z))[U ] be a polynomial whose roots 1 , . . . , n K ((z))alg
satisfy |i / j 1| < 1 for all i, j. Prove that there exists K ((z)) such
that |i / 1| for all i. Then show that this fails without the hypothesis
that K has characteristic 0.

Part III
p-adic Differential Equations on Discs
and Annuli

8
Rings of functions on discs and annuli

In Part III we focus our attention specifically on p-adic ordinary differential

equations (although most of our results apply also to complete nonarchimedean fields of residual characteristic 0). To do this with maximal generality,
one would need first to introduce a category of geometric spaces over which to
work. This would require a fair bit of discussion of either rigid analytic geometry, in the manner of Tate, or nonarchimedean analytic geometry in the manner
of Berkovich, neither of which we want either to assume or introduce. Fortunately, since we only need to consider one-dimensional spaces, we can manage
by working completely algebraically and considering differential modules over
appropriate rings.
In this chapter, we introduce those rings and collect their basic algebraic
properties. This includes the fact that they carry Newton polygons analogous to those for polynomials. Another key fact is that there is a form of the
approximation lemma (Lemma 1.3.7) valid over some of these rings.
Notation 8.0.1. Throughout this part, let K be a field of characteristic 0 that
is complete for a nontrivial nonarchimedean norm | |. (The assumption of
characteristic 0 is not used in this chapter; it will become crucial when we start
discussing differential modules again.) Let p denote the characteristic of the
residue field K . We do not assume p > 0 (as the case p = 0 may be useful for
some applications), but when p > 0 we do require the norm to be normalized
in such a way that | p| = p 1 .
Definition 8.0.2. By a piecewise affine function on an interval I we will mean
a continuous function f : I R such that I can be covered with intervals
of positive length, on each of which f restricts to an affine function (i.e., a
function of the form ax + b for some a, b R). By the compactness of a
closed interval (see Lemma 8.0.4), it is equivalent to require that each point of I
135

136

Rings of functions on discs and annuli

admit a one-sided neighborhood, on each side within I , on which f restricts to

an affine function. (If I is infinite, we allow the possibility of having infinitely
many different slopes unless we say otherwise explicitly.)
Example 8.0.3. An example specifically excluded by Definition 8.0.2 is the
function f : [0, 1] R defined by

x = '0,
1/2
(
f (x) = 1/2 1/(N + 1) + N /(N + 2) (1/N x) x 1/(N + 1), 1/N ,

N = 1, 2, . . . ;
this function is piecewise affine on (0, 1] and continuous on [0, 1], but there is
no one-sided neighborhood of 0 on which f is affine.
Lemma 8.0.4. Let I = [, ] be a (bounded) closed interval. Let S be a set
of closed subintervals of I which cover every one-sided neighborhood of every
point in I . (That is, for any [, ), there exist ( , ] and J S
such that [ , ] J and, similarly, for any (, ], there exist [, ]
and J S such that [, ] J .) Then there exists a finite subset of S with
union I .
Proof. Exercise.

8.1 Power series on closed discs and annuli

We start by introducing some rings that should be thought of as the analytic
functions on a closed disc |t| or a closed annulus |t| . As noted
above, this is more properly done in a framework of p-adic analytic geometry,
but nevertheless we choose to avoid this framework.
Definition 8.1.1. For , > 0, put


ci t i K t, t 1  : lim |ci | i = 0
K /t, t/ =
iZ


( [, ]) .

That is, consider the set of formal bidirectional power series which converge
whenever one inserts a value for t with |t| [, ] or, in other words, when
/|t| and |t|/ are both at most 1; it suffices to check for = and =
. Although formal bidirectional power series do not form a ring, the subset
K /t, t/ does form a ring under the expected operations. (In this and all
similar notation, we will omit when it is equal to 1.)
Definition 8.1.2. If = 0, the only reasonable interpretation of the previous
definition is to require ci = 0 for i < 0. When there are no negative powers

8.1 Power series on closed discs and annuli

137

of t, it is redundant to require convergence for < . In other words, we shall

define



i
i
K 0/t, t/ = K t/ =
ci t K t : lim |ci | = 0 .
i

i=0

One can also write K /t, for which the implied value of is . More
succinctly put, we identify K /t, t/ with K  1 /t 1 , t 1 / 1 . We do
not allow both and to equal 0, as this case is exceptional: no condition is
then imposed on the power series except that negative indices cannot occur.
This results in the ring K t.
Remark 8.1.3. Note that if then, within K  /t, t/,
K /t, t/ K  /t, t/ = K /t, t/.
This will serve as the basis for some gluing arguments later.
Definition 8.1.4. We will also occasionally use the intermediate ring



i
i
K t/0 =
ci t K t : sup{|ci | } < ;
i

i=0

these are the power series which converge and take bounded values on the open
disc |t| < . (The notation will make more sense once we have also defined
K t/ for > 0; see Definition 18.4.1.) Note that, for any (0, ),
K t/ K t/0 K t/.
We will most often use this construction with = 1, in which case we can
also write
K t0 = o K t oK K .
Definition 8.1.5. An analogue of the previous construction for an annulus is



i
i
i
K /t, t/0 =
ci t : ci K , lim |ci | = 0, sup{|ci | } < ;
iZ

these are the Laurent series which converge and take bounded values on the
half-open annulus |t| < . For any [, ), this ring satisfies
K /t, t/ K /t, t/0 K /t, t/.
One can also use the boundedness condition on both sides, defining



i
i
i
ci t : ci K , sup{|ci | } < , sup{|ci | } < .
K /t, t/0 =
iZ

138

Rings of functions on discs and annuli

Remark 8.1.6. The rings of bounded series behave well only in the case where
K is discretely valued; otherwise, they are not even noetherian (exercise). For
general K it is better to work with rings of analytic elements; see Section 8.5.

8.2 Gauss norms and Newton polygons

The rings K /t, t/ behave quite like polynomial rings (or Laurent polynomial rings, in the case  = 0) in one variable. The next few statements are all
instances of this analogy.
Definition 8.2.1. From the definition of K /t, t/, we see that it carries a
well-defined -Gauss norm





i
ci t  = max{|ci | i }



i
i

for any [, ]. For = = 0, this reduces to simply |c0 |. The fact that
this is a multiplicative norm follows as in Proposition 2.1.2. In the additive
version (Gauss valuation) one takes r [ log , log ] and puts



i
vr
ci t = min{v(ci ) + ri},
i

where v(c) = log |c|. There is also a -Gauss norm on K /t, t/0 ,
although it must be defined as a supremum that may fail to be achieved if
K is not discretely valued.
Definition 8.2.2. One may define the Newton polygon for an element

x = i xi t i K /t, t/ as the boundary of the lower convex hull of the set
{(i, v(xi )) : i Z, xi  = 0},
retaining only those slopes within [ log , log ].

Proposition 8.2.3. Let x = i xi t i K /t, t/ be nonzero.

(a) The Newton polygon of x has finite width.
(b) The function r  vr (x) on [ log , log ] is continuous, piecewise
affine, and concave. Moreover, even if = 0 there are only finitely
many different slopes.
(c) The function  |x| on [, ] is continuous and log-convex. The
log-convexity means that if , [, ], c [0, 1], and = c 1c
then
|x| |x|c |x|1c
.

8.2 Gauss norms and Newton polygons

139

(d) If = 0 then vr (x) is increasing on [ log , +); in other words,

for all [0, ], |x| |x| .
Part (c) should be thought of as a nonarchimedean analogue of the
Hadamard three-circle theorem.
Proof. We have (a) because there is a least i for which |ci | i is maximal, and
there is a greatest j for which |c j | j is maximal. This implies (b) because,
as in the polynomial case, we may interpret vr (x) as the y-intercept of the
supporting line of the Newton polygon of slope r . This in turn implies (c), and
(d) is a remark made earlier.
Remark 8.2.4. The analogue of Proposition 8.2.3 for x K /t, t/0 also
holds except that if K is not discretely valued then vr (x) need not be piecewise
affine in a one-sided neighborhood of r = log (exercise).
When dealing with the ring K /t, t/, the following completeness
property will be extremely useful.
Proposition 8.2.5. The rings K /t, t/ and K /t, t/0 are Frchet complete for norms | | for all [, ]. That is, if {xn }
n=0 is a sequence which
is simultaneously Cauchy under | | for all [, ] then it is convergent.
(By Proposition 8.2.3 it suffices to check the Cauchy property at = , , or
just at = in the case = 0.)
Proof. Exercise.
The completeness property is used in the construction of multiplicative
inverses, for instance.
Lemma 8.2.6
(a) A nonzero element f K t/ is a unit if and only if there exists c K
such that | f c| < | f | .
(b) A nonzero element f K /t, t/ is a unit if and only if there exist
c K and i Z such that | f ct i | < | f | for = , .
(c) If < then a nonzero element f K /t, t/0 is a unit if and only
if there exist c K and i Z such that | f ct i | < | f | for all
[, ).
Proof. If f is a unit then the Newton polygon of f has no slope in [, ];
this implies the forward implication in all cases. For the reverse implication in
cases (a) and (b), it suffices to check that if | f 1| < | f | , for = in
case (a) or for = , in case (b), then f is a unit. This holds because the
series

140

Rings of functions on discs and annuli


(1 f ) j

(8.2.6.1)

j=0

converges by Proposition 8.2.5, and its limit is the inverse of f .

For the reverse implication in case (c), we note that the series (8.2.6.1) converges in K /t, t/ for each [, ). Moreover, the terms of the sum have
bounded -norm, so the limit does also.

8.3 Factorization results

We need a number of results to the effect that elements of one ring that we
are considering can be factored into positive and negative parts. The basic
result of this form may be viewed as a form of the Weierstrass preparation
theorem.
Proposition 8.3.1 (Weierstrass preparation). Assume one of the following
two sets of conditions.

(a) Put R = K /t, t/ or R = K /t, t/0 . Given f = iZ f i t i

R and [, ], suppose that there is a unique m Z maximizing
| f m | m .

(b) Put R = K /t, t/0 and put = . Given f = iZ f i t i R,

suppose that the supremum of | f i | i is achieved by at least one i (this
is only guaranteed to hold if K is discretely valued), and let m be the
least such i.
Then there is a unique factorization f = f m t m gh, with
g R K t,

h R K t 1 ,

such that |g| = |g0 | = 1 and |h 1| < 1.

Proof. As in Theorem 2.2.1, we invoke the master factorization theorem
(Theorem 2.2.2). This gives a factorization of the desired form in the completion of R with respect to | | .
However, within this completion, define the subring R  = K /t, t/ if
R = K /t, t/ or R  = K /t, t/0 if R = K /t, t/0 . Then h is a
unit in R  K t 1  = K /t and hence in R  . We thus have g = f h 1
R  K t = R K t. We may similarly deduce that h R K t 1 .
In light of the finite-width property of the Newton polygon, the following should not be a surprise. (One can only replace K /t, t/ with
K /t, t/0 if K is discretely valued; see the exercises.)

8.3 Factorization results

141

Proposition 8.3.2 (More Weierstrass preparation). For f K /t, t/

with < +, there exists a polynomial P K [t] and a unit g
K /t, t/ such that f = Pg. In particular, K /t, t/ is a principal
ideal domain (even if = +).
Proof. If > 0 then we may apply Proposition 8.3.1(b) to f since the Newton polygon has only finitely many slopes. We may thus factor f in R =
K /t, t/0 as f m t m gh with f m K , m Z, g R K t = K t/0 ,
h R K t 1  = K /t, |g| = |g0 | = 1, and |h 1| < 1. In particular
h is a unit in K /t by Lemma 8.2.6, so
h 1 f K /t, t/ t m K t/0 t m K t/.
If = 0, the same conclusion holds with h = 1, m = 0.
Next, factor h 1 f in K  1 /t 1 , t 1 / 1 0 and argue as above, to get
an associate element of f in K /t, t/ which belongs to t m K t/
t n K /t for some integer n. This element must now belong to K [t] if = 0
or to K [t, t 1 ] if > 0; in either case we may deduce the claim.
We next wish to generalize the previous considerations to matrices, but this
will require a bit more care.
Lemma 8.3.3. Let A be an invertible n n matrix over K /t, t/. Then A
can be factored as a product of elementary matrices.
Proof. We first note that we can perform a sequence of elementary row operations after which A11 becomes a unit and Ai1 becomes zero for i = 2, . . . , n.
Namely, using Proposition 8.3.2, multiply each row by a suitable unit to replace
each Ai1 by an element of K [t]. Then perform the Euclidean algorithm in K [t]
on the Ai1 to achieve the desired result.
To finish, it suffices to note that we can now perform the usual GaussJordan
elimination over A: repeatedly apply the previous paragraph to construct a
sequence of row operations putting A into upper triangular form and then
perform column operations to eliminate entries above the diagonal.
Lemma 8.3.4. Let A be an invertible n n matrix over K t/ (resp. over
K /t, t/.) Then there exists U GLn (K [t]) (resp. U GLn (K [t, t 1 ]))
such that |U A In | < 1.
Proof. Note that in the notation of Definition 4.0.3, conjugating an elementary matrix of type (c) (scaling a row) by one of type (a) (swapping rows)
produces another of type (c), whereas conjugating an elementary matrix of
type (b) (adding a multiple of one row to another) by one of type (c) produces
another of type (b). Consequently, any sequence of elementary row operations

142

Rings of functions on discs and annuli

has the same effect as another sequence in which all steps of type (a) and (b)
happen before all steps of type (c).
By Lemma 8.3.3 we can perform a sequence of elementary row operations
on A that produces the identity matrix. By this, plus the previous paragraph,
we can perform another sequence of elementary row operations, of types
(a) and (b), on A that produces a diagonal matrix D. By appending additional
row operations of type (c) with c K (resp. c K t Z ), we can force
|D In | < 1.
Let R1 , . . . , Rm be elementary matrices as above with R1 Rm A = D;
note that those R j of type (a) and (c) are already elementary matrices over
K [t] (resp. over K [t, t 1 ]), the case of type (c) being handled by the previous
paragraph. Put
= max{1, | A|1 }

m


max{1, |R j |1 }.

j=1

R j

be an elementary matrix of type (b) over K [t] (resp.

For R j of type (b), let
1
over K [t, t ]) with |R j R j | < 1 ; for R j not of type (b), put R j = R j .
 A I | < 1, proving the claim.
Then |R1 Rm
n
We can now deduce a Weierstrass preparation theorem for matrices.
Proposition 8.3.5. Let A be an invertible n n matrix over K /t, t/. Then
there exist U GLn (K t/) and V GLn (K /t) such that A = U V .
Proof. Pick any [, ]. By Lemma 8.3.4, we can find U1
GLn (K [t, t 1 ]) such that |U1 A In | < 1. By applying the master factorization theorem (Theorem 2.2.2) in the (noncommutative) ring of n n matrices
over K /t, t/ and then arguing as in Proposition 8.3.1, we can factor U1 A
as U2 V2 with U2 GLn (K t/) and V2 GLn (K /t). We may then set
U = U11 U2 and V = V2 .
Our main application of Proposition 8.3.5 is the following gluing lemma,
which we will invoke frequently and often implicitly.
Lemma 8.3.6 (Gluing lemma). Suppose that . Let M1 be
a finite free module over K /t, t/, let M2 be a finite free module over
K  /t, t/, and suppose that we are given an isomorphism
: M1 K  /t, t/
= M2 K  /t, t/.
Then we can find a finite free module M over K /t, t/ and isomorphisms
M1
= M K /t, t/, M2
= M K  /t, t/ inducing . Moreover, M
is determined by this requirement up to unique isomorphism.

8.4 Open discs and annuli

143

Proof. We will explain only the case > 0; the case = 0 is similar. Choose
bases v1,1 , . . . , v1,n and v2,1 , . . . , v2,n of M1 and M2 , respectively. Let A be
the change-of-basis matrix from the v1,i to the v2,i , viewing both as bases of
M1 K  /t, t/
= M2 K  /t, t/ via . By Proposition 8.3.5 we can
factor A as U V with U GLn (K t/) and V GLn (K  /t).
We can then construct a finite free module M over K /t, t/ equipped
with a basis v1 , . . . , vn such that the change-of-basis matrices from this basis
to the v1,i and to the v2,i are U 1 and V , respectively. This is the desired
module.
Remark 8.3.7. If < , one can similarly glue together a finite
free module over K /t, t/ and a finite free module over K  /t, t/0 ,
whose base extensions to K  /t, t/ are isomorphic; the result is a finite
free module over K /t, t/0 . As we will not use this fact, we omit further
details.

8.4 Open discs and annuli

Although we have been discussing closed discs so far, it is quite natural also
to consider open discs. One important reason is that the antiderivative of an
analytic function on the closed disc of radius is only defined on the open
disc of radius (see exercises for Chapter 9).
Definition 8.4.1. Define the ring


ci t i : ci K , lim |ci | i = 0
K {t/} =
i

i=0


( (0, )) ;

these are the power series convergent on the open disc |t| < , with no
boundedness restriction. Note that we can write
)
K t/;
K {t/} =
(0,)

in particular, for any (0, ),

K t/0 K {t/} K t/.
Definition 8.4.2. An analogue of the previous construction for an annulus is

K /t, t/} =

iZ

ci t i : ci K , lim |ci | i = 0, lim |ci | i = 0

i

i+

( (0, )) ;

these are the Laurent series convergent on the half-open annulus |t| < .

144

Rings of functions on discs and annuli

These rings behave worse than their closed counterparts (see the exercises),
so we will only make occasional use of them. More often, we will work
with the following definition, which is motivated by considerations from rigid
analytic geometry.
Definition 8.4.3. Consider the region |t| I , for any interval I [0, +);
this could be a closed or open disc or a closed, open, or half-open annulus.
By a coherent locally free module M on this region, we will mean a sequence
of finite free modules Mi over K i /t, t/i , with [1 , 1 ] [2 , 2 ]
an increasing sequence of closed intervals with union I , together with isomorphisms Mi+1 K i /t, t/i 
= Mi . Using Lemma 8.3.6 we check that the
construction is canonically independent of the choice of sequence. However,
the resulting object is not finite in the sense of admitting a finite generating set
over the entire annulus, except when K is spherically complete. For a special
case, see Proposition 16.1.4; the notes give further discussion.

8.5 Analytic elements

An intermediate construction between open and closed discs is the ring
K t/0 of bounded power series but, as noted previously, it behaves badly
if K is not discretely valued. Another intermediate construction that behaves
somewhat better is the following.
Definition 8.5.1. Define the ring K t/an by starting with the subring of
K (t) consisting of rational functions with no poles in the disc |t| < and then
completing for the -Gauss norm. This is the ring of analytic elements on the
open disc |t| < ; it satisfies
K t/ K t/an K t/0 .
Analogously, we define the ring K /t, t/an of analytic elements on the
half-open annulus |t| < as follows. Start with the subring of
K (t) consisting of rational functions with no poles in the annulus
|t| < . Then take the Frchet completion for the -Gauss norms with
.
One may also define the ring K /t, t/an of analytic elements on the
open annulus < |t| < . For this, we start with the subring of K (t) consisting of rational functions with no poles in the annulus < |t| < and then
again take the Frchet completion for the -Gauss norms with . In
the case = this construction gives a field; we will see this field again in
Definition 9.4.1.

8.5 Analytic elements

145

For analytic elements, we have analogues of many properties asserted for

analytic functions.
Proposition 8.5.2. Let R be K t/an , K /t, t/an , or K /t, t/an .
(a) For any x R, the function r  log |x|er is continuous and
concave in r .
(b) For any x R, the Newton polygon of x has finite width.
(c) Any x R can be written as a polynomial P K [t] times a unit in R.
(d) The ring R is a principal ideal domain.
Proof. Since we are taking the Frchet completion for norms over a closed
interval, the convergence of any Cauchy sequence must be uniform in the different norms. We may thus deduce (a) from the corresponding assertions when
x K (t) with no poles in the appropriate disc or annulus.
We will check (b) for K t/an ; the other cases are analogous. Let

xi t i K t/an be any nonzero ring element. We can then choose

x = i=0
a rational function f K (t), with no poles in the disc |t| < , such that
|x f | < |x| . By (a), |x| is continuous in , so we also have |x f | < |x|
for in a neighborhood of . Consequently, those slopes of x sufficiently close
to occur with the same multiplicities as the corresponding slopes of f . But
f is a rational function, so it has no slopes in some punctured neighborhood
of . This proves that x has no slopes in some punctured neighborhood of
either, so its Newton polygon has finite width.
To check (c) and (d), we may use the same proof as in Proposition 8.3.2.
Corollary 8.5.3. The ring < K /t, t/an is a field.
Proof. By Proposition 8.5.2 every element of K /t, t/an can be written as
a unit in K /t, t/an times a polynomial. It thus suffices to observe that, for
any P K [t], we can choose so that none of the roots of P lie in the annulus
|t| < ; for such an , P is a unit in K /t, t/an .
We will use the following optimal approximation property.
Lemma 8.5.4. Let R be K t/, K /t, t/, or K t/an . Let F be the
completion of Frac(R) under | | . Then, for any f F, there exists g R
minimizing | f g| .
Proof. We may assume f
/ R, as otherwise g = f fulfills the lemma. Put
c = inf{| f g| : g R}. Since R is complete under | | we have c > 0.
Since K (t) is dense in F we can choose h K (t) such that | f h| < c.
Let P K [t] be the (monic) denominator of h. By Theorem 2.2.1 we may
factor P as P1 P2 in such a way that no irreducible factor of P1 in K [t] is a

146

Rings of functions on discs and annuli

unit in R whereas every irreducible factor of P2 in K [t] is a unit in R. (The

point is that each irreducible polynomial has only a single slope in its Newton
polygon, and whether the polynomial is a unit is determined entirely by this
slope.) Since the claims for f and f P2 are equivalent, we may assume that
P2 = 1. Using the division algorithm, write h P = g P + S with g, S K [t]
and deg(S) < deg(P). We claim that this choice of g works; that is, for any
g  R we have | f g  | | f g| .
Suppose first that R = K t/; in this case R K [t] is dense in R. Given
g  R, we may choose g  K [t] such that |g  g  | < c. Applying
Lemma 2.3.1 to the instance (h g  )P = (g g  )P + S of the division
algorithm yields
|(h g  )P| = max{|(g g  )P| , |S| } |S| .
Therefore |h g  | |S/P| = |h g| ; since | f h| , |g  g  | < c
|h g| , we also have | f g  | | f g| as claimed.
Suppose next that R = K /t, t/; in this case R K [t, t 1 ] is dense
in R. Given g  R, we may choose g  K [t] and m 0 such that |g 
g  t m | < c. Using the division algorithm to divide (ht m g  )P by P now
returns the same remainder as does dividing t m S by P. We may argue as in the
previous case once we have checked that, for S K [t] with deg(S) < deg(P),
the remainder S  upon dividing t S by P satisfies |S  | = |S| . We proceed
as in the proof of Lemma 2.3.1, as follows. Put d = deg(P) and let Sd1
be the coefficient of t d1 in S. If |S| = |Sd1 t d1 | then the constant term
of S  is equal to the remainder upon dividing Sd1 t d by P, which has norm
|S| . Otherwise |S  t (S Sd1 t d1 )| < |S| , so the claim follows
again.
Finally, suppose that R = K t/an . Given g  R, we may choose g  , Q
K [t] such that Q is monic with all roots of norm and |g  g  /Q| < c.
Using the division algorithm to divide (h Q g  )P by P now returns the same
remainder as does dividing Q S by P. If we denote the latter by S  , we may
now argue as in the previous cases once we have shown that |S  | = |S| .
Put d = deg(P); in this case, all the roots of P have norm less than , so
|P t d | < 1. Hence |S  | = |S  | , where S  is the remainder upon dividing
Q S by t d . But it is clear that |S  | = |S| : we may check this after dividing S
and t d by any common factors of t, at which point |S| and |S  | are achieved
as the respective constant terms of S and S  .
Remark 8.5.5. Note that the proof of Lemma 8.5.4 actually shows something
a bit stronger: the constructed element g R continues to minimize | f g|
even if we replace K with a complete extension L.

8.6 More approximation arguments

147

Remark 8.5.6. If K is discretely valued then the conclusion of Lemma 8.5.4

also holds for R = K t0 and = 1, since we can then write F as the
completion of K t0 [t 1 ].

8.6 More approximation arguments

We now give some variants of the approximation lemmas (Lemmas 1.3.7 and
1.5.5) that involve rings of power series.
Lemma 8.6.1. Let R be K t, K 1/t, t, K tan , or (if K is discretely valued)
K t0 equipped with the 1-Gauss norm. Let F be the completion of Frac(R)
under | |1 . Let M be a finite free R-module. Let | | M be a norm on M,
compatible with R, obtained by restriction from the supremum norm defined
by some basis of M R F. Then | | M is the supremum norm defined by some
basis of M.
Proof. (Thanks to Liang Xiao for his help with this proof.) Let N be the
o F -span of a basis of M R F whose supremum norm restricts to | | M . Put
R1 = {r R : |r |1 1} and M1 = {x M : |x| M 1}. Note that
m K R1 coincides with the subring of R consisting of series whose coefficients
all belong to m K . (If R = K t0 this holds only if K is discretely valued.)
Hence the ring R1 /m K R1 is either K t or a localization of K [t], so in any
case it is a principal ideal domain.
Note that M1 /m K M1 embeds into N /m K N , so in particular it is torsionfree as a module over R1 /m K R1 . Since Frac(R) is dense in F, we can choose
a basis y1 , . . . , yn of N which is also a basis of M R Frac(R). We can then
find an f R which is nonzero and such that M contains f y1 , . . . , f yn ;
by multiplying by a unit in K , we may normalize f so that | f |1 = 1.
Then f y1 , . . . , f yn project to elements of M1 /m K M1 which form a basis of
N /m K N over F .
We can also find a g R which is nonzero and such that M is contained
in the R-span of g 1 y1 , . . . , g 1 yn , which we again normalize so that |g|1 =
1. This means that M1 /m K M1 is contained in the (R1 /m K R1 )-submodule of
N /m K N generated by g 1 y1 , . . . , g 1 yn . Since R1 /m K R1 is a principal ideal
domain, it follows that M1 /m K M1 is finitely generated, torsion-free, and hence
finite free as a module over R1 /m K R1 . By the previous paragraph, any basis
of M1 /m K M1 over R1 /m K R1 freely generates N /m K N over F and hence
freely generates N over o F .
Let x1 , . . . , xn M1 lift a basis of M1 /m K M1 over R1 /m K R1 . For any
x M1 we have a unique representation x = r1 x1 + + rn xn with

148

Rings of functions on discs and annuli

r1 , . . . , rn o F . By Lemma 8.5.4 and Remark 8.5.6 we may choose ri R to

maximize |ri ri |1 . Choose K so that || = maxi {|ri ri |1 }; if  = 0
then the image of 1 (x r1 x1 rn xn ) in M1 /m K M1 fails to be in
the (R1 /m K R1 )-span of the images of x1 , . . . , xn . This yields a contradiction,
so we must have = 0 and ri = ri R o F = R1 for i = 1, . . . , n.
Consequently, x1 , . . . , xn form a basis of M with supremum norm | | M , as
desired.
It is more difficult to deal with the case of K t0 when K is not discretely
valued. Although we will not use that case in what follows, for completeness
we mention one result that applies to it.
Lemma 8.6.2. Suppose that K is spherically complete with value group R.
Let M be a finite free module over K t/0 for some > 0. Let | | M be a
supremum-equivalent norm on M that is compatible with K t/0 . For j =
1, 2, . . . , let | | j be the quotient seminorm on M/t j M induced by | | M .
Suppose that, for any m M,
|m| M = lim |m| j .

(8.6.2.1)

Then | | M is the supremum norm for some basis of M.

Proof. By rescaling we reduce immediately to the case = 1. Put R = K t0
and R1 = o K t, so that R1 is the set of r R having norm at most 1.
Let M1 be the set of x M having |x| M 1; then M1 is a R1 -submodule
of M. Since | | M is compatible with R, for any positive integer j we have
M1 t j M = t j M1 .
By assumption we can find a basis m 1 , . . . , m n of M which determines a
supremum norm equivalent to | | M . That is, there exist c1 , c2 > 0 such that,
for any a1 , . . . , an R,
c1 max{|ai | R } |a1 m 1 + + an m n | M c2 max{|ai | R }.
i

As in the proof of Lemma 1.3.5 it follows that, for any positive integer j and
any a1 , . . . , an R/t j R,
c1 max{|ai | R/t j R } |a1 m 1 + + an m n | j c2 max{|ai | R/t j R }. (8.6.2.2)
i

In particular, each | | j is a norm.

By Lemma 1.5.5, | |1 is the supremum norm defined by some basis
m 1,1 , . . . , m n,1 of M/t M. By Lemma 1.5.4 (applied by viewing each M/t j M
as a K -vector space), for i = 1, . . . , n and j = 1, 2, . . . we can construct
m i, j+1 M/t j+1 M lifting m i, j M/t j M and satisfying |m i, j+1 | j+1 =
|m i, j | j .

Notes

149

For each i, the inverse limit of the m i, j determines an element m i of M R

K t. However, by (8.6.2.2), if we write m i as a K t-linear combination of
m 1 , . . . , m n then each coefficient is a power series bounded in norm by c11 so
m i M. By (8.6.2.1), |m i | M = 1.
 m +
For any x M with |x| M 1, in each M/t j M we can write x = a1,
j 1



j
+ an, j m n with a1, j , . . . , an, j R/t R. We will prove by induction on j
 |, . . . , |a  | 1. This is true for j = 1 because the m , . . . , m were
that |a1,
1
n
n, j
j
chosen so that the supremum norm they define matches | |1 . Given the claim
for j 1, lift each ai, j1 R1 /t j1 R1 to bi, j R1 /t j R1 by making the

last coefficient 0, so that the norm is preserved. Then t 1 j (x i bi, j m i ) is

an element of M/t M of norm at most 1, so it is an o K -linear combination of
m 1 , . . . , m n . This completes the induction.
We thus obtain a representation x = a1 m 1 + + an m n with a1 , . . . , an
R1 . Since any element of M can be written as an element of K times an element
of M of norm at most 1, we deduce that m 1 , . . . , m n is a basis of M and that
the supremum norm defined by this basis coincides with | | M .
Question 8.6.3. Is condition (8.6.2.1) necessary?

Notes
The Hadamard three-circles theorem (Proposition 8.2.3(c)) is a special case of
the fact that the Shilov boundary of the annulus |t| consists of the two
circles |t| = and |t| = . For a considerable amplification of this remark,
including a full-blown theory of harmonic functions on Berkovich analytic
curves, see [205]. For an alternative presentation, restricted to the Berkovich
projective line but otherwise more detailed, see [12].
The gluing lemma (Lemma 8.3.6) is a special case of the gluing property of coherent sheaves on affinoid rigid analytic spaces, as specified in the
theorems of Kiehl and Tate [31, Theorems 8.2.1/1 and 9.4.2/3]. The factorization argument in the proof, however, is older still; it is the nonarchimedean
version of what is called a Birkhoff factorization over an archimedean field.
Similarly, Definition 8.4.3 corresponds to the definition of a locally free coherent sheaf on the corresponding rigid or Berkovich analytic space. Such a
sheaf is only guaranteed to be freely generated by global sections in the
case where K is spherically complete [128, Theorem 3.14]; in fact, a previous result of Lazard [156] implies that this property, even when restricted
to modules of rank 1, is in fact equivalent to the spherical completeness
of K .
We again thank Liang Xiao for his help with the proof of Lemma 8.6.1.

150

Rings of functions on discs and annuli

Exercises
(1) Prove Lemma 8.0.4. (Hint: for each point of I find an open neighborhood covered by one or two elements of S. Then reduce to the usual
compactness property of a bounded closed interval.)
(2) Verify the assertions of Remark 8.2.4. (Hint: a typical example where

1/n t n .)
piecewise affinity fails is
n=1 p
(3) Prove Proposition 8.2.5. (Hint: it may be easiest to first construct the limit
using a single [, ] and then show that this can also be done for the
other .)
(4) Prove that if K is discretely valued then, for any f K /t, t/0 , there
exists a polynomial P K [t] and a unit g K /t, t/
0 such that
f = Pg. (Hint: the Newton polygon has finite width in this case, so one
may argue as in Proposition 8.3.2.)
(5) Prove that the ring K {t} is not noetherian. (Hint: pick a sequence of points
in the open unit disc converging to the boundary, and consider the ideal of
functions vanishing on all but finitely many of these points.)
(6) Prove that if K is not trivially or discretely valued then K t0 is not noetherian. (Hint: proceed as in the previous exercise but choose the points so
that the Newton polygon of a function vanishing on all the points has finite
height.)
(7) Prove that if K is discretely valued then o K t = o K t K t is
noetherian. Otherwise it is not, because then o K itself is not noetherian.
(8) Prove that each maximal ideal of o K t is generated by m K together with
some P o K [t] whose reduction modulo m K is irreducible in K [t].
(9) State and prove an analogue of the gluing lemma (Lemma 8.3.6) for gluing
together finite free modules over K 1/t, tan and K t0 , using an isomorphism over the completion of K t0 K [t] K (t) under | |1 , to obtain a
module over K tan .

9
Radius and generic radius of convergence

In this chapter, we begin to approach a fundamental question peculiar to the

study of nonarchimedean differential modules. It was pointed out in Chapter 0
that a differential module over a nonarchimedean disc can fail to have horizontal sections even in the absence of singularities. The radius of convergence
of local horizontal sections is thus an important numerical invariant, whose
control is a key factor in being able to produce solutions of p-adic differential
equations.
Unfortunately, the radius of convergence is often difficult to compute
directly. One of Dworks fundamental insights is that one can get much better control over the radius of convergence around a so-called generic point.
The properties of the generic radius of convergence can then be used to infer
information about the actual convergence of horizontal sections. For instance,
Dworks transfer theorem asserts that the radius of convergence of a differential module over a nonarchimedean disc is no less than the generic radius of
convergence at the boundary of the disc.
However, both the radius of convergence and the generic radius of convergence are rather coarse invariants. Just as the notion of the spectral radius
is refined by the notion of the full spectrum, we can introduce subsidiary
radii of convergence and subsidiary generic radii of convergence, which detect
whether some local horizontal sections at a point converge further than others. We will devote much effort in the remainder of this part of the book to
analyzing the behavior of these refined invariants.
Hypothesis 9.0.1. Throughout this chapter we will view K /t, t/,
K /t, t/an , and so forth as differential rings with derivation d = d/dt,
formal differentiation in the variable t, unless otherwise specified (as in
Section 9.6).
151

152

Radius and generic radius of convergence

9.1 Differential modules have no torsion

We start with some simple but critical observations about the categories of
differential modules over the rings of power series introduced in Chapter 8.
Lemma 9.1.1. Let R be one of the following rings: K t/, K t/an ,
K /t, t/, K /t, t/an , K /t, t/an , or (if K is discretely valued)
K t/0 or K /t, t/0 . Then R has no nonzero proper differential ideal.
Proof. If I is a nonzero differential ideal then, by Proposition 8.3.2 (or
Proposition 8.5.2), I contains a nonzero element P K [t]. But then I
also contains d deg(P) (P), which is a nonzero element of K (because K is of
characteristic 0), so I must be the unit ideal.
Proposition 9.1.2. For R as in Lemma 9.1.1, any finite differential module
over R is free. Consequently, the finite free differential modules over R form
an abelian category.
Proof. Let M be a finite differential module over R. If m M is annihilated
by the nonzero element r R then 0 = D(r m) = r D(m) + d(r )m, and so
D(m) is annihilated by r 2 . Consequently the torsion submodule T of M is also
a differential module. Also, T is finite over R because R is a principal ideal
domain (by either Proposition 8.3.2 or Proposition 8.5.2) and hence noetherian.
Hence the annihilator I of T is a nonzero ideal of R. It is also a differential
ideal: if r I then, for any m T , 0 = D(r m) = r D(m) + d(r )m = d(r )m,
so d(r ) I . By Lemma 9.1.1, I must be the trivial ideal. Hence T = 0, so M
is torsion-free; since R is a principal ideal domain, M must also be free.
Corollary 9.1.3. For R as in Lemma 9.1.1 and for M a finite differential module over R, any finite set of horizontal sections which is linearly independent
over K forms part of a basis of M.
Proof. Let S be a finite set of horizontal sections which is linearly independent
over K . By Lemma 5.1.5, S is also linearly independent over R. In this case, S
determines an injective morphism from a trivial differential module to M. By
Proposition 9.1.2 the image of this map must be a direct summand of M as an
R-module; this yields the desired result.
Corollary 9.1.4. For R as in Lemma 9.1.1, let M be a finite differential module over R, of rank n and admitting a set S of n horizontal sections linearly
independent over K . Then M is trivial and H 0 (M) is the K -span of S.
Since we also wish to deal with open discs and annuli, we must formally
define differential modules on them.

9.2 Antidifferentiation

153

Definition 9.1.5. Consider the region |t| I , for any interval I [0, +);
this region could be a closed or open disc or a closed, open, or half-open annulus. By a finite differential module M on this region we will mean a coherent
locally free module in the sense of Definition 8.4.3, in which each module Mi
carries the structure of a differential module over K i /t, t/i  and each isomorphism Mi+1 K i /t, t/i 
= Mi is horizontal. By Proposition 9.1.2 we
see that these again form an abelian category. (It would probably be better to
call these coherent differential modules since they need not be generated by a
finite set of sections.)

9.2 Antidifferentiation
One initially surprising fact about p-adic analysis is that the relationship
between differentiation and boundary convergence is reversed from the archimedean case: whereas d carries K /t, t/ into itself, antidifferentiation
does not. Instead, one has merely the following.

Lemma 9.2.1. For any x = i xi t i K {/t, t/} with x1 = 0, there exists

y K {/t, t/} for which d(y) = x.
Proof. Exercise.
Corollary 9.2.2. Let M be the trivial differential module over K t/ (resp.
over K /t, t/). Then, for any with 0 < (resp. any , with
< < ), the map H 1 (M) H 1 (M K t/) (resp. H 1 (M)
H 1 (M K  /t, t/)) is the zero map.
This gives us an explicit description of unipotent differential modules.
(Recall that these are the successive extensions of trivial modules.)
Lemma 9.2.3. Let M be a finite unipotent differential module for the derivation t (d/dt) over K /t, t/ with 0 < < . Then, for any , with
< < , M K  /t, t/ admits a basis on which the matrix of
action of D is nilpotent and has entries in K .
Proof. Let v1 , . . . , vn be a basis of M such that, letting Mi denote the span
of v1 , . . . , vi , we have that Mi is stable under D and Mi /Mi1 is trivial for
i = 1, . . . , n. We proceed by induction on n; that being said, we may assume
that the matrix of action of D on v1 , . . . , vn1 is upper triangular nilpotent
with entries in K .
/ K , we can choose
We now write D(vn ) = c1 v1 + + cn1 vn1 . If cn1
 ,  with <  < <  < and then (by Lemma 9.2.1) choose

154

Radius and generic radius of convergence

x K   /t, t/   such that cn1 t (d x/dt) K . If we replace vn by vn =



vn1 , with cn1
K . Similarly, if
vn xvn1 then D(vn ) = c1 v1 + + cn1
/ K , we can change basis to bring cni into K
cn1 , . . . , cni+1 K but cni
while possibly changing c1 , . . . , cni1 but not cni+1 , . . . , cn1 . Repeating
this process, we ultimately obtain a basis of the desired form.

9.3 Radius of convergence on a disc

Definition 9.3.1. Let M be a finite differential module over K t/, K t/an ,
or K t/0 . Define the radius of convergence of M around 0, denoted R(M),
to be the supremum of the set of (0, ) such that M K t/ has a basis
of horizontal elements; we refer to those elements as local horizontal sections
of M. For M a finite differential module on the open disc of radius around
t = 0, define R(M) as the supremum of R(M K t/ ) over all < . For
, note that


R(M),
R(M K t/ ) =
R(M) > R(M).
Example 9.3.2. In general, if p > 0 then it is possible to have R(M) < ;
that is, there is no p-adic analogue of the fundamental theorem of ordinary
differential equations (as was noted in Example 0.4.1). For instance, consider the module M = K t/v with D(v) = v; for > p 1/( p1) we
have R(M) = p 1/( p1) because that is the radius of convergence of the
exponential series. (This is essentially Example 0.4.1 again.)
However, the local form of the fundamental theorem of ordinary differential
equations has the following analogue.
Proposition 9.3.3 ( p-adic Cauchy theorem). Let M be a finite differential
module over K t/, K t/an , or K t/0 . Then R(M) > 0.
Proof. By shrinking we reduce to the case over K t/. One can give
a direct proof of this, but instead we will deduce it from Dworks transfer
theorem (Theorem 9.6.1 below). We will give a direct proof of a slightly
stronger result later (Proposition 18.1.1); see also the notes.
Here are some easy consequences of the definition of the radius of convergence; note the parallels with properties of the spectral radius (Lemma 6.2.8).
Lemma 9.3.4. Let M, M1 , M2 be finite differential modules over K t/,
K t/an , or K t/0 .

9.4 Generic radius of convergence

155

(a) If 0 M1 M M2 0 is exact then

R(M) = min{R(M1 ), R(M2 )}.
(b) We have
R(M ) = R(M).
(c) We have
R(M1 M2 ) min{R(M1 ), R(M2 )},
with equality when R(M1 )  = R(M2 ).
Proof. For (a), it is clear that R(M) min{R(M1 ), R(M2 )}; we must check
that equality holds. Choose < min{R(M1 ), R(M2 )}, so that M1 K t/
and M2 K t/ are both trivial. For any  < , the map H 1 (M2 M1 )
H 1 ((M2 M1 ) K t/ ) is zero by Corollary 9.2.2, so M K t/  is trivial
by Lemma 5.3.3. Since we can make and  as close to min{R(M1 ), R(M2 )}
as we like, we find that R(M) min{R(M1 ), R(M2 )}.
For (b), we obtain R(M ) R(M) from the fact that if M K t/ is
trivial, then so is its dual M K t/. Since M and M enter symmetrically,
we get R(M ) = R(M).
For (c), the inequality is clear from the fact that the tensor product of two
trivial modules over K t/ is also trivial. The last assertion follows as in the
proof of Lemma 6.2.8(c).
Example 9.3.5. Assume that p > 0, and let M be the differential module of
rank 1 over K t/ defined by D(v) = v with K . Then it is an exercise
to show that
R(M) = min{, p 1/( p1) ||1 }.
A special case is provided by an important example of Dwork; see Example 17.1.4.

9.4 Generic radius of convergence

In general, the radius of convergence of a differential module is difficult to
compute. To make its computation easier, we introduce a related but simpler
invariant.
Definition 9.4.1. For > 0, let F be the completion of K (t) under the
-Gauss norm | | . Put d = d/dt on K (t); then d extends by continuity to
F , and we have

156

|d| F =

Radius and generic radius of convergence

1

|d|sp,F = lim |n!|

1/n 1


1
p 1/( p1) 1

p = 0,
p > 0.

(See Proposition 9.10.2 below for a refinement of this assertion.) It is a

common convention to define

=

p = 0,

p 1/( p1)

p > 0,

so that we may write |d|sp,F = 1 .

Remark 9.4.2. Note that F coincides with the ring K /t, t/an of analytic elements on the circle |t| = (see Definition 8.5.1). As a result, F is
commonly known as the field of analytic elements of norm .
We will also make a related construction in the case = 1. For a way to
unify the two constructions, see Definition 9.10.1.
Definition 9.4.3. Put E = K 1/t, t0 or, in other words, let E be the completion of o K ((t)) oK K for the 1-Gauss norm | |1 . If K is discretely valued,
the supremum in the Gauss norm is achieved; consequently E is a field and
its residue field is equal to K ((t)). However, none of this applies unless K
is discretely valued. (This is the same issue that arises in Remark 8.2.4.) In
any case E is complete under | |1 , there is an isometric map F1 E carrying t to t, and the supremum is achieved for elements of E in the image
of that map; this at least gives an embedding of K ((t)) into the quotient
{x E : |x|1 1}/{x E : |x|1 < 1}.
Definition 9.4.4. Let (V, D) be a nonzero finite differential module over F
or E. We define the generic radius of convergence (or for short, the generic
radius) of V to be
R(V ) = |D|1
sp,V ;
note that R(V ) > 0. We will see later (in Proposition 9.7.5) that this does
indeed compute the radius of convergence of horizontal sections of V on a
generic disc. (If V is the zero module, set R(V ) = .)
Remark 9.4.5. Note that the map F1 E is isometric, and |d|sp,E =
= |d|sp,F1 . Consequently, for any finite differential module V over F1 ,
Corollary 6.2.7 implies that |D|sp,V E = |D|sp,V .

9.5 Some examples in rank 1

157

We can translate some basic properties of the spectral radius (Lemma 6.2.8)
into properties of generic radii of convergence, leading to the following analogue of Lemma 9.3.4. Alternatively, one can first check Proposition 9.7.5 and
then simply invoke Lemma 9.3.4 itself around a generic point.
Lemma 9.4.6. Let V, V1 , V2 be nonzero finite differential modules over F .
(a) For an exact sequence 0 V1 V V2 0,
R(V ) = min{R(V1 ), R(V2 )}.
(b) We have
R(V ) = R(V ).
(c) We also have
R(V1 V2 ) min{R(V1 ), R(V2 )},
with equality when R(V1 )  = R(V2 ).
Definition 9.4.7. In some situations it is more natural to consider the intrinsic
generic radius of convergence, or for short the intrinsic radius, defined as
I R(V ) = 1 R(V ) =

|d|sp,F
(0, 1].
|D|sp,V

(The upper bound of 1 comes from Lemma 6.2.4.) To emphasize the difference,
we may refer to the unadorned generic radius of convergence defined earlier
as the extrinsic generic radius of convergence. (See Proposition 9.7.6 and the
notes for some reasons why the intrinsic radius deserves its name.)
Remark 9.4.8. For I an interval, M a differential module on the annulus
|t| I , and I , it is unambiguous to refer to the generic radius of convergence R(M F ) of M at radius . This is defined by first making a base
change to K /t, t/ for some closed subinterval [, ] of I containing .
Since the resulting module M F does not depend on the choice of [, ],
neither does its generic radius.

9.5 Some examples in rank 1

Assume p > 0 for these examples. See also Example 9.9.3.
Example 9.5.1. In Example 9.3.5 we have D s (v) = s v for all nonnegative
integers s. By Lemma 6.2.5 we have
R(M F ) = min{, p 1/( p1) ||1 } = R(M).

158

Radius and generic radius of convergence

Another important class of examples is given as follows.

Example 9.5.2. For K , let V be the differential module of rank 1 over
F defined by D(v) = t 1 v. It is an exercise to show that I R(V ) = 1 if and
only if Z p .
We can classify Example 9.5.2 further as follows.
Proposition 9.5.3. We have V
= V if and only if  Z.
Proof. Note that V
= V if and only if V is trivial, so we may reduce
/ Z p ; by
to the case  = 0. By Example 9.5.2, V is nontrivial whenever
direct inspection, V is trivial whenever Z.
It remains to deduce a contradiction, assuming that V is trivial, Z p ,
and
/ Z. There is no harm in enlarging K now, so we may assume that K
contains a scalar of norm ; by rescaling, we may reduce to the case = 1.
We now have f F1 such that t (d f /dt) = f ; by multiplying by an element
of K we can force | f |1 = 1.
Let 1 be an integer such that 1 (mod p). Then


 d( f t 1 ) 

 = |( 1 ) f t 1 1 |1 p 1 .


dt
1

i
Using the embedding F1 E, we may expand f =
iZ f i t with
maxi {| f i |} = 1. The previous calculation then forces | f i | p 1 unless
i 1 (mod p).
By considering the reduction of f modulo p n and arguing similarly, we
/ Z,
find that | f i | p 1 unless i (mod p n ) for all n. But, since
this means that the image of f in K ((t)) cannot have any terms at all, a
contradiction.

9.6 Transfer theorems

A fundamental relationship between the radius of convergence and the generic
radius of convergence is given by the following theorem. In the language of
Dwork, it is a transfer theorem, because it transfers convergence information from one disc to another. (Note that Proposition 9.3.3, which asserts that
R(M) > 0, is an immediate corollary.)
Theorem 9.6.1 (Dwork). For any nonzero finite differential module M over
K t/ or K t/an , R(M) R(M F ). That is, the radius of convergence
is equal to at least the generic radius.

9.6 Transfer theorems

159

Proof. Suppose that < and < |D|1

sp,M . Fix a supremum norm | | M
on M that is compatible with | | . We claim that, for any x M, the Taylor
series
y=


(t)i
i=0

i!

D i (x)

(9.6.1.1)

converges under || M . To see this, pick  > 0 such that 1 (|D|sp,M + ) < 1.
By Lemma 6.1.8 there exists c > 0 such that |D i (x)| M c(|D|sp,M +
)i |x| M for all i. The ith term of the sum defining y thus has norm at most
i i c(|D|sp,M + )i |x| M , which tends to 0 as i .
By differentiating the series expression as in Remark 5.8.4 we find that
D(y) = 0, so y is a horizontal section of M K t/. If we run this construction over a basis of M, we obtain horizontal sections of M K t/ whose
reductions modulo t also form a basis; these sections are thus K -linearly independent and so form a basis of M K t/ by Proposition 9.1.2. This proves
the claim.
Here is a simple example of how one may apply the transfer theorem; it is
conditional on one result, Theorem 11.3.2.
Example 9.6.2. Recall the hypergeometric differential equation
y  +

c (a + b + 1)z) 
ab
y
y = 0,
z(1 z
z(1 z)

considered in Chapter 0. In general, one solution in the ring K ((z)) is given by

the hypergeometric series
F(a, b; c; z) =


a(a + 1) (a + i)b(b + 1) (b + i)
i=0

c(c + 1) (c + i)i!

zi .

Let us now restrict to the case a, b, c Z p Q. Let m be the denominator of

c. In the ring K ((z 1/m )), the general solution is
AF(a, b; c; z) + Bz 1c F(a + 1 c, b + 1 c; 2 c; z)

(A, B K ),
(9.6.2.1)

and it converges for |z| < 1 (see the exercises for Chapter 0).
We now pass to the associated differential module M of rank 2 which is
defined over the ring K /z, zan for any > 0. From (9.6.2.1) we see that,
for any (0, 1), M K /z, z/ has a filtration in which one quotient is a
trivial module and the other has the form D(w) = z 1 w for some Z p Q.
From Example 9.5.2 and Lemma 9.4.6 we deduce that R(M F ) = 1 for

160

Radius and generic radius of convergence

(0, 1). By Theorem 11.3.2(a), R(M F ) is continuous at = 1 so that

R(M F1 ) = 1 also.
We now expand in power series around another value, z 0 o K , that is not
congruent to 0 or 1 modulo m K . We get another differential module N over
K z z 0 an such that M F1
= N F1 and so, in particular, R(N F1 ) = 1.
By Theorem 9.6.1, R(N ) = 1; that is, the general solution of the hypergeometric differential equation at z = z 0 converges in the disc |z z 0 | < 1.
Remark 9.6.3. In Example 9.6.2 one cannot directly prove that R(N F1 ) = 1
using Theorem 6.5.3, because of the limitation on slopes therein. An alternate
approach that works is to construct a Frobenius structure on N ; see Part V.

9.7 Geometric interpretation

As promised, here is a construction that explains the terminology generic
radius of convergence.
Definition 9.7.1. Let L be a complete extension of K . A generic point of L
of norm relative to K is an element t L, with |t | = , such that there is
no z K alg with |z t | < . For instance, t itself is a generic point of F of
norm .
Remark 9.7.2. If t L is a generic point then evaluation at t gives an
isometry K [t] L for the -Gauss norm on K [t]. To see this, it suffices to
check after replacing K by a completed algebraic closure and enlarging L to

contain this enlarged K . Then any P K (t) can be factored as c i (t z i )
for some z i K , and for each i we have |z i t | because t is a generic
point. Consequently,

|P(t )| = |c|
|t z i |
i

= |c|

max{, |z i |}

= |c|

|t z i |

= |P| .
Definition 9.7.3. Let L be a complete extension of K . For any t L with
|t | = , the substitution t  t + (t t ) induces an isometric map
K [t] L(t t )/. However, if (and only if) t is a generic point then the
composition of this map with the reduction modulo t t is again an isometry,

9.7 Geometric interpretation

161

by Remark 9.7.2. Hence in this case the map K [t] L(t t )/ extends
to an isometry F L(t t )/an .
Remark 9.7.4. In Berkovichs theory of nonarchimedean analytic geometry,
the geometric interpretation of the above construction is that the analytic space
corresponding to F is obtained from the closed disc of radius by removing
the open disc of radius centered around each point of K alg . As a result, it still
contains any open disc of radius that does not meet K alg .
Proposition 9.7.5. Let V be a finite differential module over F , and put V  =
V F L(t t )/an . Then the generic radius of convergence of V is equal
to the radius of convergence of V  .
Proof. Let G be the completion of L(t t ) for the -Gauss norm; then the
map F G is an isometry for any < . By Corollary 6.2.7,
|D|sp,V G = max{|d|sp,G , |D|sp,V } = max{1 , |D|sp,V }.
On one hand, this implies R(V ) R(V  ) by the application of Theorem 9.6.1
to V L(t t )/ for a sequence of values of converging to .
On the other hand, pick any < R(V  ); then V G is a trivial differential
module, so the spectral radius of D on it is 1 . We thus have
|D|sp,V 1 ,
so R(V ) . This yields R(V ) R(V  ).
Here is an example illustrating both the use of the geometric interpretation
and a good transformation property of the intrinsic normalization.
Proposition 9.7.6. Let m be a nonzero integer not divisible by p, and let
f m : F F 1/m be the substitution t  t m . Then, for any finite differential
module V over F , I R(V ) = I R( f m (V )).
Proof. Let m be a primitive mth root of unity in K alg . Then the claim follows
from the geometric interpretation plus the fact that
|t t mi | < c for some i {0, . . . , m 1} |t m tm | < c m

(c (0, 1)),
(9.7.6.1)

whose proof is left as an exercise.

Remark 9.7.7. Be aware that in Proposition 9.7.6 we are not quite performing
the standard base change. In explicit terms, if dm denotes the derivation with
respect to t on F 1/m and d1 denotes the derivation with respect to t on F
extended via f m then

162

Radius and generic radius of convergence

dm = mt m1 d1 .

We will encounter this issue again when we perform Frobenius pullback and
pushforward in Chapter 10. One may view it as an argument in favor of a
coordinate-free perspective (Proposition 6.3.1).
Remark 9.7.8. A similar construction can be made for E. Let L be the
completion of o K ((t1 )) oK K for the 1-Gauss norm. Then the substitution
t  t1 + (t t1 ) induces an isometry o K ((t)) o L t t1  for the 1-Gauss
norm that extends to an isometric embedding of E into the completion of
o L t t1  oL L for the 1-Gauss norm.

9.8 Subsidiary radii

It is sometimes important to consider not only the generic radius of convergence but also some secondary invariants.
Definition 9.8.1. Let V be a finite differential module over F . Let V1 , . . . , Vm
be the JordanHlder constituents of V . We define the multiset of subsidiary
generic radii of convergence, for short the subsidiary radii, to consist of
R(Vi ) with multiplicity dim Vi for i = 1, . . . , m. We will always list these
in increasing order, as s1 sn , so that s1 = R(V ). We also have
intrinsic subsidiary generic radii of convergence, obtained by multiplying the
subsidiary radii by 1 .
Remark 9.8.2. If we replace F by C((z)) in the definition of intrinsic subsidiary generic radii of convergence, then the negative logarithm of the product
of the radii equals the irregularity of V . Thus our analysis of the variation of
subsidiary radii, in the remainder of this part, will also imply results about the
variation of irregularity. See [144] for an application of this.
Remark 9.8.3. It is not immediate from the definition how to interpret the subsidiary radii. We will give an interpretation in a later chapter (Theorem 11.9.2).

9.9 Another example in rank 1

We introduce one more important example in rank 1, in the case p > 0.
Definition 9.9.1. For p > 0, the ArtinHasse exponential series is the formal
power series
 i
 tp
E p (t) = exp
.
pi
i=0

9.9 Another example in rank 1

163

Proposition 9.9.2. We have E p (t) (Z p Q)t.

Proof. This follows from the formal identity

E p (x) =
(1 x n )(n)/n ,

(9.9.2.1)

n1, pn

in which (n) is the Mbius function. That is, (n) equals (1)e if n is the
product of e 0 distinct primes, and it equals 0 otherwise.
Example 9.9.3 (Matsuda). Let h be a nonnegative integer, and suppose that K
contains a primitive p h+1 th root of unity . Let Mh be the differential module
of rank 1 on the whole t-line, the action of D on a generator v being given by
D(v) =

h !


" i
i
p 1 t p 1 v;

i=0

note that M1 is isomorphic to the module of Example 9.3.5 for = 1. On

the open disc |t| < 1, Mh admits the horizontal section
 h

 1 pi i
E p (t)
p
t
v = exp
v
(9.9.3.1)
E p ( t)
pi
i=0

by Proposition 9.9.2, so R(Mh ) 1. Since this horizontal section is bounded

it also gives a horizontal section in the open unit disc around a generic point of
norm 1, so R(Mh F1 ) = 1. (This can also be seen by arguing that Mh is trivial
on any disc of the form |t| < for < 1, so that R(Mh F ) = 1, and then
using the continuity of the generic radius of convergence as in Example 9.6.2.)
By Proposition 9.7.6 the same remains true if we pull back along the map
t  ct m , for any positive integer m not divisible by p and any c o
K . (This
can be generalized further; see the notes.)
Remark 9.9.4. Note that


E p (t p )
E p (t) p
= exp((1 ) pt)
.
E p ( t)
E p ( p t p )
p

Consequently, on some disc of radius greater than 1, Mh

pullback of Mh1 along t  t p .

is isomorphic to the

Remark 9.9.5. It is possible to prove directly that, for [1, +)

sufficiently close to 1,
I R(Mh F ) = p

(9.9.5.1)

(exercise). We can also do this using variational properties of the generic

radius; see the exercises for Chapter 11.

164

Radius and generic radius of convergence

9.10 Comparison with the coordinate-free definition

It is worth reconciling our definition of the generic radius of convergence with
the coordinate-free formula in Proposition 6.3.1. In the process, we can give a
more unified treatment of F and E.
Definition 9.10.1. Let F be a complete nonarchimedean field of characteristic
0 equipped with a bounded derivation d. For u F, we say that d is of rational
type for the parameter u if the following conditions hold.
(a) We have d(u) = 1 and |ud| F = 1.
(b) For each positive integer n, |d n /n!| F |d|nF .
For instance, this holds for F = F with d = d/dt and u = t, and similarly
for F = E if K is discretely valued. Note also that condition (b) is vacuously
true if F is of characteristic 0.
Proposition 9.10.2. Set notation as in Definition 9.10.1. Then, for any
nonnegative integer n and any c0 , . . . , cn F, we have

 n
 u i 


ci d i  = max{|ci |}.
(9.10.2.1)


i
i! 
i=0

Proof. The left-hand side of (9.10.2.1) is less than or equal to the right-hand
side because, by hypothesis, |u i d i /i!| F |u|i |d|iF 1 for all i. Conversely,
let j {0, . . . , n} be the minimal index for which |c j | = maxi {|ci |}; since

ui i j
j
ci u j
ci d (u ) =
i
i!
and F is of characteristic 0, we have


 i
= |c j u j |
 u i j 
ci d (u ) < |c j u j |

 i!

= 0

(i = j),
(i < j),
(i > j).

n
ci u i d i (u j )/i!| = |c j u j |, so the left-hand side of (9.10.2.1) is
Hence | i=0
greater than or equal to the right-hand side.
Corollary 9.10.3. Set notation as in Definition 9.10.1. Then, for any nonzero
finite differential module V over F, we have
|D|sp,V = |d|sp,F lim |Ds |1/s ,
s

where Ds is defined as in Proposition 6.3.1.

Notes

165

Proof. This follows from Proposition 6.3.1, since (6.3.1.2) holds by virtue of
Proposition 9.10.2.

Notes
As noted in [80, Appendix III], which can be consulted for more information, the p-adic Cauchy theorem (Proposition 9.3.3) was originally proved
by Lutz [163]. (Note that this publication long predates Dworks work.) See
Proposition 18.1.1 for a related result.
The idea of restricting a p-adic differential module to a generic disc originated in the work of Dwork [75], although in retrospect the base change
involved is quite natural in Berkovichs framework of nonarchimedean analytic geometry. For instance, the Berkovich space associated with a closed
unit disc contains generic points corresponding to each disc of each possible radius less than or equal to 1; the residue field of the generic point
of the disc |t| is precisely F . (This point of view was adopted
by Baldassarri and Di Vizio in [11] and by Baldassarri in [10].) Our
definition of the generic radius of convergence is taken from Christol and
Dwork [49].
The intrinsic generic radius of convergence (the original terminology) was
introduced in [140], where it is called the toric normalization in light of
Proposition 9.7.6.
Previously the subsidiary radii (the original terminology) have not been
studied much; the one reference we found is the work of Young [222]. We will
give Youngs interpretation of the subsidiary radii as the radii of convergence
of certain horizontal sections, in a refined form, as Theorem 11.9.2.
Our description of the ArtinHasse exponential follows Robert [191, 7.2].
Matsudas example, from [168], is an explicit instance of a general construction introduced by Robba [190]. In turn, Matsudas example can be greatly
generalized as shown by Pulita [185], building on work of Chinellato. One
obtains an analogous construction from any LubinTate group over Q p ; Matsudas example arises from the multiplicative group. (The introduction to [185]
provides a detailed historical discussion.)
A slightly different notion of the generic radius of convergence was introduced by Baldassarri and Di Vizio [11] and studied further by Baldassarri [10].
The difference between that definition and ours appears only in the case where
our definition gets truncated. For instance, for a differential module on a closed
disc of radius , our generic radius of convergence at radius (0, ) is only
allowed to take values up to . However, Baldassarri and Di Vizio allow values

166

Radius and generic radius of convergence

up to since the domain of definition of the original module includes a generic

disc of radius .
The approach of Baldassarri and Di Vizio gives a more refined invariant than
ours, since ours can be recovered by truncation. Thus, their approach is likely
to have certain advantages in applications. One serious disadvantage is that, at
the moment, there is no good theory of subsidiary generic radii of convergence
in the BaldassarriDi Vizio framework. The correct definition is presumably
that suggested by Theorem 11.9.2, but it is not clear how to prove any useful
properties for it.
The notion of a derivation of rational type was introduced in [145, 1.4]
as a way to isolate those features of the field F which are needed in the
spectral theory of differential modules. For instance, it is shown in [145]
that the rational type is preserved under unramified or tamely ramified field
extensions. The coordinate-free interpretation of the generic radius of convergence (Corollary 9.10.3) is useful in the study of the irregularity of
higher-dimensional flat meromorphic connections [144].

Exercises
(1) Prove Lemma 9.2.1. Then exhibit an example showing that the cokernel of d/dt on K /t, t/ is not spanned over K by t 1 . That is,
antidifferentiation with respect to t is not well defined on K /t, t/.
(2) Prove Example 9.3.5. (Hint: use the exponential series to construct a
horizontal section.)
(3) Prove Example 9.5.2. (Hint: consider the cases Z p , o K Z p ,
and
/ o K separately.)
(4) In Example 9.5.2 give an explicit formula for I R(V ) in terms of and
the minimum distance from to an integer.
(5) Prove (9.7.6.1). (Hint: you may find it easier to start with the cases where
m > 0 and where m = 1 and then deduce the general result from them.)
(6) The following considerations illustrate the pitfalls of using td/dt instead
of d/dt in the p-adic setting.
(a) Verify that |td/dt|sp,F  = |t| |d/dt|sp,F .
(b) Show that the inequality (6.3.1.1) of Proposition 6.3.1 can be strict
for F = F and d = td/dt. (Hint: use Example 9.5.2.)
(7) With notation as in Proposition 9.7.6, show that all the intrinsic subsidiary
radii of V match those of f m (V ), not just the generic radii.
(8) Here is an off-centered analogue of Proposition 9.7.6 suggested by
Liang Xiao (compare with Theorem 10.8.2 below). Let m be a nonzero
integer not divisible by p. Given (0, 1], let f m : F F be the

Exercises

167

map t  (t + 1)m 1. Then, for any finite differential module V over

F , R(V ) = R( f m (V )). (As in the previous exercise, one also obtains
equality for the other subsidiary radii.)
(9) Prove (9.9.2.1). (Hint: take logarithms.)
(10) Prove (9.9.5.1) by analyzing an explicit horizontal section around
a generic point. A similar argument is given in the proof of [39,
Proposition 1.5.1]. (Hint: use the equality |1 | = p p
Example 2.1.6.)

h+1 /( p1)

from

10
Frobenius pullback and pushforward

In this chapter, we introduce Dworks technique of Frobenius descent to analyze the generic radius of convergence and subsidiary radii of a differential
module, in the range where Newton polygons do not apply. In one direction
we introduce a somewhat refined form of the Frobenius antecedents introduced
by Christol and Dwork. These fail to apply in an important boundary case; we
remedy this by introducing the new notion of Frobenius descendants, which
covers the boundary case.
Using these results, we are able to improve a number of results from
Chapter 6 in the special case of differential modules over F . For instance
we get a full decomposition by spectral radius, extending the visible decomposition theorem (Theorem 6.6.1) and the refined visible decomposition theorem
(Theorem 6.8.2). We will use these results again to study the variation of subsidiary radii, and decomposition by subsidiary radii, in the remainder of this
part.
Notation 10.0.1. Throughout this chapter we retain Hypothesis 9.0.1. We also
continue to use F to denote the completion of K (t) for the -Gauss norm
viewed as a differential field with respect to d = d/dt, unless otherwise
specified.
Notation 10.0.2. Throughout this chapter we also assume p > 0 unless
otherwise specified.

10.1 Why Frobenius descent?

Remark 10.1.1. It may be helpful to review the current state of affairs, in order
to clarify why we need to descend along a Frobenius morphism.
168

10.2 pth powers and roots

169

Let V be a finite differential module over F . Then the possible values

of the spectral radius |D|sp,V are the real numbers greater than or equal to
|d|sp,F = p 1/( p1) 1 , corresponding to generic radii of convergence less
than or equal to . However, in calculating the spectral radius using the Newton
polygon of a twisted polynomial, we cannot distinguish values less than or
equal to the operator norm |d| F = 1 . In particular, we cannot use this technique to prove a decomposition theorem for differential modules that separates
components of spectral radius between p 1/( p1) 1 and 1 .
One way in which one might want to get around this is to consider not d
but a high power of d, in particularly the p n th power. The trouble with this is
that iterating a derivation does not give another derivation but something much
more complicated. Thus, instead we will try to differentiate with respect to
n
t p , rather than t, in order to increase the spectral radius into the range where
Newton polygons become useful.

10.2 pth powers and roots

We first make some calculations in answer to the following question: if two
p-adic numbers are close together, how close are their pth powers or their pth
roots? (See also [80, V.6] and [42, Proposition 4.6.4].)
Remark 10.2.1. We observed previously, in (9.7.6.1), using slightly different
notation, that when m is a positive integer coprime to p, for m a primitive mth
root of unity we have
|t mi | < || for some i {0, . . . , m 1} |t m m | < ||m

( (0, 1)).

This breaks down for m = p, because |1 p | = p 1/( p1) < 1 by

Example 2.1.6. Instead, we have the following bounds.
Lemma 10.2.2. Pick t, K .
(a) For (0, 1), if |t | || then
|t | max{ , p
p

}| | =
p

p | p |

p 1/( p1) ,

p 1 | p |

p 1/( p1) .

(b) Suppose that p K . If |t p p | | p | then there exists m

{0, . . . , p 1} such that

1/ p || p p/( p1) ,
|t pm | min{1/ p , p}|| =
p||
p p/( p1) .
Moreover, if p p/( p1) , we may always take m = 0.

170

Frobenius pullback and pushforward

We will use repeatedly, and without comment, the fact that

 max{ p , p 1 },

 min{1/ p , p}

are strictly increasing functions from [0, 1] to [0, 1] that are inverse to each
other.
Proof. There is no harm in assuming that p K for both parts. For (a), factor
 p1
t p p as (t ) m=1 (t pm ), and write
t pm = (t ) + (1 pm ).
If |t | p 1/( p1) || then t is the dominant term; otherwise (1 pm )
dominates. This gives the claimed bounds.
For (b), consider the Newton polygon of
p1 

p i
p
p
t c =
(t ) pi c,
i
i=0

viewed as a polynomial in t . Suppose that |c| = | p |. If p p/( p1)

then the terms (t ) p and c dominate, and all roots have norm 1/ p ||. Otherwise the terms (t ) p , p(t ) p1 , and c dominate, so one root has
norm p|| and the others are larger. Repeating with replaced by pm for
m = 0, . . . , p 1 gives p distinct roots. This accounts for all the roots.
Corollary 10.2.3. Let T : K t p p  K t  be the substitution
p1 

p i
(t ) pi .
t p p 
i
i=0
p
p
)/(| |) for

some (0, 1) then T ( f ) K (t

(a) If f
)/( ||) for  = min{1/ p , p}.
(b) If T ( f ) K (t )/(||) for some ( p1/( p1) , 1) then f
K (t p p )/( | p |) for  = p .
(c) Suppose that K contains a primitive pth root of unity p . For m =
0, . . . , p 1, let Tm : K t p p  K t pm  be the substi
p1  p im i
m pi . If for some
tution t p p 
i=0 i p (t p )
(0, p 1/( p1) ] one has Tm ( f ) K (t pm )/(||), for m =
0, . . . , p 1, then f K (t p p )/( | p |) for  = p 1 .
K (t p

10.3 Frobenius pullback and pushforward operations

We now define Frobenius pullback and pushforward operations, and show how
they affect the generic radius of convergence.

10.3 Frobenius pullback and pushforward operations

171

Definition 10.3.1. Let F be the completion of K (t p ) for the p -Gauss norm,

viewed as a subfield of F and equipped with the derivation d  = d/d(t p ). We
then have
d=

d(t p ) 
d = pt p1 d  .
dt

Given a finite differential module (V  , D  ) over F , we may view V  =

V  F F as a differential module over F with
D(v f ) = pt p1 D  (v) f + v d( f ).
Note that this is not the usual base-change operation, because the restriction of
d to F is not d  ; this is the same situation as that encountered in Remark 9.7.7.
Lemma 10.3.2. Let (V  , D  ) be a finite differential module over F . Then
I R( V  ) min{I R(V  )1/ p , p I R(V  )}.
Proof. For any < I R(V  ), any complete extension L of K , and any
p
generic point t L relative to K of norm , V  L(t p t )/( p )
admits a basis of horizontal sections. By Corollary 10.2.3(a), ( V  ) L(t
t )/(min{1/ p , p}) does likewise.
Here is a important example, for which Lemma 10.3.2 gives a strict
inequality.
Definition 10.3.3. For m = 0, . . . , p 1, let Wm be the differential module
over F with one generator v such that
D(v) =

m p
t v.
p

(The generator v is taken to behave as t m = (t p )m/ p .) From the Newton polygon associated to v, we may read off I R(Wm ) = p p/( p1) for m  = 0 since
this is within the range of applicability of Theorem 6.5.3. Note that the inequality of Lemma 10.3.2 is strict in this case, since Wm is trivial and so has
intrinsic radius 1.
Definition 10.3.4. For V a differential module over F , define the Frobenius
descendant of V as the module V obtained from V by restriction along
F F . The module V is viewed as a differential module over F with
differential D  = p 1 t 1 p D. Note that this operation commutes with duals,
but not with tensor products because of a rank mismatch: the rank of V is p
times that of V . See, however, Lemma 10.3.6(f) below.

172

Frobenius pullback and pushforward

Remark 10.3.5. The definition of the Frobenius descendant extends to differential modules over K /t, t/ for > 0 but not for = 0, since we must
divide by a power of t to express D  in terms of D. The underlying problem
is that (in geometric terms) the map ramifies at the point t = 0. We will see
one way to deal with this problem in the discussion of off-centered Frobenius
descendants (Section 10.8).
The operation enjoys a number of properties which are useful and
reasonably easy to verify.
Lemma 10.3.6
(a) For V a differential module over F , there are canonical isomorphisms
m : ( V ) Wm
= V

(m = 0, . . . , p 1).

(b) For V a differential module over F , a submodule U of V is itself

the Frobenius descendant of a submodule of V if and only if m (U
Wm ) = U for m = 0, . . . , p 1.
(c) For V  a differential module over F , there is a canonical isomorphism
V
=

p1
%

(V  Wm ).

m=0

(d) For V a differential module over F , there is a canonical isomorphism

V
= V p.
(e) For V a differential module over F , there are canonical bijections
H i (V )
= H i ( V )

(i = 0, 1).

(f) For differential modules V1 , V2 over F , there is a canonical isomorphism

V1 V2
= ( (V1 V2 )) p .
Proof. Exercise.

10.4 Frobenius antecedents

An important counterpart to the construction of Frobenius descendants is the
construction of Frobenius antecedents; this inverts the pullback operation
when the intrinsic radius is sufficiently large.

10.4 Frobenius antecedents

173

Definition 10.4.1. Let (V, D) be a finite differential module over F .

A Frobenius antecedent of V is a differential module (V  , D  ) over F
such that I R(V  ) > p p/( p1) , together with an isomorphism V
= V .
By Lemma 10.3.2 a necessary condition for the existence of a Frobenius
antecedent is that I R(V ) > p 1/( p1) ; Theorem 10.4.2 below implies that
this condition is also sufficient.
Theorem 10.4.2 (after Christol and Dwork). Let (V, D) be a finite differential module over F such that I R(V ) > p 1/( p1) . Then there exists a unique
Frobenius antecedent V  of V . Moreover, I R(V  ) = I R(V ) p .
Proof. We will assume that p K , as otherwise we could check everything
by adjoining p and then performing a Galois descent at the end.
We first check existence. Since |D|sp,V < 1 , we may define an action of
Z/ pZ on V using a Taylor series:
pm (x) =


( pm t t)i
i=0

i!

D i (x)

(x V, m Z/ pZ).

Note that the maps P j : V V defined by

P j (v) =

p1
1  i j i
p p (v)
p

( j = 0, . . . , p 1)

i=0

are F -linear projectors onto the generalized eigenspaces for the characters of
Z/ pZ. Note also that these eigenspaces are permuted by multiplication by t, so
they must all have the same dimension. We may conclude that the fixed space
V  is an F -subspace of V and the natural map V  V is an isomorphism.
(This calculation amounts to a simple instance of the HilbertNoether theorem,
i.e., of Galois descent.)
By applying the Z/ pZ-action to a basis of horizontal sections of V in the
generic disc |t t | for some ( p 1/( p1) , I R(V )) and then invoking
Corollary 10.2.3(b), we may construct horizontal sections of V  in the generic
p
disc |t p t | p p . Hence I R(V  ) I R(V ) p > p p/( p1) .
To check uniqueness, suppose that V
= V  with I R(V  ),
= V 


p/(
p1)
. By Lemma 10.3.6(c) we have
I R(V ) > p
p1
p1
V
= m=0 (V  Wm )
= m=0 (V  Wm ).

For m = 1, . . . , p 1, we have I R(Wm ) = p p/( p1) ; since I R(V  ) >

I R(Wm ), we have I R(V  Wm ) = p p/( p1) by Lemma 9.4.6(c). Since
I R(V  ) > p p/( p1) , the factor V  W0 must be contained in V  W0 and
vice versa. Hence V 
= V  .

174

Frobenius pullback and pushforward

For the last assertion, note that the proof of existence gives I R(V  )
I R(V ) p , whereas Lemma 10.3.2 gives the reverse inequality.
Corollary 10.4.3. Let V  be a differential module over F such that I R(V  ) >
p p/( p1) . Then V  is the Frobenius antecedent of V  , so I R(V  ) =
I R( V  ) p .
Proof. By Lemma 10.3.2, I R( V  ) I R(V  )1/ p > p 1/( p1) , so V 
has a unique Frobenius antecedent by Theorem 10.4.2. Since I R(V  ) >
p p/( p1) , V  is that antecedent.
The construction of Frobenius antecedents carries over to discs and annuli
as follows.
Theorem 10.4.4. Let M be a finite differential module over K /t, t/
(we may allow = 0) such that I R(M F ) > p 1/( p1) for
[, ] (or, equivalently, for = and = ). Then there exists a
unique differential module M  over K  p /t p , t p / p  such that M = M 
K /t, t/ and I R(M  F ) > p p/( p1) for [, ]; this M  also
satisfies I R(M  F ) = I R(M F ) p for [, ]. (This theorem
also holds with K /t, t/, K  p /t p , t p / p  replaced with K /t, t/an ,
K  p /t p , t p / p an , respectively, by a similar proof.)
Proof. Define the projectors P j as in the proof of Theorem 10.4.2, and let
M  be the image of P0 . By arguing as in Theorem 10.4.2 we may show that
M  K /t, t/[t 1 ]
= M K /t, t/[t 1 ]. However, the quotient of

M by the image of M K /t, t/ is a differential module and so cannot
have any t-torsion, by Proposition 9.1.2. Hence M  K /t, t/
= M, and
we may continue as in Theorem 10.4.2.

10.5 Frobenius descendants and subsidiary radii

We saw in Lemma 10.3.2 that we can only weakly control the behavior of
the generic radius of convergence under Frobenius pullback. Under Frobenius
pushforward, we can do much better; we can control not only the generic radius
of convergence but also the subsidiary radii. This will lead to a refinement of
Lemma 10.3.2; see Corollary 10.5.4. (Example 11.7.2 gives an explicit case.)
Theorem 10.5.1. Let V be a finite differential module over F with intrinsic
subsidiary radii s1 sn . Then the intrinsic subsidiary radii of V
comprise the multiset
 p
n
$
{si , p p/( p1) ( p 1 times)} si > p 1/( p1) ,
i=1

{ p 1 si ( p times)}

si p 1/( p1)

10.5 Frobenius descendants and subsidiary radii

175

(note that in the upper expression the second element occurs p 1 times altogether and in the lower expression the first element occurs p times altogether).
In particular,
I R( V ) = min{ p 1 I R(V ), p p/( p1) }.
Proof. It suffices to consider V as irreducible. First suppose that I R(V ) >
p 1/( p1) . Let V  be the Frobenius antecedent of V (as per Theorem 10.4.2);
p1
note that V  is also irreducible. By Lemma 10.3.6(c), V
= m=0 (V  Wm ).

Since each Wm has rank 1, V Wm is also irreducible. Since I R(V  ) =
I R(V ) p by Theorem 10.4.2 and I R(V  Wm ) = p p/( p1) for m  = 0 by
Lemma 9.4.6(c), we have the claim.
Next suppose that I R(V ) p 1/( p1) . We show first that
I R( V ) p 1 I R(V ) = max{I R(V ) p , p 1 I R(V )}.
For t a generic point of radius and (0, p 1/( p1) ), the module V
p
L(t p t )/( p 1 p ) splits as the direct sum of V L(t pm t )/()
over m = 0, . . . , p 1. If < I R(V ), by applying Corollary 10.2.3(c) we
obtain I R( V ) p 1 .
Next, let W  be any irreducible subquotient of V ; then I R(W  )
I R( V ), so Lemma 10.3.2 gives on the one hand
I R( W  ) min{I R(W  )1/ p , p I R(W  )} min{I R( V )1/ p , p I R( V )} I R(V ).
(10.5.1.1)

On the other hand W  is a subquotient of V , which by

Lemma 10.3.6(d) is isomorphic to V p . Since V is irreducible, each Jordan
Hlder constituent of W  must be isomorphic to V , yielding I R( W  ) =
I R(V ). That forces each inequality in (10.5.1.1) to be an equality; thus, in particular, I R(W  ) and I R( V ) have the same image under the injective map
s  min{s 1/ p , ps}. We conclude that I R(W  ) = I R( V ) = p 1 I R(V ),
proving the claim.
Remark 10.5.2. One might be tempted to think that the verification that
I R( V ) p 1 I R(V ) within the proof of Theorem 10.5.1 should carry over
to the case I R(V ) > p 1/( p1) , in which case it would lead to the false conclusion I R( V ) I R(V ) p . What breaks down in this case is that pushing
forward a basis of local horizontal sections of V gives only dim V local horizontal sections of V ; what these span is precisely the Frobenius antecedent
of V .
Corollary 10.5.3. Let s1 sn be the intrinsic subsidiary radii of V .
(a) For i such that si p 1/( p1) , the product of the i p smallest intrinsic
p
p
subsidiary radii of V is equal to p i p s1 si .

176

Frobenius pullback and pushforward

(b) For i such that either i = n or si+1 p 1/( p1) , the product of the
i p + ( p 1)(n i) smallest intrinsic subsidiary radii of V is equal
p
p
to p in s1 si .
In particular, the product of the intrinsic subsidiary radii of V is
p
p
p np s1 sn .
Note that both conditions apply when si = p 1/( p1) ; this will be important
later (see Remark 11.6.2).
As promised, we will now obtain a refined version of Lemma 10.3.2.
Corollary 10.5.4. Let V  be a differential module over F such that I R(V  )  =
p p/( p1) . Then I R( V  ) = min{I R(V  )1/ p , p I R(V  )}. (This fails when
I R(V  ) = p p/( p1) , e.g., for V  = Wm .)
Proof. In the case I R(V  ) > p p/( p1) , this holds by Corollary 10.4.3. Othp1
erwise, by Lemma 10.3.6(c), V 
= m=0 (V  Wm ) and I R(V  Wm ) =


I R(V ) since I R(V ) < I R(Wm ). Hence, by Theorem 10.5.1,
I R(V  ) = I R( V  ) = min{ p 1 I R( V  ), p p/( p1) }.
We get a contradiction if the right-hand side equals p p/( p1) , so it must be
the case that I R(V  ) = p 1 I R( V  ) p p/( p1) , proving the claim.

10.6 Decomposition by spectral radius

As our first application of Frobenius descent, we will extend the visible decomposition theorem, and its refined form, in the special case of a differential
module over F . We then deduce some consequences analogous to the consequences of the visible decomposition theorem. To do this, we must use
Frobenius descendants to cross the bound on the spectral radius. This cannot
be done using Frobenius antecedents alone, as they give no information in the
boundary case I R(V ) = p 1/( p1) .
In this section, we may suppress the hypothesis p > 0, since the case p = 0
is already covered by the original decomposition theorems.
Proposition 10.6.1. Let V1 , V2 be irreducible finite differential modules over
F , with I R(V1 )  = I R(V2 ). Then H 1 (V1 V2 ) = 0.
Proof. We may assume that I R(V2 ) > I R(V1 ); note that I R(V1 ) = I R(V1 )
by Lemma 9.4.6(b). If p = 0, or if p > 0 and I R(V1 ) < p 1/( p1) , then
any short exact sequence 0 V2 V V1 0 splits by the visible decomposition theorem (Theorem 6.6.1), yielding the desired vanishing by
Lemma 5.3.3. Thus we may assume hereafter that p > 0.

10.6 Decomposition by spectral radius

177

Suppose that p > 0 and I R(V1 ) = p 1/( p1) . Let V2 be the Frobenius
antecedent of V2 ; it also is irreducible, and I R(V2 ) = I R(V2 ) p > p p/( p1)
by Theorem 10.4.2. By Theorem 10.5.1, each irreducible subquotient W of
V1 satisfies I R(W ) = p p/( p1) ; hence H 1 (W V2 ) = 0 by the previous
case, so H 1 ( V1 V2 ) = 0 by the snake lemma.
By Lemma 10.3.6(a), (c),
p1
V1 V2
= m=0 ( V1 Wm V2 )

= ( V1 V  ) p .
2

This yields H 1 ( V1 V2 ) = 0. By Lemma 10.3.6(f), (V1 V2 ) is a direct

summand of V1 V2 , so H 1 ( (V1 V2 )) = 0. By Lemma 10.3.6(e),
H 1 (V1 V2 ) = H 1 ( (V1 V2 )) = 0.
In the general case, 1 I R(V2 ) > I R(V1 ). If I R(V1 ) > p 1/( p1)
then Theorem 10.4.2 implies that V1 , V2 have Frobenius antecedents V1 , V2 .
In addition, for any extension 0 V1 V V2 0, the module V
satisfies I R(V ) > p 1/( p1) by Lemma 9.4.6, so Theorem 10.4.2 implies that
the whole sequence is itself the pullback of an extension 0 V1 V 
(V2 ) 0. To show that V always splits, it suffices to do so for V  ; that is, we
may reduce from V1 , V2 to V1 , V2 . By repeating this enough times, we reach
the situation where I R(V1 ) p 1/( p1) . We may then apply the previous
cases.
From here, the proof of the following theorem is purely formal.
Theorem 10.6.2 (Strong decomposition theorem). Let V be a finite differential module over F . Then there exists a decomposition
%
Vs ,
V =
s(0,1]

where every subquotient Ws of Vs satisfies I R(Ws ) = s.

Proof. We induct on dim V ; we need consider only reducible V . Choose a
short exact sequence 0 U1 V U2 0 with U2 irreducible.
Decompose U1 as s(0,1] U1,s , where every subquotient Ws of U1,s satisfies
I R(Ws ) = s. For each s  = I R(U2 ), we have H 1 (U2 U1,s ) = 0 by repeated
application of Proposition 10.6.1 plus the snake lemma. By Lemma 5.3.3 we
have
%
V = V
U1,s ,
s= I R(U2 )

where 0 U1,I R(U2 ) V  U2 0 is exact.

178

Frobenius pullback and pushforward

As with the visible decomposition theorem, we obtain the following corollaries.

Corollary 10.6.3. Let V be a finite differential module over F whose intrinsic
subsidiary radii are all less than 1. Then H 0 (V ) = H 1 (V ) = 0.
Proof. It is clear that H 0 (V ) vanishes, as otherwise V would have a submodule with intrinsic generic radius of convergence equal to 1. To see that
H 1 (V ) = 0, it suffices by Lemma 5.3.3 to show that any short exact sequence
0 V W X 0 with X trivial is split. This follows by applying
Theorem 10.6.2 to W : in the resulting decomposition, V and X must project
into distinct summands, so W
= V X.
Corollary 10.6.4. With V = s(0,1] Vs as in Theorem 10.6.2, we have
H i (V ) = H i (V1 ) for i = 0, 1.
This suggests that difficulties in computing H 0 and H 1 arise in the case
of intrinsic generic radius 1. We will pursue a closer study of this case in
Chapter 13.
Using the strong decomposition theorem, we obtain a refined version of
Corollary 6.2.9 that extends Corollary 6.6.3.
Corollary 10.6.5. If V1 and V2 are irreducible finite differential modules over
F and I R(V1 ) < I R(V2 ) then every irreducible subquotient W of V1 V2
satisfies I R(W ) = I R(V1 ).
Proof. Decompose V1 V2 = s(0,1] Vs according to Theorem 10.6.2; by
Lemma 9.4.6(c), Vs = 0 whenever s < I R(V1 ). If some Vs with s > I R(V1 )
were nonzero then V1 V2 would have an irreducible submodule of intrinsic
radius greater than I R(V1 ), in violation of Corollary 6.2.9.
We should also mention the following related result extending Corollary 6.2.7
and Corollary 6.5.5.
Proposition 10.6.6. Let F E be an isometric embedding of complete nonarchimedean differential fields such that |d| F = |d| E and |d|sp,F = |d|sp,E .
Then, for any finite differential module V over F , the full spectra of V and
V F E are equal.
Proof. We may assume that V is irreducible and has spectral radius s. We first
check the case where E is a finite extension of F ; we may assume that E is
Galois over F . Let W be any submodule of V F E; then V F E is a
quotient of the direct sum of the Galois conjugates of W , all of which have the
same spectral radius. Hence W and V F E have the same spectral radius,

10.6 Decomposition by spectral radius

179

which by Corollary 6.2.7 is s. Moreover, every constituent of V F E is a

conjugate of W and so has spectral radius s. This proves the claim.
In the general case, we may use the previous paragraph to add a primitive pth root of unity p to F in the case p > 0. If p = 0 or s >
then Theorem 6.5.3 implies the desired result. Otherwise we may apply
Theorem 10.5.1, replacing F by F and E by the fixed field under the action
of Z/ pZ (given by a Taylor series). Repeating this finitely many times, we
arrive back at the case where Theorem 6.5.3 becomes applicable.
We end this discussion by extending the refined visible decomposition
theorem (Theorem 6.8.2) to the full spectrum.
Theorem 10.6.7 (Refined strong decomposition theorem). Let V be a finite
differential module over F such that no subquotient of V has intrinsic radius
equal to 1. Then, for some finite tamely ramified extension E of F , V F E
admits a (unique) direct sum decomposition
%
V F E =
Vi ,
i

with each Vi refined, such that for i  = j we have Vi  V j in the sense of

Definition 6.2.12.
Proof. By Theorem 10.6.2 we may reduce immediately to the case where V
is pure and I R(V ) < 1. If p = 0, or if p > 0 and I R(V ) < p 1/( p1) , then
Theorem 6.8.2 gives the claim.
Next we consider the case I R(V ) = p 1/( p1) ; by Theorem 10.5.1, V is
again pure with I R( V ) = p p/( p1) . By the previous paragraph, for some
finite tamely extension E  of F we obtain a decomposition of ( V ) F E 
into nonequivalent refined submodules. For V1 , V2 appearing in this decomposition, we declare V1 and V2 to be weakly equivalent if V1 is equivalent
to V2 E  (Wm F E  ) for some m. This is again an equivalence relation;
by Lemma 10.3.6(b), if we group the summands of ( V ) F E  into weak
equivalence classes then the resulting decomposition descends to V F E for
E = F F E  .
Let X be a summand of V F E in this decomposition; it remains to check
that X is refined. By Lemma 10.3.6(f) we have
( X ) ( X )
= ( (X X )) p .
Hence, by the previous construction, (X X ) decomposes as a direct sum
in which each summand can be twisted by a suitable Wm F E  to raise its
intrinsic radius above p p/( p1) . However, by Lemma 10.3.6(a), (X X )

180

Frobenius pullback and pushforward

is isomorphic to its own twist by any Wm F E  . Hence on the one hand its
intrinsic subsidiary radii must comprise a multiset of p rank(X )2 elements in
which exactly rank(X )2 of the elements are greater than p p/( p1) . On the
other hand, since I R(X X ) I R(X ) = p 1/( p1) by Lemma 9.4.6(c),
Theorem 10.5.1 implies that the intrinsic subsidiary radii of (X X )
include at most rank(X )2 elements greater than p p/( p1) . In fact, equality
occurs if and only if all the intrinsic subsidiary radii of X X are greater
than p 1/( p1) . Hence X is refined.
Finally, we handle the case p 1/( p1) < I R(V ) < 1 by induction on the
h
smallest integer h such that I R(V ) p p /( p1) . The case h = 0 is handled by the arguments above. If h > 0 then, by Theorem 10.4.2, V has a
h+1 /( p1)
.
Frobenius antecedent W for which I R(W ) = I R(V )1/ p p p
By the induction hypothesis, W admits a decomposition into nonequivalent
refined submodules; pulling back by a Frobenius morphism then gives the
desired decomposition of V , by Corollary 10.5.4.

10.7 Integrality of the generic radius

The relationship between the generic radius of convergence and Newton polygons in the visible range (given in Theorem 6.5.3) suggests that the generic
radius of convergence should satisfy some sort of integrality property. On the
one hand we can infer such a property using Frobenius antecedents; on the
other hand, a certain price must be paid.
Theorem 10.7.1. Let V be a finite differential module over F with intrinsic
subsidiary radii s1 sn . Let m be the largest integer such that sm =
I R(V ). Then, for any nonnegative integer h,
s1 < p p

h /( p1)

s1m |K | p Z .

Proof. For h = 0 we can read this off from the case of a Newton polygon
(i.e., we can invoke Theorem 6.5.3). To reduce from h to h 1, if I R(V ) >
p 1/( p1) then replace V by its Frobenius antecedent (Theorem 10.4.2); if
I R(V ) = p 1/( p1) , apply and invoke Corollary 10.5.3.
Here is an example to show that the exponent p h in the conclusion s1m
h
|K | p Z of Theorem 10.7.1 is not spurious.
Example 10.7.2. Suppose that K satisfies | | = p 1/( p1) . Pick K
and 0 < such that, for [, ],
p 1/( p1) < || p < p p/( p1) .

10.8 Off-center Frobenius antecedents and descendants

181

Let M be the differential module over K /t, t/ that is generated by v satisfying D(v) = p t p1 v. Then M
= M  , where M  is the differential
module over K  p /t p , t p / p  with generator w and D  (w) = (t p )2 w.
We deduce that
|D  | M  F = p 1/( p1) || 2 p > p .
Hence we have
I R(M  F ) = ||1 p ,
I R(M F ) = ||1/ p ,
where the first equality follows by Theorem 6.5.3 and the second follows from
the first by Corollary 10.4.3.
Remark 10.7.3. Another way to understand Example 10.7.2 is by means of
the Dwork exponential series; see Definition 17.1.3.
Question 10.7.4. What is the correct extension of Theorem 10.7.1 for the
remaining subsidiary radii? (This should not be difficult.)

10.8 Off-center Frobenius antecedents and descendants

Since pushing forward along a Frobenius morphism does not work well
on a disc (Remark 10.3.5), we must also consider off-center Frobenius
antecedents and descendants. Although this can be done rather more generally,
we will stick to one case that is sufficient for our purposes.
Definition 10.8.1. For ( p 1/( p1) , 1], let F be the completion of K ((t
1) p 1) under the p -Gauss norm. Note that this coincides with the restriction
of the -Gauss norm on K (t), because |((t 1) p 1) t p | < |t p | . (One
could allow K ((t ) p p ) for any K of norm 1, but there is no
significant loss of generality in rescaling t to reduce to the case = 1.) For
brevity, write u = (t 1) p 1. Equip F with the derivation
d  =

d
1
=
d.
du
du/dt

Given a differential module V  over F , we may view V  = V  F

as a differential module over F . Given a differential module V over F , we
may view the restriction V of V along F F as a differential module
over F .

182

Frobenius pullback and pushforward

The main point here is that du/dt K t . Consequently, we can extend
both and not just to annuli but also discs. This is needed to establish the
monotonicity property for subsidiary radii (Theorem 11.3.2(d)).
We may apply Lemma 10.2.2 with replaced by + 1, keeping in mind
that | + 1| = 1 for || < 1. This has the net effect that everything that holds
for also holds for , except that the intrinsic generic radius of convergence
must be replaced by the extrinsic one. Rather than rederive everything, we
simply state the analogues of Theorems 10.4.2 and 10.5.1 and leave the proofs
as exercises.
Theorem 10.8.2. Let (V, D) be a finite differential module over F such that
R(V ) > p 1/( p1) . Then there exists a unique differential module (V  , D  )
over F equipped with an isomorphism V
= V  such that R(V  ) >

p/(
p1)


. For this V , one has in fact R(V ) = R(V ) p .
p
Theorem 10.8.3. Let V be a finite differential module over F with extrinsic subsidiary radii s1 , . . . , sn . Then the extrinsic subsidiary radii of V
comprise the multiset
 p
n
$
{si , p p/( p1) ( p 1 times)} si > p 1/( p1) ,
i=1

{ p 1 si ( p times)}

si p 1/( p1) .

Remark 10.8.4. Note that one cannot expect Theorem 10.8.3 to hold for <
p 1/( p1) , as in that case p p/( p1) is too large to appear as a subsidiary radius
of V .

Notes
Lemma 10.2.2 is taken from [128, 5.3] with some typos corrected.
The Frobenius antecedent theorem of Christol and Dwork [49, Thorme 5.4]
is slightly weaker than the one given here: it only applies for I R(V ) > p 1/ p .
The discrepancy is due to the introduction of cyclic vectors, which create some
regular singularities which can only eliminated under the stronger hypothesis. Much closer to the statement of Theorem 10.4.2 is [128, Theorem 6.13];
however, uniqueness is only asserted there when I R(V  ) I R(V ) p .
The concept of the Frobenius descendant, and the results deduced using
it, are original. This includes Theorem 10.5.1 and its off-center analogue
(Theorem 10.8.3), the strong decomposition theorem (Theorem 10.6.2), and
the refined strong decomposition theorem (Theorem 10.6.7). The latter was
suggested by Liang Xiao; see the notes for Chapter 6 for the motivation
for this.

Exercises

183

The notions of Frobenius antecedents and descendants extend to derivations

of rational type, with one caveat: the assertion of Theorem 10.7.1 does not
carry over. See [145, Theorem 1.4.21] for the correct statement.

Exercises
(1) Prove Lemma 10.3.6.
(2) Prove that, for any finite differential module V  over F with I R(V  ) >
p p/( p1) , H 0 (V  ) = H 0 ( V  ). (The example V  = Wm shows that the
bound on I R(V  ) cannot be relaxed.)
(3) Derive Theorem 10.8.2 by imitating the proof of Theorem 10.4.2.
(4) Derive Theorem 10.8.3 by imitating the proof of Theorem 10.5.1.

11
Variation of generic and subsidiary radii

In this chapter, we apply the tools developed in the preceding chapters to study
the variation of the generic radius of convergence, and of the subsidiary radii,
associated with a differential module on a disc or annulus. We have already
seen some instances where this study is needed to deduce consequences about
the convergence of solutions of p-adic differential equations (Examples 9.6.2
and 9.9.3).
The statements we will formulate are modeled on statements governing the
variation of the Newton polygon of a polynomial over a ring of power series as
we vary the choice of Gauss norm on the power series ring. The guiding principle is that, in the visible spectrum, one should be able to relate the variation of
subsidiary radii to the variation of Newton polygons via matrices of action of
the derivation on suitable bases. This includes the relationship between subsidiary radii and Newton polygons for cyclic vectors (Theorem 6.5.3), but
trying to use that approach directly creates no end of difficulties because cyclic
vectors only exist in general for differential modules over fields. We will implement the guiding principle in a somewhat more robust manner than before,
using the discussion of matrix inequalities in Chapter 6.
To this principle we must add the techniques of descent along a Frobenius
morphism introduced in Chapter 10, including the off-centered variant. This
allows us to overcome the limitation to the visible spectrum.
As corollaries of this analysis, we deduce some facts about the true radius of
convergence of a differential module on a disc. We also establish a geometric
interpretation of subsidiary radii in terms of the convergence of local horizontal
sections around a generic point, extending a result of Young. We will build
further on this work when we discuss decomposition theorems in Chapter 12.
Throughout this chapter we retain Notation 10.0.1 but we do not assume that
p > 0. We will continue to use the convention
184

11.1 Harmonicity of the valuation function


=

p = 0,

p 1/( p1)

p > 0.

185

11.1 Harmonicity of the valuation function

For f K /t, t/ and r [ log , log ], the function r  vr ( f )
is continuous, piecewise affine, and (by Proposition 8.2.3(b)) concave in r .
However, one can make an even more precise statement; for simplicity, we
will write this out explicitly only for r = 0.
Definition 11.1.1. For in some extension of k, let be a lift of in some
complete extension L of K . For 1 , define the substitution
T : K /t, t/ Ltan ,

t  t + .

(This map extends to K /t, t/0 if < 1 < .) The function r 

vr (T ( f )) on [0, +) is continuous and piecewise affine; moreover, its righthand slope at r = 0 does not depend on the choice of the field L or the lift
of . We call this slope s ( f ). For 1 < (resp. < 1), define s ( f ) (resp.
s0 ( f )) to be the left-hand (resp. right-hand) slope of the function r  vr ( f )
at r = 0.
Then we have the following harmonicity property.
Proposition 11.1.2. For 0 < 1 < and f K /t, t/ nonzero,
we have

s ( f ).
s ( f ) =
alg

Proof. Without loss of generality we may assume that | f |1 = 1. The quotient

of o F1 K /t, t/ by the ideal generated by m K is isomorphic to K [t, t 1 ];
let f be the image of f in this quotient. Then s is the order of vanishing of f
at , whereas s is the pole order of f at . The desired equality then follows
from the fact that a rational function has as many zeroes as poles (counted
including the multiplicity of each pole).
Remark 11.1.3. Note that s ( f ) 0 for  = 0; thus Proposition 11.1.2 does
alg
/ K
indeed recover the concavity inequality s s0 . Also, s ( f ) = 0 if
because the zeroes and poles of a rational function with coefficients in K must
be algebraic over K .

186

Variation of generic and subsidiary radii

11.2 Variation of Newton polygons

Before proceeding to differential modules, we study the variation of the
Newton polygon of a polynomial over K /t, t/ or K /t, t/an when
measured with respect to different Gauss valuations. We begin with this both
because it motivates the statements of the results for differential modules and
also because it will be used heavily in the proofs of those statements.
Theorem 11.2.1. Let P K /t, t/an [T ] be a polynomial of degree n. For
r [ log , log ], put vr () = log | |er . Let NPr (P) be the Newton
polygon of P under vr . Let f 1 (P, r ), . . . , f n (P, r ) be the slopes of NPr (P)
(listed with multiplicity) in increasing order. For i = 1, . . . , n, put Fi (P, r ) =
f 1 (P, r ) + + f i (P, r ).
(a) (Linearity) For i = 1, . . . , n, the functions f i (P, r ) and Fi (P, r ) are
continuous and piecewise affine in r . Moreover, even if = 0 there are
only finitely many different slopes.
(b) (Integrality) If i = n or f i (r0 ) < f i+1 (r0 ) then the slopes of Fi (P, r )
in some neighborhood of r = r0 belong to Z. Consequently, the slopes
of each f i (P, r ) and Fi (P, r ) belong to
1
1
Z Z.
1
n
(c) (Superharmonicity) Suppose that < 1 < . For i = 1, . . . , n, let
s,i (P) and s0,i (P) be the left-hand and right-hand slopes of Fi (P, r )
alg
at r = 0. For ( K ) , let s,i (P) be the right-hand slope of
Fi (T (P), r ) at r = 0. Then

s,i (P)
s,i (P),
alg

with equality if i = n or f i (P, 0) < f i+1 (P, 0).

(d) (Monotonicity) Suppose that P is monic and = 0. For i = 1, . . . , n,
the slopes of Fi (P, r ) are nonnegative.
(e) (Concavity) Suppose that P is monic. For i = 1, . . . , n, the function
Fi (P, r ) is concave.

n
Pi T i with Pi K /t, t/an . By Proposition 8.5.2
Proof. Write P = i=0
the function vr (Pi ) is continuous and concave in r and piecewise affine with
slopes in Z. Moreover, even if = 0 there are only finitely many different
slopes.
For s R and r [ log , log ], put
vs,r (P) = min{vr (Pi ) + is};
i

11.2 Variation of Newton polygons

187

that is, vs,r (P) is the y-intercept of the supporting line of NPr (P) of slope
s. Since vs,r (P) is the minimum of finitely many functions of the pair (r, s),
each of which is continuous, piecewise affine with only finitely many different
slopes, and concave, it also enjoys these properties. (Compare Remark 2.1.7.)
Note that Fi (P, r ) is the difference between the y-coordinates of the points
of NPr (P) having x-coordinates i n and n. That is,
Fi (P, r ) = sup{vs,r (P) (n i)s} vr (Pn ).

(11.2.1.1)

The supremum in (11.2.1.1) is achieved by some s whose denominator is

bounded by n. For any given r0 , for r in some neighborhood of r0 there can be
only finitely many values of s with denominator bounded by n achieving the
supremum in (11.2.1.1) for at least one value of r in the neighborhood. Consequently, Fi (P, r ) is continuous and piecewise affine with only finitely many
different slopes, proving (a).
If i = n or f i (P, r0 ) < f i+1 (P, r0 ) then the point of NPr0 (P) having
x-coordinate i n is a vertex, and likewise for r in some neighborhood of r0 .
In that case, for r near r0 ,
Fi (P, r ) = vr (Pni ) vr (Pn ),

(11.2.1.2)

proving (b).
Assume that < 1 < . Then Proposition 11.1.2 implies that

s (Pi )
(i = 0, . . . , n).
s (Pi ) =
alg

If i = n or f i (P, 0) < f i+1 (P, 0) then the above equation imply plus
(11.2.1.2) that the desired inequality is in fact an equality. Otherwise, let j, k
be the least and greatest indices for which f j (P, 0) = f i (P, 0) = f k (P, 0);
then j i < k, and the convexity of the Newton polygon implies that
Fi (P, r )

k i
i j +1
F j1 (P, r ) +
Fk (P, r ),
k j +1
k j +1

(11.2.1.3)

with equality for r = 0. From this plus piecewise affinity we deduce that
k i
i j +1
s (Pn j+1 ) +
s (Pnk ),
k j +1
k j +1
k i
i j +1
s,i (P)
s (Pn j+1 ) +
s (Pnk )
k j +1
k j +1

s,i (P)

alg

( K ),

yielding (c).
Assume that = 0 and that P is monic. Then each vr (Pi ) is a nondecreasing
function of r , as is each vs,r (P). Since vr (Pn ) = 0, Fi (P, r ) is nondecreasing
by (11.2.1.1), proving (d).

188

Variation of generic and subsidiary radii

To prove (e), one can reduce to working locally around r = 0 and then
deduce the claim from (c) and (d) (because the latter implies that s,i (P) 0
for  = 0). However, one can also prove (e) directly as follows. Assume that
P is monic, so that Pn = 1 and (11.2.1.1) reduces to
Fi (P, r ) = sup{vs,r (P) (n i)s}.
s

It is not immediately clear from this that Fi (P, r ) is concave, since we are taking the supremum rather than the infimum of a collection of concave functions.
To get around this, pick r1 , r2 [ log , log ] and put r3 = ur1 +(1u)r2
for some u [0, 1]. For j {1, 2}, choose an s j achieving the supremum in
(11.2.1.1) for r = r j . Put s3 = us1 + (1 u)s2 ; then, using the concavity of
vs,r (P) in both s and r , we have
Fi (P, r3 ) vs3 ,r3 (P) (n i)s3
u(vs1 ,r1 (P) (n i)s1 ) + (1 u)(vs2 ,r2 (P) (n i)s2 )
= u Fi (P, r1 ) + (1 u)Fi (P, r2 ).
This yields concavity for Fi (P, r ), proving (e).
Remark 11.2.2. A more geometric interpretation of the previous proof can

j
be given by writing each Pi as
j Pi, j t and considering the lower convex
hull of the set of points {(i, j, v(Pi, j ))} in R3 . We leave elaboration of this
point to the reader.
Remark 11.2.3. If i = n or f i (P, r0 ) <
implies that

f i+1 (P, r0 ) then (11.2.1.2)

f 1 (P, r0 ) + + f i (P, r0 ) v(K ) + Zr0 .

This fact does not analogize to subsidiary radii, because one has to replace
v(K ) by its p-divisible closure. See Theorem 10.7.1 and Example 10.7.2.
Remark 11.2.4. The conclusions of Theorem 11.2.1 carry over if we replace
f i (P, r ) by min{ f i (P, r ), ar + b} for any fixed a, b R (except for (b), for
which we need a, b Z). This is so because
i

j=1

min{ f i (P, r ), ar + b} = sup {vs,r (P) (n i)s} vr (Pn );

sar +b

in other words, the height of the relevant point is determined by the supporting
lines of slope less than or equal to ar +b, rather than by all slopes. Note that, in
the notation of the proof of Theorem 11.2.1(e), the inequality s j ar j + b for
j = 1, 2 implies the same for j = 3, so the proof of concavity goes through.
For bounded elements, we obtain a similar but slightly weaker conclusion.

11.3 Variation of subsidiary radii: statements

189

Theorem 11.2.5. Let P K /t, t/0 [T ] be a polynomial of degree n.

Then the conclusions of Theorem 11.2.1 continue to hold except that, if K is
not discrete, in (a) the functions f i (P, r ) and Fi (P, r ) are only piecewise affine
on the interior of [ log , log ] (with possibly infinitely many slopes).
Proof. The revised statement of (a) holds by Remark 8.2.4. For the other parts,
one may simply apply Theorem 11.2.1 over the ring K  /t, t/[T ] for all ,
with < < (in the case  = 0) or 0 = = < (in the case
= 0).

11.3 Variation of subsidiary radii: statements

In order to state the analogue of Theorem 11.2.1 for the subsidiary radii of
a differential module on a disc or annulus, we must set some corresponding
notation.
Notation 11.3.1. Let M be a finite free differential module of rank n
over K /t, t/, K /t, t/an , or K /t, t/an . For [, ], let
R1 (M, ), . . . , Rn (M, ) be the extrinsic subsidiary radii of M F in
increasing order, so that R1 (M, ) = R(M F ) is the generic radius of
convergence of M F . For r [ log , log ], define
f i (M, r ) = log Ri (M, er ),
so that f i (M, r ) r for all r . Put Fi (M, r ) = f 1 (M, r ) + + f i (M, r ).
We now have the following results, whose proofs are distributed across the
remainder of this chapter (Lemmas 11.5.1, 11.6.1, 11.6.3, and 11.7.1). Note
that there is an overall sign discrepancy with Theorem 11.2.1, so that concavity
becomes convexity and so forth. There are also some exceptions made in cases
where f i (M, r ) = r .
Theorem 11.3.2. Let M be a finite free differential module of rank n over
K /t, t/, K /t, t/an , or K /t, t/an .
(a) (Linearity) For i = 1, . . . , n, the functions f i (M, r ) and Fi (M, r ) are
continuous and piecewise affine. Moreover, even if = 0 there are only
finitely many different slopes.
(b) (Integrality) If i = n or f i (M, r0 ) > f i+1 (M, r0 ) then the slopes of
Fi (M, r ) in some neighborhood of r0 belong to Z. Consequently, the
slopes of each f i (M, r ) and Fi (M, r ) belong to
1
1
Z Z.
1
n

190

Variation of generic and subsidiary radii

(c) (Subharmonicity) Suppose that < 1 < and that f i (M, 0) > 0. For
i = 1, . . . , n, let s,i (M) and s0,i (M) be the left-hand and right-hand
alg
slopes of Fi (M, r ) at r = 0. For ( K ) , let s,i (M) be the

right-hand slope of Fi (T (M), r ) at r = 0. Then


s,i (M),
s,i (M)
alg

with equality if either i = n and f n (M, 0) > 0 or i < n and

f i (M, 0) > f i+1 (M, 0).
(d) (Monotonicity) Suppose that = 0. For i = 1, . . . , n, for any point r0
where f i (M, r0 ) > r0 the slopes of Fi (M, r ) are nonpositive in some
neighborhood of r0 . (Remember that if = 0 then f i (M, r ) = r for r
sufficiently large, by Proposition 9.3.3; see also Proposition 11.8.1.)
(e) (Convexity) For i = 1, . . . , n, the function Fi (M, r ) is convex.
Remark 11.3.3. Note that f i (M, r ) and Fi (M, r ) are defined using the extrinsic normalization. However, if we switch to the intrinsic normalization then
everything in Theorem 11.3.2 stays the same except for (d), in which the upper
bound on the slopes in a neighborhood of r0 changes from 0 to 1.
Remark 11.3.4. Suppose instead that M is a finite free differential module
of rank n over K /t, t/0 (resp. over K /t, t/0 ). We may then apply
Theorem 11.3.2 to M K  /t, t/ for all , with < < (in
the case  = 0 if we are working over K /t, t/0 ) or = <
(otherwise). This implies that all the conclusions of Theorem 11.3.2 continue
to hold for M itself, except that in (a) the functions f i (M, r ) and Fi (M, r ) are
defined only on ( log , log ] (resp. on ( log , log )). One can also
show that they extend continuously, and continue to be piecewise affine (with
finitely many slopes) if K is discrete, on the whole of [ log , log ]; see
Remark 11.6.5.

11.4 Convexity for the generic radius

As a prelude to tackling Theorem 11.3.2, we give a quick proof of subharmonicity, monotonicity, and convexity (parts (c)(e) of Theorem 11.3.2) for the
function f 1 corresponding to the generic radius of convergence. This argument
applies to both discs and annuli and can be used in place of the full strength
of Theorem 11.3.2 for many purposes; indeed, this is true for numerous results
which predate Theorem 11.3.2. See the notes for further details.

11.5 Measuring small radii

191

Proof of Theorem 11.3.2(c), (d), (e) for i = 1. Choose a basis of M, and let
Ds be the matrix of action of D s on this basis. Then recall from Lemma 6.2.5
that
R1 (M, ) = min{, lim inf |Ds |1/s
}.

1/s

For each s, the function r  log |Ds |er is convex in r by Proposition 8.2.3(b). This implies the convexity of

1/s
f 1 (M, r ) = max r, log + lim sup( log |Ds |er ) .
s

Similarly, we may deduce (c) by applying Proposition 11.1.2 to each Ds . If

1/s
= 0 then the function r  log |Ds |er is nonincreasing, yielding (d).
Remark 11.4.1. To improve upon this result, we will try to read off the generic
radius of convergence, and maybe even the other subsidiary radii, from the
Newton polygon of a cyclic vector. In order to do this, we have to circumvent
two obstacles.
(a) In general one can only construct cyclic vectors for differential
modules over differential fields, not over differential rings. (While
Theorem 5.7.3 produces cyclic vectors over certain rings, it only does
so locally for the Zariski topology, which appears to be insufficient for
this purpose.)
(b) If p > 0 some subsidiary radii may be greater than p1/( p1) , in
which case Newton polygons will not detect them.
The first problem will be addressed by using a cyclic vector over a fraction field
to establish linearity, integrality, and subharmonicity and then using a carefully
chosen lattice to deduce monotonicity and convexity. The second problem will
be addressed using Frobenius descendants.

11.5 Measuring small radii

In this section, we address concern (a) from Remark 11.4.1, using both cyclic
vectors and matrix inequalities.
Lemma 11.5.1. For any i {1, . . . , n} and any r0 such that f i (M, r0 ) >
r0 log , Theorem 11.3.2 holds in a neighborhood of r0 .
Proof. If M is defined over R, put F = Frac R. Choose a cyclic vector for
M R F to obtain an isomorphism M R F
= F{T }/F{T }P for some

192

Variation of generic and subsidiary radii

monic twisted polynomial P over F. We may then apply Corollary 6.5.4 and
Theorem 11.2.1 to deduce (a), (b), (c) of Theorem 11.3.2.
To deduce (d), we may work in a right-hand neighborhood of a single value
r0 of r . There is no harm in enlarging K (by Proposition 10.6.6), so we may
assume that v(K ) = R. Then we may reduce to the case r0 = 0 by replacing t
by t for some K .
Since v(K ) = R, we may pick c1 , . . . , cn K such that
log |c j | = min{ log f j (M, 0), 0}

( j = 1, . . . , n).

Let S F[U ] be the untwisted polynomial with the same coefficients as P,

and let 1 , . . . , n be the roots of S. By Corollary 6.5.4,
|ci | = max{|i |, 1} = max{e f j (M,0) , 1}

(i = 1, . . . , n).

We now construct a basis of M R F as in Theorem 6.5.3. Let B0 be the basis

of M R F given by
1
1
j
cn
cn1
jT

( j = 0, . . . , n 1).

Let N0 be the matrix of action of D on B0 ; it is a conjugated companion matrix,

of the form appearing in Proposition 4.3.10, corresponding to S. In particular,
the singular values of N0 are |c1 |, . . . , |cn1 |, |1 n /(c1 cn1 )|. The
latter equals |cn | if |cn | > 1 and otherwise is less than or equal to 1.
By Lemma 8.6.1 the supremum norm defined by B0 is also defined by
some basis B1 of M R K tan . Let N1 be the matrix of action of D on B1 .
Theorem 6.7.4 implies that, for r close to 0, the visible spectrum of M R Fer
is the multiset of those norms of eigenvalues of the characteristic polynomial
of N1 which exceed er . We may then deduce (d) from Theorem 11.2.1(d).
(Alternatively, one may replace K by a spherical completion, then tensor M
with K t0 , and use Lemma 8.6.2.)
We can deduce (e) from (c) and (d), as noted in the proof of
Theorem 11.2.1(e). It can also be proved directly as follows. We may again
assume that r0 = 0. We may also assume that < 1 < , as otherwise there
is nothing to check at r0 . Define B0 as above. This time, apply Lemma 8.6.1 to
construct a basis B1 of M R K 1/t, t defining the same supremum norm as
B0 . We may approximate B1 with a basis B1 of M R K  /t, t/ for some
< 1 < defining the same supremum norm with respect to
| |1 . (Here we use the facts that K /t, t/ is dense in K 1/t, t and that
any element of K /t, t/ which becomes a unit in K 1/t, t is already a
unit in K  /t, t/ for some < 1 < . The latter holds because
if the Newton polygon of an element of K  /t, t/ has no slope equal to 0,
then it also has no slopes in some neighborhood of 0.) Let N1 be the matrix of

11.6 Larger radii

193

action of D on B1 . Applying Theorem 6.7.4 to N1 , we may deduce (e) from

Theorem 11.2.1(e).

11.6 Larger radii

Next we address concern (b) from Remark 11.4.1, considering the cases
f i (M, r0 ) > r0 and f i (M, r0 ) = r0 separately. We temporarily omit monotonicity, as it requires a slightly different argument. (We must also say more
about Theorem 11.3.2(a); see Remark 11.6.4 below.)
Lemma 11.6.1. For any i {1, . . . , n} and any r0 such that f i (M, r0 ) > r0 ,
statements (a)(c), (e) of Theorem 11.3.2 hold in a neighborhood of r0 .
Proof. This holds by Lemma 11.5.1 in the case p = 0, so we may assume
p > 0 throughout the proof. For each nonnegative integer j, we will prove the
claim for r0 such that f i (M, r0 ) > r0 + (1/ p j ( p 1)) log p, by induction on
j; the base case j = 0 is precisely Lemma 11.5.1, so we may assume j > 0
hereafter. As in the proof of Lemma 11.5.1, we may reduce to the case r0 = 0.
Let R1 ( p ), . . . , R pn ( p ) be the subsidiary radii of M F in increasing
order. (The normalization is chosen in this way because the series variable in F is t p , which has norm p .) Put gi ( pr ) = log Ri (e pr ). By
Theorem 10.5.1, the list g1 ( pr ), . . . , g pn ( pr ) consists of

n
p
1
$
log p ( p 1 times)} f i (M, r ) r + p1
log p,
{ p f i (M, r ), pr + p1
i=1

{log p + ( p 1)r + f i (M, r ) ( p times)}

f i (M, r ) r +

1
p1

log p.

Thus we may deduce (a) from the induction hypothesis.

To check (b), (c), (e), it suffices to handle the cases where i = n or
f i (M, 0) > f i+1 (M, 0). (As in the proof of Theorem 11.2.1(c), we may linearly interpolate to establish convexity and subharmonicity in the other cases.)
In these cases, as in Corollary 10.5.3, we have at least one of f i (M, 0) >
(1/( p 1)) log p, in which case in some neighborhood of r = 0 we have
g1 ( pr ) + + g pi ( pr ) = p Fi (M, r ) + pi log p + ( p 1)i pr, (11.6.1.1)
or f i+1 (M, 0) < (1/( p 1)) log p, or i = n, in which last case we have in
some neighborhood of r = 0
g1 ( pr ) + + g pi+( p1)(ni) ( pr ) = p Fi (M, r ) + pn log p + ( p 1)npr.
(11.6.1.2)
Moreover, f i (M, 0) > (1/ p j ( p 1))) log p if and only if g pi (0) >
(1/ p j1 ( p 1)) log p.

194

Variation of generic and subsidiary radii

If f i (M, 0) > (1/( p 1)) log p, apply (11.6.1.1) and the induction hypothesis to write piecewise
Fi (M, r ) = (1/ p)(g1 ( pr ) + + g pi ( pr ) pi log p ( p 1)i pr )
= (1/ p)(m( pr ) + )
= mr + (1/ p)
for some m Z. (Note that is not guaranteed to be in pv(K ); this explains
Example 10.7.2.) Otherwise, we may apply (11.6.1.2) to write piecewise
Fi (M, r) = (1/ p)(g1 ( pr ) + + g pi+( p1)(ni) ( pr ) pn log p ( p 1)npr)
= (1/ p)(m( pr ) + )
= mr + (1/ p)

for some m Z.
Remark 11.6.2. In the proof of Lemma 11.6.1, note the importance of the
fact that the domains of applicability of (11.6.1.1) and (11.6.1.2) overlap: if
f i (M, 0) = (1/( p 1)) log p then (11.6.1.1) is valid for r = 0 but possibly
not for nearby r values.
The case f i (M, r0 ) = r0 remains inaccessible even using descent along a
Frobenius morphism, so we make an ad hoc argument.
Lemma 11.6.3. For any i {1, . . . , n} and any r0 such that f i (M, r0 ) = r0 ,
Theorem 11.3.2 holds in a neighborhood of r0 .
Proof. As in the proof of Lemma 11.5.1, it suffices to consider the case r0 =
0. We first check continuity. For this, note that the proofs of Lemma 11.5.1
and 11.6.1 show that, for any c > 0, the function max{ f i (M, r ), r + c} is
continuous at r = 0. Consequently, for any  > 0, we can find 0 < < /2
such that
| max{ f i (M, r ), r + /4}| < /2
(|r | < ).
For such r ,  < < r fi (M, r ) < ; this yields continuity.
Next we check convexity. Using Remark 11.2.4, the proofs of Lemmas 11.5.1 and 11.6.1 show that, for any c > 0, the function

i
j=1 max{ f j (M, r ), r + c} is convex. (The key point is that the domain over
which this holds does not depend on c.) Since this function tends to Fi (M, r )
as c tends to 0, we may deduce (e).
We now check piecewise affinity by induction on i. Given that
f 1 (M, r ), . . . , f i1 (M, r ) are affine in a one-sided neighborhood of r = 0,
say [, 0], and given that f i (M, 0) = 0, it suffices to check the linearity

11.7 Monotonicity

195

of f i (M, r ) r in some [  , 0]. By (e), the set of r [, 0] for which

f i (M, r ) r 0 is connected. Since f i (M, r ) r 0 always, it follows that
if f i (M, r0 ) = 0 for a single r0 [, 0] then f i (M, r ) = 0 for r [r0 , 0]
so, in particular, f i (M, r ) r is linear in a one-sided neighborhood of 0. Otherwise the slopes of f i (M, r ) r in [, 0) form a sequence of discrete values
which are negative and nondecreasing (by (e)). This sequence must then stabilize, so f i (M, r ) r is again linear in a one-sided neighborhood of 0. This
proves (a).
To prove (b), note that when f i (M, 0) = 0 the input hypothesis can only
hold if i = n. Suppose that we wish to check the integrality of the right slope
of Fn (M, r ) (the argument for the left-hand slope is analogous). If f 1 (M, r )
r, . . . , f n (M, r ) r are identically zero in a right-hand neighborhood of 0 then
we have nothing to check. Otherwise, let j be the greatest integer such that
f j (M, r ) r is not identically zero in a right-hand neighborhood of 0; we then
deduce (b) by applying Lemma 11.6.1 with i replaced by j.
Since (c) and (d) make no assertion at r = 0 in the case f i (0) = 0, we
are done.
Remark 11.6.4. Lemmas 11.6.1 and 11.6.3 fail to establish the last assertion of Theorem 11.3.2(a), i.e., if = 0 then each f i (M, r ) has only finitely
many differential slopes. However, this is easy to see from parts (b) and (e) of
the theorem. Namely, each Fi (M, r ) has slopes which are discrete and nondecreasing but also bounded above by i because f i (M, r ) = r for r large
(by Proposition 9.3.3). Hence each Fi (M, r ) has only finitely many different
slopes, as does each f i (M, r ).
Remark 11.6.5. Suppose that M is a finite free differential module of rank
n over K /t, t/0 (resp. over K /t, t/0 ). As noted in Remark 11.3.4,
Theorem 11.3.2 implies that the functions f i (M, r ) and Fi (M, r ) are continuous and piecewise affine on ( log , log ] (resp. on ( log , log )).
However, by imitating the proofs of Lemmas 11.5.1, 11.6.1, and 11.6.3, using
Theorem 11.2.5 in place of Theorem 11.2.1, we see that f i (M, r ) and Fi (M, r )
extend continuously to [ log , log ]. Moreover, if K is discretely valued
and = 1, the limits of the e fi (M,r ) as r 0+ are the subsidiary radii of
M E.

11.7 Monotonicity
To complete the proof of Theorem 11.3.2 we must prove (d) for p > 0 without
the restriction f i (M, r0 ) > r0 log . The reason why we do not have (d) as
part of Lemma 11.6.1 is that passing from M to M introduces a singularity

196

Variation of generic and subsidiary radii

at t = 0 (Remark 10.3.5), so we cannot hope to infer monotonicity on M.

To fix this, we must use off-center Frobenius descendants.
Lemma 11.7.1. If = 0 and f i (M, r0 ) > r0 then the slope of f i (M, r ) in a
right-hand neighborhood of r0 is nonpositive.
Proof. We may assume p > 0, as otherwise Lemma 11.5.1 implies the result.
We proceed as in the proof of Lemma 11.6.1 but using the off-center Frobenius
instead of . Again, we may assume that r0 = 0 and that i = n or
f i (M, 0) > f i+1 (M, 0) (reducing to the latter case by linear interpolation).
Let R1 ( p ), . . . , Rn ( p ) be the subsidiary radii of M F in increasing order. Put gi ( pr ) = log Ri (e pr ). By Theorem 10.8.3, if f i (M, 0) >
(1/( p 1)) log p then
g1 ( pr ) + + gi p ( pr ) = p Fi (M, r ) + i p log p,
whereas if f i+1 (M, 0) < (1/( p 1)) log p or i = n then
g1 ( pr ) + + gi p+( p1)(ni) ( pr ) = p Fi (M, r ) + np log p.
Moreover, f i (M, 0) > (1/ p j ( p 1)) log p if and only if gi p (0) >
(1/ p j1 ( p 1)) log p. To conclude, we may proceed as in Lemma 11.6.1.
Example 11.7.2. To see in action the discrepancy between the behavior of the
centered and off-center Frobenius descendants, we consider an example suggested by Liang Xiao. (All verifications are left as an exercise.) Take > 1,
and let M be the differential module over K t/ with a single generator
v satisfying D(v) = t p1 v. Pick any (0, 1), so that we may form
p1
M on K /t p , t p /. Then M splits as m=0 (M  Wm ), where M 



1
has a single generator v satisfying D (v ) = p v, and Wm is defined as in
alg
Definition 10.3.3. One then computes, for m  = 0 and K ,
s,1 (M  ) = 0,
s,1 (M  ) = 0,
s,1 (M  Wm ) = 0,
s0,1 (M  Wm ) = 1,
sm,1 (M  Wm ) = 1,
s,1 (M  Wm ) = 0
This yields

(  = 0, m).

s, p ( M) = 0,
s0, p ( M) = p 1,
s, p ( M) = 1
s, p ( M) = 0

( F
p ),
(
/ F p ),

11.8 Radius versus generic radius

197

and in turn
s,1 (M) = p + 1,
s0,1 (M) = 0,
s,1 (M) = 1
s,1 (M) = 0

( F
p ),
(
/ F p ).

11.8 Radius versus generic radius

As promised, we can recover some information about the radius of convergence from the properties of the generic radius of convergence.
Proposition 11.8.1. Let M be a finite differential module over K t/ or
K t/an for some > 0. Then the radius of convergence of M equals er ,
for r the smallest value in [ log , +) such that f 1 (r ) = r . Consequently
f 1 (r  ) = r  for all r  r .
Proof. By Theorem 9.6.1 the radius of convergence of M is on the one hand
at least the generic radius of convergence of M Fer , which by hypothesis
equals er . On the other hand, if > er then by hypothesis f 1 ( log ) >
log , or in other words R(M F ) < . This means that M K t/
cannot be trivial, so the radius of convergence cannot exceed . This proves
the desired result.
Corollary 11.8.2. Let M be a finite differential module over K t/ or
K t/an for some > 0. Then the radius of convergence of M belongs to
the divisible closure of the multiplicative value group of K .
Proof. By Theorem 11.3.2(a), (b) and Theorem 10.7.1, the function f 1 (r )
is piecewise of the form ar + b with a Q and b Qv(K ). By
Proposition 11.8.1 the radius of convergence of M equals er , for r the smallest value such that f 1 (r ) = r . To the left of this r , f 1 must be piecewise affine
with slope  = 1; by comparing the left and right limits at r we deduce that
r = ar + b for some rational a  = 1 and some b Qv(K ). Since this gives
r = b/(a 1), we deduce the claim.
One should be able to control the denominators better, as implied in the
following question. (The p = 0 analogue is easy; see the exercises.)
Question 11.8.3. Assume that p > 0. Let M be a finite differential module over K t/ for some > 0. Does there necessarily exist j
{1, . . . , rank(M)} such that the jth power of the radius of convergence of M
belongs to the p-divisible closure of the multiplicative value group of K ?

198

Variation of generic and subsidiary radii

We also have a criterion to establish when the radius of convergence equals

the generic radius.
Corollary 11.8.4. Let M be a finite differential module over K t/ or
K t/an for some > 0 such that, for some (0, ), R(M F ) is
constant for [, ]. Then R(M) = R(M F ).
Proof. The hypothesis implies that f 1 (M, r ) is constant in a right neighborhood of r = log . By Theorem 11.3.2(d), f 1 (M, r ), which is piecewise
affine, remains constant until it becomes equal to r . By Proposition 11.8.1, we
deduce the claim.

11.9 Subsidiary radii as radii of optimal convergence

The subsidiary generic radii of convergence can be interpreted as the radii of
convergence of a well-chosen basis of local horizontal sections at a generic
point. The argument is a variation on Corollary 11.8.4.
Definition 11.9.1. Let M be a differential module of rank n over K t/
or K t/an or on the open disc of radius . For i = 1, . . . , n, the ith
radius of optimal convergence of M at 0 is the supremum of those
[0, ) for which there exist n i linearly independent horizontal sections of
M K t/. (Remember that by Corollary 9.1.3 it is equivalent to require
linear independence of the horizontal sections over K or over K t/.)
Note that there exists a basis of local horizontal sections s1 , . . . , sn of M
such that si has radius of convergence equal to the ith radius of optimal convergence of M at 0: once si+1 , . . . , sn have been chosen, there must be at
least a one-dimensional space of choices left for si . Such a basis is sometimes
called an optimal basis of local horizontal sections. (It might be more consistent with our earlier terminology to refer to the radii of optimal convergence
as the subsidiary radii of convergence, but we have refrained from doing so
to avoid confusion with the subsidiary generic radii of convergence, which we
commonly abbreviate to subsidiary radii.)
The following generalizes Proposition 9.7.5.
Theorem 11.9.2 (after Young). Let (V, D) be a differential module over F
of dimension n with subsidiary radii r1 rn . Let L be a complete
extension of K , let t be a generic point of L relative to K of norm , and put
V  = V F L(t t )/an . Then the radii of optimal convergence of V  are
also r1 rn .

Notes

199

Proof. We first produce a basis s1 , . . . , sn for which si converges in the

open disc of radius ri around t for i = 1, . . . , n. For this, we may apply
Theorem 10.6.2 to decompose V into components each with a single subsidiary radius and thus reduce to the case r1 = = rn = r . By the geometric
interpretation of the generic radius (Proposition 9.7.5), each JordanHlder
constituent of V admits a basis of local horizontal sections on a generic disc of
radius r . By Lemma 6.2.8(a) the same is true for V itself.
It remains to check that there cannot exist n i linearly independent local
horizontal sections converging on a disc of radius strictly greater than ri . We
will prove this by induction on n. Let m be the largest integer such that r1 = rm .
Let V1 be the component of V of subsidiary radius r1 , so that dim V1 = m.
We will check that no local horizontal section of V1 at t can have radius of
convergence strictly greater than r1 .
Put
f i (r ) = f i (V1 F L(tt )/an , r )

(i = 1, . . . , m; r [ log , +)).

By Theorem 11.3.2(c) the f i (r ) are constant in a neighborhood of r = log .

By Theorem 11.3.2(c), (e) and induction on i,

log ri 0 < r log ri ,
f i (r ) =
r
r log ri .
However, if there were a local horizontal section of V1 at t which converged
on a closed disc of radius for some (r1 , ) then V1 F L(t t )/
would have a trivial submodule, and so it would have as one of its subsidiary
radii. This would force f n (r ) = r for r = log < log ri , a contradiction.
We conclude that any local horizontal section of V that projects nontrivially onto V1 has radius of convergence at most r1 . If i m, we are done;
otherwise, any linearly independent set of n i local horizontal sections converging in an open disc of radius greater than ri must project to zero in V1 . We
may thus reduce to applying the induction hypothesis to the complementary
component.

Notes
The harmonicity property of functions on annuli (Proposition 11.1.2) may
be best viewed as a theory of subharmonic functions on one-dimensional
Berkovich analytic spaces. Such a theory has been developed by Thuillier
[205] with a view towards applications in Arakelov theory. A related development is the work of Favre and Jonsson concerning potential theory on

200

Variation of generic and subsidiary radii

the valuative tree, with applications to the theory of plurisubharmonic singularities on complex surfaces. See [87] and [88] and also a more recent
paper by Boucksom, Favre, and Jonsson [33] giving some higher-dimensional
generalizations.
For the function f 1 (M, r ) = F1 (M, r ) representing the generic radius of
convergence, Christol and Dwork established convexity [49, Proposition 2.4]
(using essentially the same short proof as that given here) and continuity
at endpoints [49, Thorme 2.5] (see also [80, Appendix I]) in the case of
a module over a full annulus. In the case of analytic elements, continuity
and piecewise affinity were conjectured by Dwork and proved by Pons [181,
Thorme 2.2]. The analogous results for the higher Fi (M, r ) are original.
When restricted to intrinsic subsidiary radii less than , Theorem 11.9.2 is
a result of Young [222, Theorem 3.1]. Youngs proof is an explicit calculation
using twisted polynomials and cyclic vectors.
As suggested earlier (see the notes for Chapter 9), Youngs definition of
the radii of optimal convergence suggests an analogue in the framework of
Baldassarri and Di Vizio. For instance, given a differential module on a closed
disc of radius , the radii of optimal convergence at a generic point of radius
(0, ) should be defined in terms of an optimal basis of local horizontal
sections which are allowed to extend all the way across the disc of radius
rather than just across the disc of radius . However, we do not know how to
prove the appropriate analogue of Theorem 11.3.2 for these quantities, since
working over F loses all information beyond radius . Accomplishing this is
an important open problem.1

Exercises
(1) Give an example to show that, in Theorem 11.2.1, f 2 need not be concave
(even though f 1 and f 1 + f 2 are concave).
(2) Verify Example 11.7.2.
(3) Give another proof of (9.9.5.1) as follows. First find one value 0 for
which Theorem 6.5.3 implies that I R(M F0 ) = 0b . Then use
Theorem 11.3.2 to show that (9.9.5.1) holds for [1, 0 ].
(4) State, then answer affirmatively (using Theorem 6.5.3), the analogue of
Question 11.8.3 for p = 0.

1 Added in proof: A solution to this problem has been announced by Baldassarri.

12
Decomposition by subsidiary radii

In the previous chapter we established a number of important variational properties of the subsidiary radii of a differential module over a disc or annulus.
In this chapter, we continue the analysis by showing that, under suitable conditions, one can separate a differential module into components of different
subsidiary radii. That is, we can globalize the decompositions by spectral
radius provided by the strong decomposition theorem, if a certain numerical
criterion is satisfied.
As in the previous chapter, our discussion begins with some observations about power series, in this case identifying criteria for invertibility.
We use these in order to set up a Hensel lifting argument to give the desired
decompositions; again we must start with the visible (see Definition 6.5.1) case
and then extend using Frobenius descendants. We end up with a number of distinct statements, covering open and closed discs and annuli as well as analytic
elements.
As a corollary of these results we recover an important theorem of Christol and Mebkhout. That result gives a decomposition by subsidiary radii on
an annulus in a neighborhood of a boundary radius at which the module is
solvable, that is, all the intrinsic subsidiary radii tend to 1. (It is not necessary to assume that the annulus is closed at this boundary.) One may view our
results as a collection of quantitative refinements of the ChristolMebkhout
theorem.
Note that nothing is this chapter is useful if the intrinsic subsidiary radii are
everywhere equal to 1. We will tackle this case in Chapter 13.
Throughout this chapter, besides Notation 10.0.1 we also retain Notation 11.3.1.

201

202

Decomposition by subsidiary radii

12.1 Metrical detection of units

One can identify the units in rings such as K /t, t/ rather easily in terms of
power series coefficients (Lemma 8.2.6). However, for the present application
we need an alternate characterization based on metric data, i.e., Gauss norms.
Definition 12.1.1. For f K /t, t/ with 1 , define the
discrepancy of f at r = 0 as the sum

disc( f, 0) =
s ( f );
alg

( K )

note that disc( f, 0) 0 because it is a sum of nonnegative terms. We define

disc( f, r ) for general r [ log , log ] by rescaling: assume without loss
of generality that K contains a scalar c of norm er , let Tc : K /t, t/
K (er )/t, t/(er ) be the substitution t  ct, and then put
disc( f, r ) = disc(Tc ( f ), 0).
Lemma 12.1.2. For x K t/ nonzero and c K of norm , x is a unit if
and only if s0 (Tc (x)) = disc(x, log ) = 0.
Proof. We may reduce to the case = 1 and |x|1 = 1. In this case, by
Lemma 8.2.6, x is a unit if and only if its image modulo m K in K [t] is a
unit. As noted in Proposition 11.1.2, the order of vanishing of this image at
alg
K is precisely s (x); this proves the claim.
There is also a variant for bounded series and analytic elements which is
slightly simpler.
Lemma 12.1.3. For x K t/0 or x K t/an nonzero, x is a unit if and
only if s0 (Tc (x)) = 0.
Proof. Again, we reduce to the case = 1 and |x|1 = 1. Since |x| is nonincreasing, s0 (Tc (x)) = 0 if and only if |x| = 1 for all (0, 1). This is true
if and only if the constant term of x has norm 1, which happens if and only if
x is a unit in o K t or o K t F1 (since these are both local rings).
For annuli, it is more convenient to prove a weak criterion first.
Lemma 12.1.4. For x (0,) K /t, t/ nonzero, x is a unit if and only
if we have disc(x, log ) = 0.
Proof. We may again reduce to the case = 1 and |x|1 = 1. In this case, by
Lemma 8.2.6, x is a unit if and only if its image modulo m K in K [t, t 1 ] is a
unit. We then argue as in Lemma 12.1.2.

12.2 Decomposition over a closed disc

203

One may then deduce the following.

Lemma 12.1.5. For > 0 and for x K /t, t/ nonzero, x is a unit
if and only if the function r  vr (x) is affine on [ log , log ] and
disc(x, log ) = disc(x, log ) = 0.
Proof. Note that, by Proposition 11.1.2, r  vr (x) is affine on [ log ,
log ] if and only if disc(x, r ) = 0 for r ( log , log ). We may thus
reformulate the desired result as follows: x is a unit if and only if disc(x, r ) = 0
for all r [ log , log ].
If x is a unit then disc(x, r ) = 0 for all r [ log , log ] by
Lemma 12.1.4. Conversely, given the latter condition, to check that x is a
unit, it suffices by Remark 8.1.3 to check that x is a unit in K i /t, t/i 
for a finite collection of closed intervals [i , i ] with union [, ]. However,
Lemma 12.1.4 implies that one can cover a one-sided neighborhood of any element of [, ] with such an interval; the compactness of [, ] (Lemma 8.0.4)
then yields the claim.
Remark 12.1.6. Another statement in this vein is the fact that (0,)
K /t, t/an is a field (Corollary 8.5.3).

12.2 Decomposition over a closed disc

We consider decomposition by subsidiary radii first in the case of a closed disc.
The numerical criterion in this case involves an analogue of the discrepancy
function from the previous section.
Definition 12.2.1. Let M be a finite differential module over K /t, t/ with
1 . Define the ith discrepancy of M at r = 0 as

s,i (M);
disci (M, 0) =
alg

( K )

it is always nonnegative by Theorem 11.3.2(d). Extend the definition to define

disci (M, r ) for general r [ log , log ] as in Definition 12.1.1.
Theorem 12.2.2. Let M be a finite differential module over K t/ of rank n.
Suppose that the following conditions hold for some i {1, . . . , n 1}.
(a) We have f i (M, log ) > f i+1 (M, log ).
(b) The function Fi (M, r ) is constant for r in a neighborhood of log .
(c) We have disci (M, log ) = 0.

204

Decomposition by subsidiary radii

Then there is a direct sum decomposition of M inducing, for each (0, ],

the decomposition of M F separating the first i subsidiary radii from the
others.
Before proving Theorem 12.2.2 we record some observations which will
simplify the proof.
Remark 12.2.3. To prove Theorem 12.2.2 it suffices to lift to M the decomposition of M F separating the first i subsidiary radii. Namely, suppose
that M1 M2 is this decomposition and that M1 F accounts for the first i
subsidiary radii of M F . By Theorem 11.3.2(a), for r in a neighborhood of
log , M1 Fer accounts for the first i subsidiary radii of M Fer . Consequently, for r in a neighborhood of log we have f j (M, r ) = f j (M1 , r )
for j = 1, . . . , i and f j (M, r ) = f ji (M2 , r ) for j = i + 1, . . . , n.
By Theorem 11.3.2(d), (e), the function Fi (M1 , r ) is convex and nonincreasing as long as f i (M1 , r ) > r . For r in a neighborhood of log ,
Fi (M1 , r ) = Fi (M, r ) by the previous paragraph, and Fi (M, r ) is constant
by hypothesis. Hence Fi (M1 , r ) must remain constant until the first value of
r1 for which f i (M1 , r1 ) = r1 . Then, for log r r1 , f i (M1 , r ) =
Fi (M1 , r ) Fi1 (M1 , r ) is nondecreasing by Theorem 11.3.2(d), whereas for
log r f 1 (M2 , r ) is nonincreasing until it becomes equal to r and then
stays equal to r thereafter. Hence f i (M1 , r ) > f 1 (M2 , r ) for log r < r1
and r1 = f i (M1 , r1 ) f 1 (M2 , r1 ) r1 ; thus also f i (M1 , r ) r = f 1 (M2 , r )
for r r1 . Consequently, M1 Fer accounts for the first i subsidiary radii of
M Fer for all r .
We next make a trivial but quite useful observation.
Lemma 12.2.4. Let R, S, T be subrings of a common ring U with S T = R.
Let M be a finite free R-module. Then the intersection (M R S) (M R T )
inside M R U is equal to M itself.
This also holds when M is only locally free; see the exercises.
Remark 12.2.5. Lemma 12.2.4 allows us to replace K by a complete extension L in the course of proving Theorem 12.2.2; thus, inside the completion of
L(t) for the -Gauss norm we have
F Lt/ = K t/
(exercise). Thus, obtaining matching decompositions of M K t/ F and
M K t/ Lt/ gives a corresponding decomposition of M itself.

12.2 Decomposition over a closed disc

205

For annuli, we will use the related fact that, for any [, ], inside the
completion of L(t) for the -Gauss norm we have
F L/t, t/ = K /t, t/
(exercise).
We also need a lemma about polynomials over K t.

Lemma 12.2.6. Let P = i Pi T i and Q = i Q i T i be polynomials over

K t satisfying the following conditions.
(a) We have |P 1|1 < 1.
(b) For m = deg(Q), Q m is a unit and |Q|1 = |Q m |1 .
Then P and Q generate the unit ideal in K t[T ].
Proof. We may assume without loss of generality that Q m = 1. The hypothesis
that |Q|1 = |Q m |1 = 1 implies that all the slopes of the Newton polygon of Q
under | |1 are nonnegative. Hence, by Lemma 2.3.1, if R K t[T ] and S is
the remainder upon dividing R by Q then |S|1 |R|1 .
Let Si denote the remainder upon dividing (1 P)i by Q. By the previ

Si converges and its limit S satisfies P S 1

ous paragraph, the series i=0
(mod Q). This proves the claim.
We are now ready to make the key step of establishing Theorem 12.2.2 in
the visible range (see Definition 6.5.1).
Lemma 12.2.7. Theorem 12.2.2 holds if f i ( log ) > log log .
Proof. By invoking Remark 12.2.5 to justify enlarging K and then rescaling,
we may reduce to the case = 1. We may also assume that K has value
group R.
Set the notation as in that part of the proof of Lemma 11.5.1 dealing with
Theorem 11.3.2(d). Let Q(T ) be the characteristic polynomial of N1 , so that
the Newton polygon of Q computes f 1 (M, r ), . . . , f i (M, r ) in a neighborhood of r = 0 (from the same part of the proof of Lemma 11.5.1). Condition
(a) of Theorem 12.2.2 implies that these Newton polygons all have a vertex,
with x-coordinate n i, whose position is determined by the coefficient of
T ni in Q. Conditions (b) and (c) then imply that this coefficient satisfies
the hypothesis of Lemma 12.1.2 and so is a unit in K t. We can thus apply
Theorem 2.2.1 to factor Q as Q 2 Q 1 , so that the roots of Q 1 are the i largest
roots of Q under | |1 .
Use the basis B1 to identify M with K tn , so that we may view N1 as a
K t-linear endomorphism of M. By Lemma 12.2.6 applied after rescaling,
Q 1 and Q 2 generate the unit ideal in K t[T ]. Hence M splits as a direct sum

206

Decomposition by subsidiary radii

M1 M2 of modules (but not differential modules) with Mi = ker(Q i (N1 )).
We now invoke some of the approximation lemmas, as follows. Equip M with
the supremum norm compatible with | |1 defined by B1 , and then equip
M1 , M2 with the induced quotient norms; these are both supremum-equivalent
by Lemma 1.3.5. By Lemma 1.3.7, for any c > 1 we can approximate these
norms to within a factor of at most c by supremum norms defined by bases
of M1 K t F1 and M2 K t F1 . By Lemma 8.6.1 the latter norms are also
defined by bases B1,1 , B1,2 of M1 , M2 .
Let U be the change-of-basis matrix from B1 to B1,1 B1,2 , so that U 1 N1 U
is a block diagonal matrix and |U |1 , |U 1 |1 c. Then the matrix of action of
D on B1,1 B1,2 is U 1 N1 U + U 1 d(U ). If we write this in block form as


A B
C D
then, by taking c sufficiently close to 1, we may force the following conditions
to hold.
(a) The matrix A is invertible and |A1 |1 max{|d|1 , |B|1 , |C|1 , |D|1 } < 1.
(b) The Newton slopes of A under | |1 account for the first i subsidiary
radii of M K t F1 .
As in the proof of Lemma 6.7.3 we may now use Lemma 6.7.1 to produce a
submodule of M accounting for the last n i subsidiary radii of M K t F1 .
By repeating this argument for M , we obtain a submodule of M accounting
for the first i subsidiary radii of M K t F1 . By Remark 12.2.3 this suffices to
prove the desired result.
To prove Theorem 12.2.2 in general, we must use Frobenius descendants
again.
Proof of Theorem 12.2.2. The claim holds by Lemma 12.2.7 if p = 0, so we
may assume p > 0 throughout the proof. It suffices to prove that, for = 1,
Theorem 12.2.2 holds if f i (M, 0) > 1/( p j ( p 1)) log p for each nonnegative
integer j; we again proceed by induction on j, the base case j = 0 being
provided by Lemma 12.2.7.
Suppose that f i (M, 0) > 1/( p j ( p 1)) log p for some j > 0. Let M1 M2
be the decomposition of M separating the subsidiary radii less than or equal
to e p fi (M,0) into M1 (which exists by the induction hypothesis). This might
not be induced by a decomposition of M1 , because some factors of subsidiary
radius p p/( p1) that are needed in M2 are instead grouped into M1 . To correct
this, consider instead the decomposition
(0 (M1 ) p1 (M1 )) (0 (M2 ) + + p1 (M2 )),

12.3 Decomposition over a closed annulus

207

where m is defined as in Lemma 10.3.6(a). By Lemma 10.3.6(b) this decomposition is induced by a decomposition of M. By Theorem 10.5.1, it has the
desired effect.
Remark 12.2.8. As in Lemma 11.5.1 one can prove Lemma 12.2.7 using
Lemma 8.6.2 in place of Lemma 8.6.1, at the expense of some extra complications. First, one must replace K by a spherical completion with value group R
and algebraically closed residue field (by invoking Theorem 1.5.3). One then
obtains the desired decomposition only over K t0 . By gluing with the original decomposition (defined over F1 ), we recover a decomposition over K tan .
To remove poles in the disc |t | < 1 we apply the same argument with M
replaced by T (M) (in the sense of Definition 11.1.1).

12.3 Decomposition over a closed annulus

Over a closed annulus, one has a decomposition theorem of a shape somewhat
different from that over a closed disc. Fortunately, the proof is essentially the
same as for Theorem 12.2.2.
Theorem 12.3.1. Let M be a finite differential module over K /t, t/ of
rank n, for some 0 < . Suppose that the following conditions hold for
some i {1, . . . , n 1}.
(a) We have f i (M, r ) > f i+1 (M, r ) for log r log .
(b) The function Fi (M, r ) is affine for log r log .
(c) We have disci (M, log ) = disci (M, log ) = 0.
Then there is a direct sum decomposition of M inducing, for each [, ],
the decomposition of M F separating the first i subsidiary radii from the
others.
First we prove a lemma which looks somewhat more like Theorem 12.2.2.
Lemma 12.3.2. Let M be a finite differential module over K /t, t/ of
rank n. Suppose that the following conditions hold for some i {1, . . . , n 1}.
(a) We have f i (M, log ) > f i+1 (M, log ).
(b) We have disci (M, log ) = 0.
Then, for some [, ), there is a direct sum decomposition of M
K  /t, t/ inducing, for each [ , ], the decomposition of M F
separating the first i subsidiary radii from the others.
Proof. Using Remark 12.2.5 again, we may enlarge K and then reduce
to the case = 1. Moreover, it suffices to consider the case where

208

Decomposition by subsidiary radii

f i (M, 0) > log , as we may reduce the general case to this one as in the
proof of Theorem 12.2.2.
Again set the notation as in that part of the proof of Lemma 11.5.1 dealing with Theorem 11.3.2(d), but take = 1; as in the part dealing with
Theorem 11.3.2(e), we can find a basis B1 of M K /t,t K  /t, t for some
[, 1) defining the same supremum norm as B0 . Let N1 be the matrix of
action of D on B1 . By conditions (a) and (b) of the lemma plus Lemma 12.1.4,
the coefficient of T ni in the characteristic polynomial of N1 is a unit in
K  /t, t for some [, 1). We may thus continue as in the proof of
Lemma 12.2.7.
To prove Theorem 12.3.1 from Lemma 12.3.2, we proceed as in the proof
of Lemma 12.1.5. However, we must first give an alternate formulation of the
hypotheses of the theorem.
Remark 12.3.3. In the statement of Theorem 12.3.1, given condition (a) we
may reformulate conditions (b) and (c) together as the following condition.
(a) We have disci (M, r ) = 0 for log r log .
To see this, note that if < 1 < then condition (a) implies that equality
holds in Theorem 11.3.2(c), so we have s0,i (M) s,i (M) = disci (M, 0).
Consequently Fi (M, r ) is affine in a neighborhood of 0 if and only if
disci (M, 0) = 0. By rescaling, we obtain the desired equivalence. (Compare
the proof of Lemma 12.1.5.)
Proof of Theorem 12.3.1. The case = proceeds exactly as in
Theorem 12.2.2, so we will assume < hereafter. By Remark 12.3.3,
if M satisfies the given hypothesis then so does M K  /t, t/ for each
closed subinterval [ , ] [, ]. For each (, ], Lemma 12.3.2 implies
that, for some [, ), M K  /t, t/ admits a decomposition with
the desired property. Similarly, for each [, ) and for some (, ],
M K /t, t/  admits a decomposition with the desired property.
By the compactness of [, ] (Lemma 8.0.4), we can cover [, ] with
finitely many intervals [i , i ] for which M K i /t, t/i  admits a decomposition with the desired property. Since the decomposition of M K i /t, t/i 
is uniquely determined by the induced decomposition over F for any single
[i , i ], these decompositions agree on overlaps of the covering intervals. By the gluing lemma (Lemma 8.3.6), we obtain a decomposition of M
itself.
Remark 12.3.4. As in Remark 12.2.8, to prove Lemma 12.3.2 one may use
Lemma 8.6.2 in place of Lemma 8.6.1. Again, one enlarges K to be spherically complete with value group R and algebraically closed residue field. One

12.4 Decomposition over an open disc or annulus

209

then notes that F = <1 K  /t, tan is a field (Corollary 8.5.3), so one can
approximate the basis B0 of M K /t,t F1 with a basis B1 of M K /t,t F
defining the same supremum norm. One then obtains a decomposition of
M K /t,t K  /t, tan for some , and one can remove the unwanted poles
as in Remark 12.2.8.

12.4 Decomposition over an open disc or annulus

Over open discs, we have similar decomposition theorems but without the
discrepancy conditions at endpoints.
Theorem 12.4.1. Let M be a finite differential module of rank n over the
open disc of radius . Suppose that the following conditions hold for some
i {1, . . . , n 1} and some (0, ).
(a) The function Fi (M, r ) is constant for log < r log .
(b) We have f i (M, r ) > f i+1 (M, r ) for log < r log .
Then M admits a unique decomposition separating the first i subsidiary radii
of M F for [ , ).
Proof. As in Remark 12.3.3, note that (a) and subharmonicity
(Theorem 11.3.2(c)) imply that disci (M, ) = 0 for ( , ). Thus, for
any such , we may apply Theorem 12.2.2 to M K t/; doing so for all
such (or a sequence increasing to ) yields the desired result.
Similarly, for open annuli we obtain a decomposition theorem without a
discrepancy condition at endpoints.
Theorem 12.4.2. Let M be a finite differential module of rank n over the
open annulus of inner radius and outer radius . Suppose that the following
conditions hold for some i {1, . . . , n 1}.
(a) The function Fi (M, r ) is affine for log < r < log .
(b) We have f i (M, r ) > f i+1 (M, r ) for log < r < log .
Then M admits a unique decomposition separating the first i subsidiary radii
of M F for any (, ).
Proof. By Remark 12.3.3 the conditions of Theorem 12.4.2 are satisfied by
M K  /t, t/ whenever < < . Gluing together the resulting
decompositions yields the desired result.
Remark 12.4.3. One can also obtain a decomposition theorem for a half-open
annulus, by covering the half-open annulus with an open annulus and a closed
annulus and then gluing together the decompositions given by Theorems 12.3.1

210

Decomposition by subsidiary radii

and 12.4.2. Similarly, one can obtain decomposition theorems on more exotic
subspaces of the affine line by gluing; the reader knowledgeable enough to be
interested in such statements should at this point have no trouble formulating
and deriving them. See also Remark 12.5.3 below.

12.5 Partial decomposition over a closed disc or annulus

We next state decomposition theorems which apply to a closed disc or annulus
without requiring a discrepancy condition. The price one must pay is that one
must work with analytic elements.
Theorem 12.5.1. Let M be a finite differential module of rank n over
K t/an . Suppose that the following conditions hold for some i {1, . . . ,
n 1}.
(a) The function Fi (M, r ) is constant in a neighborhood of r = log .
(b) We have f i (M, log ) > f i+1 (M, log ).
Then M admits a direct sum decomposition separating the first i subsidiary
radii of M F for (0, ].
Proof. As for Theorem 12.2.2, except that Lemma 12.1.2 is replaced by
Lemma 12.1.3.
Theorem 12.5.2. Let M be a finite differential module of rank n over
K /t, t/an , for some < . Suppose that the following conditions hold
for some i {1, . . . , n 1}.
(a) The function Fi (M, r ) is affine for log r log .
(b) We have f i (M, r ) > f i+1 (M, r ) for log r log .
Then M admits a direct sum decomposition separating the first i subsidiary
radii of M F for [, ].
Proof. To obtain a decomposition of M K  /t, t/an for some (, ),
we may proceed as in Theorem 12.3.1, Lemma 12.1.4 being replaced by
Remark 12.4.3. We will omit further details.
Remark 12.5.3. Readers familiar with affinoid algebras should be able to
extend the results of this section to cases where M is defined over the ring
of analytic elements for a disc contained in a one-dimensional affinoid space.
(We leave even the definition of this ring as an unstated exercise.)
If K is discrete, one can also extend to modules defined over K t/0 or
K /t, t/0 . However, if K is not discretely valued then one runs into various difficulties associated with the fact that E is no longer a field; compare
Remark 8.2.4.

12.6 Modules solvable at a boundary

211

Remark 12.5.4. If p = 0, it should be possible to get a decomposition theorem over K /t, t/0 even without assuming that f i (M, r ) >
f i+1 (M, r ) for r { log , log }; a statement along these lines (for K
discretely valued) appears in [144]. However, this fails completely if p > 0;
one can generate numerous counterexamples using the theory of isocrystals
(Chapter 23).

12.6 Modules solvable at a boundary

One of the most important special cases of our decomposition theorems occurs
in the following setting, which occurs frequently in applications.
Definition 12.6.1. Let M be a finite differential module on the half-open
annulus with closed inner radius and open outer radius . We say that
M is solvable at if R(M F ) as or, equivalently, if
I R(M F ) 1 as . (In a similar definition the roles of the inner
and outer radius are reversed; we will not refer to that definition here.)
Lemma 12.6.2. Let M be a finite differential module on the half-open annulus
with closed inner radius and open outer radius which is solvable at .
There exist b1 bn [0, +) such that, for [, ) sufficiently
close to , the intrinsic subsidiary radii of M F are (/)b1 , . . . , (/)bn .
Moreover, if i = n or bi > bi+1 then b1 + + bi Z.
Proof. For r ( log )+ , Fi (M, r ) ir is a convex function
(Theorem 11.3.2(e)) with slopes in a discrete subset of R (Theorem 11.3.2(a),
(b)). Moreover, it is nonnegative and its limit is 0; this implies that the slopes
are all nonnegative. Hence these slopes must eventually stabilize; that is, each
f i (M, r ) becomes linear in a neighborhood of log . This provides the existence of b1 , . . . , bn ; by Theorem 11.3.2(b), if i = n or bi > bi+1 then
b1 + + bi Z.
Definition 12.6.3. Let M be a finite differential module on the half-open annulus with closed inner radius and open outer radius which is solvable at .
The quantities b1 , . . . , bn defined by Lemma 12.6.2 are called the differential
slopes of M at . (They are also called ramification numbers; the reason for
this will become clear when we consider quasiconstant differential modules in
Chapter 19. See specifically Theorem 19.4.1.)
We now recover a decomposition theorem of Christol and Mebkhout; see
the notes for further discussion. We will see several applications of this result
later in the book.

212

Decomposition by subsidiary radii

Theorem 12.6.4 (ChristolMebkhout). Let M be a finite differential module

on the half-open annulus with closed inner radius and open outer radius
which is solvable at . Then, for any sufficiently large [, ), the restriction of M to the open annulus with inner radius and outer radius splits
uniquely as a direct sum b[0,+) Mb , such that for each b [0, +) and for
all [ , ) the intrinsic subsidiary radii of Mb F are all equal to (/)b .
Proof. By Lemma 12.6.2, this is a case where Theorem 12.4.2 may be
applied.
Remark 12.6.5. For some differential modules for which one has fairly
explicit series expansions for local horizontal sections, one may be able to
establish solvability at a boundary by explicit estimates. (A related strategy
appears in Example 9.6.2.) However, it is more common for solvability to be
established by proving the existence of a Frobenius structure; this notion will
be introduced in Chapter 17.
Remark 12.6.6. Be aware that, in Theorem 12.6.4, if M is the restriction of a
differential module over the closed annulus with inner radius and outer radius
, or over the ring of analytic elements on the open annulus, the decomposition
of M given by the theorem need not descend back to this original structure.
Compare Remark 12.5.4.

12.7 Solvable modules of rank 1

We now give a partial classification of modules of rank 1 on an open annulus
which are solvable at a boundary. Assume that p > 0.
Definition 12.7.1. Fix a coherent system of p-power roots of 1 in K alg ; that
is, for each h, the chosen p h+1 th root of unity should be a pth root of the
h
chosen p h th root of unity. For c o
K and n a positive integer write n = mp
with m coprime to p, and let Mn,c be the pullback of the module Mh defined
in Example 9.9.3 along the map t  ct m (using the chosen p h+1 th root of
unity); this module is solvable at 1.
Theorem 12.7.2. Let b be a positive integer. Assume that K contains the p h th
roots of unity for all h log p b. Let M denote a finite differential module of
rank 1 on a half-open annulus with open outer radius 1 which is solvable
at 1 with differential slope b. Then there exist c1 , . . . , cb {0} o
K and
nonnegative integers j1 , . . . , jb such that
M ( j1 ) (M1,c1 ) ( jb ) (Mb,cb )
has differential slope 0.

12.7 Solvable modules of rank 1

213

We will refine this statement a little later by eliminating the Frobenius

pullbacks (Theorem 17.1.6).
Proof. By Theorem 11.3.2(b) we have b Z. It thus makes sense to proceed
by induction on b; we may assume that b > 0. Pick 0 < < < 1 such that,
for some nonnegative integer j,
p p

j+1 /( p1)

< I R(M F ) = b < I R(M F ) = b < p p

j /( p1)

By Theorem 10.4.4, M admits a j-fold Frobenius antecedent N on

j
j
j
j
K  p /t p , t p / p . Pick a generator v of N , and put D(v) = nv. Then
j
|n| p j = p 1/( p1) p (b+1) for [, ], so in this range n is dominated
by a term of the form n b1 (t p )b1 with |n b1 | = p 1/( p1) . Let be the
chosen pth root of unity, write b = p h m with m coprime to p, and take
j

cb =

n b1
;
m( 1)

then
I R((N Mb,cb ) F p j ) > p

jb

( [, ]).

We may use the log-concavity of the intrinsic radius (Theorem 11.3.2(e)) to

deduce that the differential slope of N Mb,cb is strictly less than b. Thus the
induction hypothesis gives the desired result.
Corollary 12.7.3. Let M denote a finite differential module of rank 1 on a halfopen annulus with open outer radius 1 which is solvable at 1 with differential
slope b > 0. If M p has differential slope 0 then b is not divisible by p.
Proof. There is no harm in enlarging K in such a way that the hypotheses of Theorem 12.7.2 are satisfied; thus we may assume that M
=
( j1 ) (M1,c1 ) ( jb ) (Mb,cb ). By Remark 9.9.4 and Theorem 10.4.2, for
cb o
K:
p
if b is coprime to p then Mb,cb has differential slope 0;
p

if b is divisible by p then Mb,cb has differential slope b/ p.

Moreover, these differential slopes are preserved under Frobenius pullback, by
Theorem 10.4.2. This implies the desired result as follows. If b were divisible by p then ( jb ) (Mb,cb ) p would have differential slope b/ p whereas
( j1 ) (M1,c1 ) p ( jb1 ) (Mb1,cb1 ) p would have differential slope
strictly less than b/ p. By Lemma 9.4.6(c), M p would have differential slope
b/ p > 0, a contradiction.

214

Decomposition by subsidiary radii

12.8 Clean modules

It is reasonable to ask whether the refined strong decomposition theorem
(Theorem 10.6.7) admits an analogue over a disc or annulus. The following discussion provides a possible answer to this question, though not a
definitive one.
Lemma 12.8.1. Let M be a finite free differential module of rank n over
K /t, t/an . Suppose that Fn (M, r ) is affine on [ log , log ]. Let
j {0, . . . , n} be the smallest integer such that f i (M, r ) is affine on
[ log , log ] for all i > j. If 0 < j < n then f j (M, r ) > f j+1 (M, r )
for all r ( log , log ); if j = n then f j (M, r ) > r for all
r ( log , log ).
Proof. There is nothing to check if j = 0, and j = 1 is impossible, so we
may assume that j > 1. Note that F j (M, r ) = Fn (M, r ) f j+1 (M, r )
f n (M, r ) is affine by hypothesis, and that F j1 (M, r ) is convex by
Theorem 11.3.2(e), so f j (M, r ) = F j (M, r ) F j1 (M, r ) is concave. Moreover f j (M, r ) is bounded below by the affine function f j+1 (M, r ) if j < n or
by the affine function r if j = n. Hence if this inequality becomes an equality for any interior point of [ log , log ] then it must hold identically,
contrary to the choice of j. This proves the claim.
Definition 12.8.2. Let M be a finite differential module of rank n over
K /t, t/an . We say that M is clean if Fn (M, r ) and Fn 2 (M M, r ) are
both affine on [ log , log ].
Theorem 12.8.3. Let M be a finite clean differential module of rank n over
K /t, t/an .
(a) For i {1, . . . , n}, Fi (M, r ) is affine (and so f i (M, r ) is affine).
(b) For i {1, . . . , n 1}, either f i (M, r ) = f i+1 (M, r ) for all
r ( log , log ) or f i (M, r ) > f i+1 (M, r ) for all r
( log , log ). In addition, either f n (M, r ) = r for all r
( log , log ) or f i (M, r ) > r for all r ( log , log ).
(c) For i {1, . . . , n 1}, if f i (M, r ) > f i+1 (M, r ) for r
{ log , log } then M admits a direct sum decomposition separating the first i subsidiary radii of M F , for [, ].
Proof. To check (a) we may replace K by a field with algebraically closed
residue field and value group R. Take j as in Lemma 12.8.1, and suppose by
way of contradiction that j > 0 (which in turn forces j > 1). Pick a point r0
( log , log ) at which f j (M, r ) fails to be affine or, equivalently, where
F j1 (M, r ) fails to be affine; we may rescale to reduce to the case r0 = 0.

12.8 Clean modules

215

Then, by Theorem 11.3.2(c), on the one hand we must have s, j1 (M) < 0
for some K .
On the other hand, we have f j (M, 0) > 0 if j > 0 (by Lemma 12.8.1) and
hence s, j (M) = 0 (by Theorem 11.3.2(c) again). Similarly, if h {0, . . . , n 2 }
denotes the smallest index for which f i (M M, r ) = r identically for i > h,
we have f h (M M, 0) > 0, if h > 0, and s,h (M M) = 0.
Choose a lift of in K , and set N = T (M) as a differential module over
K tan . Note that Fh (N N , r ) is constant for r in a right-hand neighborhood
of 0; f h (N N , 0) > 0 if h > 0; and f i (N N , r ) = r for i > h. By
Theorem 12.5.1 there exists a direct sum decomposition P0 P1 of N N
such that, for each (0, 1), P0 F accounts for the first h subsidiary
radii of (N N ) F . By Theorem 9.6.1, P1 restricts to a trivial differential
module over the open unit disc.
For any (0, 1), any direct sum decomposition of N F is defined
by projectors which are horizontal elements of (N N ) F . For sufficiently close to 1 the subsidiary radii of P0 F are all strictly less than
(by Lemma 12.8.1 again), so the projectors must belong to P1 F . Since P1
is trivial on the open unit disc, the projectors must also extend to horizontal
elements of N N over the open unit disc. That is, they define a direct sum
decomposition of N over the open unit disc.
It follows that, over the open unit disc, N admits a direct sum decomposition i Ni in which, for each i and each (0, 1) sufficiently close to 1,
Ni F has only a single subsidiary radius. Namely, given any decomposition not satisfying this condition, we can apply Theorem 10.6.2 and then the
previous paragraph to obtain a finer decomposition. (This can only be repeated
as many times as the rank of N .)
Let S be the set of indices i for which limr 0+ f 1 (Ni , r ) f j (M, 0). Since
f j (M, 0) > 0, in a neighborhood of r = 0 we have that f 1 (Ni , r ) is affine
and nonincreasing for each i S, by Theorem 11.3.2(a), (d), and F j (N , r )
is a positive linear combination of these functions. However, the right-hand
slope of F j (N , r ) at r = 0 is s, j (M) = 0, so f 1 (Ni , r ) must be constant in a
neighborhood of r = 0 for each i S. In a neighborhood of r = 0, F j1 (N , r )
is a nonnegative linear combination of the f 1 (Ni , r ) for i S, so it also has
right-hand slope 0 at r = 0. But this contradicts the fact that s, j1 (M) < 0.
This contradiction leads to the conclusion that j = 0, which implies (a).
Given (a), we may deduce (b) from the fact that f i (M, r ) and f i+1 (M, r ) are
affine functions on [ log , log ] satisfying f i (M, r ) f i+1 (M, r ) (or,
in the case i = n, the same argument with f i+1 (M, r ) replaced by r ). Given
(a) and (b), the hypothesis of (c) implies that f i (M, r ) > f i+1 (M, r ) for all
r [ log , log ], so the claim follows from Theorem 12.5.2.

216

Decomposition by subsidiary radii

Notes
Some results described here have also been obtained recently by Christol [46].
However, since we did not have access to Christols definitive manuscript at
the time of this writing, we are unable to provide detailed references.
Our results on modules solvable at a boundary are originally due to Christol and Mebkhout [51, 52]. In particular, Lemma 12.6.2 for the generic
radius is [51, Thorme 4.2.1] and the decomposition theorem (which implies
Lemma 12.6.2 in general) is [52, Corollaire 2.41]. However, the proof technique of Christol and Mebkhout is significantly different from ours: they
construct the desired decomposition by exhibiting convergent sequences for
a certain topology on the ring of differential operators. This does not appear to
give quantitative results; that is, the range over which the decomposition occurs
is not controlled, although we are not sure whether this is an intrinsic limitation of the method. (Keep in mind that the our approach here crucially uses our
method of Frobenius descendants, which was not available when [51, 52] were
written.)
Note also that Christol and Mebkhout work directly with a differential module on an open annulus as a ring-theoretic object; this requires a freeness result
of the following form. If K is spherically complete, any coherent locally free
module on the half-open annulus with closed inner radius and open outer
radius is induced by a finite free module over the ring [,) K /t, t/.
(That is, any coherent locally free sheaf on this annulus is freely generated by
global sections.) See the notes for Chapter 8 for further discussion.
A partial extension of the work of Christol and Mebkhout can be found in a
paper of Pons [181]. However, in the absence of a theory of Frobenius descendants, Pons was forced to impose somewhat awkward hypotheses in order to
avoid the exceptional value p 1/( p1) for the generic radius of convergence.
For the attribution of Theorem 12.7.2, see the notes for Chapter 17.
The notion of a clean differential module is original and was motivated
partly by discussions with Liang Xiao. It is a first attempt to model in p-adic
differential theory a higher-rank analogue of Katos notion of cleanness for
a rank-1 tale sheaf [119]. Similar considerations in residual characteristic 0
appear in [144].

Exercises
(1) Prove an analogue of Lemma 12.2.4 in which M is required only to be
locally free (in the sense of Remark 5.3.4).
(2) Let L be a complete extension of K . Prove that, for any > 0, within the
completion of L(t) for the -Gauss norm we have

Exercises
F Lt/ = K t/,

217

F L/t, t/ = K /t, t/.

(Hint: use Remark 8.5.5.)

(3) Use the previous exercise to deduce that, for any [, ], within the
completion of L(t) for the -Gauss norm we have
F L/t, t/ = K /t, t/.
(4) (a) Prove that, for x K t0 nonzero, the following are equivalent: (i) the
element x is a unit; (ii) the function vr (x) is constant in a neighborhood
of r = 0; (iii) vr (x) is constant for all r [0, +).
(b) Prove that, for x (0,1) K /t, t0 nonzero, x is a unit if and
only if the function r  vr (x) is affine in some neighborhood of
r = 0. In fact, this always happens if K is discretely valued, since
the Newton polygon of any x K /t, t0 has finite width in this
case (see the exercises for Chapter 8). Hence, in this case the ring
(0,1) K /t, t0 is a field. We will encounter this field again under
the name of the bounded Robba ring; see Definition 15.1.2.

13
p-adic exponents

In this chapter we study p-adic differential modules in a situation left untreated

by our preceding analysis, namely when the intrinsic generic radius of convergence is equal to 1 everywhere. (This condition is commonly called the
Robba condition.) This setting is loosely analogous to the study of regular singularities of formal meromorphic differential modules considered in
Chapter 7; in particular, there is a meaningful theory of p-adic exponents in
this setting.
However, some basic considerations indicate that p-adic exponents must
necessarily be more complicated than the exponents considered in Chapter 7.
For instance, the p-adic analogue of the Fuchs theorem (Theorem 7.3.8)
can fail unless we impose a further condition: that the differences between
exponents must not be p-adic Liouville numbers.
With this in mind we may proceed to construct p-adic exponents for differential modules satisfying the Robba condition. Such modules carry an action
of the group of p-power roots of unity via Taylor series; under favorable circumstances the module splits into isotypical components for the characters of
this group. We may identify these characters with elements of Z p , and these
give the exponents.
Throughout this chapter, we retain Notation 10.0.1 and assume that p > 0.

13.1 p-adic Liouville numbers

Definition 13.1.1. For K , the type of , denoted type(), is the radius of
convergence of the power series


m=0,m=

218

tm
.
m

13.1 p-adic Liouville numbers

219

This cannot exceed 1, as there are infinitely many m for which | m| =

max{||, 1}. (For instance, we may take m pZ p if || 1 and m 1 +
/ Z p , then | m| is bounded below, so
pZ p if || < 1.) Moreover, if
type() = 1. Thus we will concentrate mostly on Z p .
Definition 13.1.2. We say that K is a p-adic Liouville number if either
or has type less than 1 and a p-adic non-Liouville number otherwise. The
explicit reference to and is not superfluous, as they may have different
types (exercise).
The following alternative characterization of type may be helpful.
Definition 13.1.3. For Z p , let (m) be the unique integer in {0, . . . , p m 1}
congruent to modulo p m .
Proposition 13.1.4. For Z p not a nonnegative integer,

(m)
1
= lim inf
.
log p type() m+ m

(13.1.4.1)

In particular, has type 1 if and only if (m) /m + as m +.

Proof. It suffices to check that, for 0 < < 1, we have
lim sup(m + (m) log p ) =

(13.1.4.2)

lim sup(m + (m) log p ) = +

(13.1.4.3)

when < type() and

when > type(). Namely (13.1.4.2) implies that m + (m) log p 0

for all large m, and so lim infm (m) /m 1/(log p ), whereas
(13.1.4.3) implies that m + (m) log p 0 for infinitely many m, and so
lim infm (m) /m 1/(log p ).
Suppose first that type() > > 0; then, as s , s /| s| 0 or
equivalently v p ( s) + s log p . (Here v p denotes the renormalized valuation with v p ( p) = 1.) Since is not a nonnegative integer we have
(m) + as m +, and so
v p ( (m) ) + (m) log p .

220

p-adic exponents

The left-hand side does not increase if we replace v p ( (m) ) by m, so we

may deduce (13.1.4.2).
Suppose next that type() < < 1; then we may choose a sequence s j such
that, as j , v p ( s j ) + s j log p +. Put m j = v p ( s j ), so that
s j (m j ) . Then
m j + (m j ) log p +,
yielding (13.1.4.3).
The alternate characterization is convenient for such verifications as the fact
that rational numbers are non-Liouville (exercise), or the following stronger
result [80, Proposition VI.1.1], whose proof we omit.
Proposition 13.1.5. Any element of Z p algebraic over Q is non-Liouville.
Later, we will encounter the p-adic Liouville property in yet another,
apparently different, form. (See the exercises for an alternate proof of this
lemma.)
Lemma 13.1.6. For not a nonnegative integer, we have the following
equality of formal power series:


m=0


xm
(x)m 1
= ex
.
(1 )(2 ) (m )
m! m
m=0

Proof. The coefficient of x m on the right-hand side is a sum of the form

m
i=0 ci /(i ) for some ci Q. It is thus a rational function of of the
form P()/((1 ) (m )), where P has coefficients in Q and degree
at most m. To check that in fact P() = 1 identically, we need only check this
for = 0, . . . , m.
In other words, to check the original identity it suffices to check after multiplying both sides by i and evaluating at = i for each nonnegative integer
i. On the left-hand side we obtain


m=i

x m
.
(1)i1 i!(m i)!

On the right-hand side we obtain

ex
which is the same thing.

(x)i
,
i!

13.2 p-adic regular singularities

221

Corollary 13.1.7. If K is not a nonnegative integer and type() = 1 then

the series


m=0

xm
(1 )(2 ) (m )

has radius of convergence at least p 1/( p1) .

13.2 p-adic regular singularities

We now consider a p-adic analogue of Theorem 7.3.8. Unlike its archimedean
analogue, it requires a hypothesis on exponents beyond simply that they are
weakly prepared (which means that no two eigenvalues of the constant matrix
differ by a nonzero integer).
Definition 13.2.1. We say that a finite set is p-adic non-Liouville if its
elements are p-adic non-Liouville numbers. We say the set has p-adic nonLiouville differences if the difference between any two elements of the set is a
p-adic non-Liouville number.

Theorem 13.2.2 ( p-adic Fuchs theorem for discs). Let N = i=0

Ni t i be
an n n matrix over K t/ for some > 0. Assume that N0 has weakly
prepared eigenvalues with p-adic non-Liouville differences. Then there exists
> 0 such that the fundamental solution matrix for N , which exists and is
unique by Proposition 7.3.6, has its entries in K t/  (the same holds for its
inverse).
Proof. Recall that the fundamental solution matrix U is computed by the
recursion (7.3.6.1):
N0 Ui Ui N0 + iUi =

i


N j Ui j

(i > 0).

j=1

There is no harm in enlarging K to include the eigenvalues 1 , . . . , n of N0 .

By Lemma 7.3.5 the map X  i + N0 X X N0 has eigenvalues g h +
i for g, h {1, . . . , n}. View the operator X  N0 X X N0 as a linear
transformation T on the space V of n n matrices over K . The matrix of
action of (i + T )1 on some basis of V can be written as a matrix of cofactors
of i + T divided by the determinant of i + T . If we fix the basis of V then the
entries of i + T are bounded independently of i, as are the cofactors; we thus
obtain a bound of the form

222

p-adic exponents

|Ui | i c|N | max{|U j | j }

j<i

n


|g h + i|1

g,h=1

for some c > 0 not depending on i. (For a somewhat more careful argument in
this vein, see Proposition 18.1.1.)
Thus, to conclude the theorem it suffices to verify that, for each g, h
{1, . . . , n}, the quantity = g h has the property that
m


max{1, | i|1 }

i=1

grows at worst exponentially. If

/ Z p then | i|1 is bounded above and
the claim is verified. Otherwise Corollary 13.1.7 and the hypothesis that is a
p-adic non-Liouville number give the desired estimate.
By a slight modification of the argument (which we omit), one may obtain
the following result of Clark [54, Theorem 3], which may viewed as a p-adic
analogue of Corollary 7.3.11.
Theorem 13.2.3 (Clark). Let M be a finite differential module over K t/
for the derivation td/dt, with a regular singularity at 0 whose exponents are
p-adic non-Liouville numbers. Then, for any x M and y M K t/ K t
such that D(y) = x, we have D(y) M K t/ K t/ for some > 0.
Remark 13.2.4. The conclusion of Theorem 13.2.2 remains true, with the
same proof, if it is assumed only that the pairwise differences between eigenvalues of N0 all have type greater than 0. However, it is possible for the
conclusion to fail otherwise. To construct a counterexample put a = b = 1,
choose c Z p with type(c) = 0, and consider the differential module
associated with the hypergeometric differential equation (0.3.2.2). Then the
eigenvalues of N0 are 0 and c, whose difference has type 0; correspondingly,
the hypergeometric series (0.3.2.1), i.e.,
F(a, b; c; z) =


a(a + 1) (a + i)b(b + 1) (b + i)
i=0

c(c + 1) (c + i)i!

zi ,

gives rise to a formal horizontal section with radius of convergence 0.

13.3 The Robba condition

Given a finite differential module on an annulus for the derivation td/dt, we
would like to be able to tell whether it extends over a disc with a regular singularity. It turns out that when the exponents of that singularity are constrained

13.4 Abstract p-adic exponents

223

to lie in Z p (e.g., if they are rational numbers with denominators prime to p),
one gets a strong necessary condition from the generic radius of convergence
function.
Definition 13.3.1. Let M be a finite differential module on the disc or annulus |t| I , for I an interval. We say that M satisfies the Robba condition if
I R(M F ) = 1 for all nonzero I .
Proposition 13.3.2. Let M be a finite differential module on the open disc
of radius for the derivation td/dt satisfying the Robba condition in some
annulus. Then the eigenvalues of the action of D on M/t M belong to Z p .

Proof. Let N = i=0

Ni t i be the matrix of action of D on some basis of M.
/ Z p ; there is no harm in enlarging K to
Suppose that N0 has an eigenvalue
force K . Choose v M such that the image of v in M/t M is a nonzero
eigenvector of N0 of eigenvalue . Let D  be the derivation corresponding to
d  = d/dt instead of td/dt. Then with notation as in Example 9.5.2 we have,
for any < ,



1/( p1)
,
max lim sup |(D  )s v|1/s
, |d |sp,F |D |sp,V , > p
s

so that I R(M F ) < 1 by Lemma 6.2.5.

We will establish a partial converse to Proposition 13.3.2 later
(Theorem 13.7.1). In the interim, we mention the following easy result.
Proposition 13.3.3. Let M be a finite differential module, on the open disc of
radius for the derivation td/dt, such that the matrix of action N0 of D on
some basis of M has entries in K . Then M satisfies the Robba condition if and
only if N0 has eigenvalues in Z p .
Proof. Exercise, or see [80, Corollary IV.7.6].

13.4 Abstract p-adic exponents

In the previous section, we considered a finite differential module on an annulus for the derivation td/dt and saw that the Robba condition was necessary
for extending the module over a disc with a regular singularity at t = 0. Moreover, the exponents of that regular singularity must belong to Z p . We may then
ask whether it is possible to identify these exponents by looking only at the
original annulus.
The answer to this question is complicated by the fact that the exponents are
only well defined as elements of the quotient Z p /Z. This means that we cannot

224

p-adic exponents

hope to identify them using purely p-adic considerations; in fact, we must use
archimedean considerations to identify them. Here are those considerations.
Definition 13.4.1. We will say that A, B Znp are equivalent (or sometimes
strongly equivalent) if there exists a permutation of {1, . . . , n} such that
Ai B (i) Z for i = 1, . . . , n. This is evidently an equivalence relation.
Definition 13.4.2. We say that A, B Znp are weakly equivalent if there exists
a constant c > 0, a sequence 1 , 2 , . . . of permutations of {1, . . . , n}, and
signs i,m {1} such that
(i,m (Ai Bm (i) ))(m) cm

(i = 1, . . . , n; m = 1, 2, . . . ).

In other words, Ai Bm (i) has a representative modulo p m of size at most

cm. Again, this is clearly an equivalence relation and equivalence implies weak
equivalence.
Lemma 13.4.3. If A, B Z p (regarded as 1-tuples) are weakly equivalent
then they are equivalent.
Proof. For some c > 0, we have
|1,m+1 (1,m+1 (A B))(m+1) 1,m (1,m (A B))(m) | 2cm + c,
and the left-hand side is an integer divisible by p m . For m large enough, we
have p m > 2cm + c and so
1,m+1 (1,m+1 (A B))(m+1) = 1,m (1,m (A B))(m) .
Hence, for m large enough, 1,m is constant and 1,m (A B) is a constant
nonnegative integer.
Corollary 13.4.4. Suppose that A Znp is weakly equivalent to h A for some
positive integer h. Then A (Z p Q)n .
Proof. We are given that, for some c > 0, some permutations m , and some
signs i,m ,
(i,m (Ai h Am (i) ))(m) cm.
The order of m divides n!, so we have
((Ai h n! Ai ))(m) n!cm
for some choice of sign (depending on i, m). That is, for each i the 1-tuple
consisting of (h n! 1)Ai is weakly equivalent to zero. By Lemma 13.4.3,
(h n! 1)Ai Z, so Ai Z p Q.

13.5 Exponents for annuli

225

Proposition 13.4.5. Suppose that A, B Znp are weakly equivalent and that
B has p-adic non-Liouville differences. Then A and B are equivalent.
Proof. There is no harm in replacing B by an equivalent tuple in which
Bi B j Z if and only if Bi = B j . For some c and m , we have, for all m,
((Ai Bm (i) ))(m) cm,
((Ai Bm+1 (i) ))(m+1) c(m + 1),
and so
((Bm (i) Bm+1 (i) ))(m) 2cm + c.
By hypothesis, the difference Bm (i) Bm+1 (i) is either zero or a p-adic
non-Liouville number which is not an integer; for m large, the previous
inequality is inconsistent with the second option, by Proposition 13.1.4, so
Bm (i) = Bm+1 (i) . That is, for m large we may take m = for some fixed
, so
((Ai B (i) ))(m) cm

(m = 1, 2, . . . ).

By Lemma 13.4.3 Ai B (i) Z, so A and B are equivalent.

Example 13.4.6. It is easy to give examples of sets which are weakly
equivalent but not equivalent; here is an example of Dwork (from [161, Example 2.3.2]). Let be an increasing function on the nonnegative integers such
that, for some  > 0, we have (i + 1) p (i)  for all i 0. Put
=


i=0

(2i)

p (2i+1) .

i=0

One may verify (exercise) that the pairs (, ) and ( , 0) are weakly
equivalent but not equivalent. In this case , , + , are all p-adic
Liouville numbers.

13.5 Exponents for annuli

We now give an abstract definition of the p-adic exponents of a differential module on an annulus satisfying the Robba condition, after a motivating
remark.
Remark 13.5.1. Let N be a differential module over K t/ for the derivation
td/dt, such that the matrix of action, N0 , of D on N has entries in K and
eigenvalues in Z p . By Proposition 13.3.3, N satisfies the Robba condition.

226

p-adic exponents

Our best hope at this point would be to prove that any differential module M
over K /t, t/ satisfying the Robba condition is isomorphic to such an N .
This turns out to be rather too much to ask for, owing to difficulties arising with
Liouville numbers (not to mention the closed boundary), but for the moment
let us postulate the existence of an isomorphism N K /t, t/
= M and
see what it tell us.
Assume (for clarity) that K contains the group p of all p-power roots of
unity (i.e., all roots of unity of orders that are powers of p) in K alg . Since M
and N satisfy the Robba condition, a Taylor series construction gives an action
of p ; recall that we have used the p-power action already in the construction of Frobenius antecedents (Theorem 10.4.2). The K -valued characters of
p may be naturally identified with Z p , by identifying a Z p with the map
carrying p to a ; we can use the structure of N to decompose M into
character spaces for this action. Namely, perform a shearing transformation if
needed (Proposition 7.3.10) to ensure that N0 has prepared eigenvalues. Then
choose the basis of N so that the matrix N0 splits into blocks corresponding
to individual eigenspaces V j for the action of D (so, in particular, the K -span
of the chosen basis of N equals j V j ). Let j Z p be the eigenvalue corresponding to V j . We may then identify M with ( j V j ) K K /t, t/; under
this identification t i V j is an eigenspace with character j + i.
The strategy for constructing p-adic exponents is then to turn this argument
on its head, by first constructing the eigenspaces for the actions of the finite
subgroups of p and then extracting elements of these which stabilize at a
basis of M. One might expect an obstruction to arise because the full action of
1 + m K by Taylor series need not be semisimple, but this does not occur; see
the exercises for a statement that may help to give the reason.
Following the strategy discussed in the previous remark, we now introduce
the definition of exponents.
Definition 13.5.2. Let M be a finite differential module of rank n over
K /t, t/ satisfying the Robba condition. For m > 0, let pm denote the
group of pth roots of unity in K alg , and put p = m>0 pm . For p ,
define the action of on M K /t,t/ K ( )/t, t/ via the Taylor
series
(x) =


( t t)i i
D (x);
i!
i=0

this series converges because of the Robba condition. Fix a basis e1 , . . . , en

of M. An exponent for M is an element A Znp for which there exists

13.5 Exponents for annuli

227

a sequence {Sm,A }
m=1 of n n matrices over K /t, t/ satisfying the
following conditions.

(a) For j = 1, . . . , n and m = 1, 2, . . . , put vm,A, j =

i (Sm,A )i j ei .
Then, for all pm , (vm,A, j ) = A j vm,A, j .
(b) For some k > 0, we have |Sm,A | p km for all m and all [, ].
(This condition will be used to obtain some sort of convergence for the
Sm,A .) 

(c) We have det(Sm,A ) 1 for all m and all [, ]. (This condition
prevents the Sm,A from converging to zero.)
We make the following observations. w
(i) The property of being an exponent does not depend on the choice of
basis, although the matrices Sm do depend on the basis.
(ii) If A is an exponent for M then so is any B Znp equivalent to A.
(iii) To obtain (b), it is enough to have |Sm,A | cp km for some
c, k > 0, as we can enlarge k to absorb c. In fact it is enough to
check this just for = , , by the Hadamard three circles theorem
(Proposition 8.2.3(c)).
Before constructing exponents in general, we note the following extension
of Remark 13.5.1.
Proposition 13.5.3. Let M be a differential module of rank n over K t/
for the derivation td/dt, such that the eigenvalues of D on M/t M are in Z p .
Then, for any (0, ), these eigenvalues form an exponent for M K t/
K /t, t/ (which satisfies the Robba condition by Proposition 13.3.3).
Proof. By applying shearing transformations (Proposition 7.3.10) as needed,
we may reduce to the case where the eigenvalues of D on M/t M are prepared.
Let e1 , . . . , en be a basis of M reducing modulo t to a basis of M/t M consisting of bases for the generalized eigenspaces of D. For j = 1, . . . , n and
m = 1, 2, . . . , let A j be the eigenvalue corresponding to e j , and put

vm,A, j = p m
A j (e j ).
pm

Then the resulting matrices Sm,A have the desired property (exercise).
For the general construction of exponents, we need the following lemma due
to Dwork and Robba. It will be easiest to postpone its proof until Chapter 18,
when we will derive it as a corollary of some explicit convergence bounds on
solutions of p-adic differential systems (Corollary 18.2.5).

228

p-adic exponents

Lemma 13.5.4. Let V be a finite differential module of rank n over F for

the derivation d/dt, such that I R(V ) = 1. Choose a basis of V and, for
i = 1, 2, . . . , let Ni be the matrix of action of D i /i! on this basis. Then
|Ni | i p (n1)

log p i"

max{1, |N1 |n1

}

(i = 1, 2, . . . ).

We now give the general construction of exponents, using a discrete Fourier

transform for the group Z/ p m Z. (Compare [79, Lemma 3.1 and Corollary 3.3]
or [161, Proposition 3.1.1].)
Theorem 13.5.5. Let M be a finite differential module of rank n over
K /t, t/ satisfying the Robba condition. Then there exists an exponent
for M.
Proof. Fix a basis e1 , . . . , en of M. For any A Znp and any positive integer
m, we wish to define the matrix Sm,A such that vm,A, j corresponds to the projection of e j into the eigenspace of pm for the character A j . To achieve this,
we let E( ) be the matrix of action of p , and put

E( ) Diag( A1 , . . . , An ).
Sm,A = p m
pm

This matrix is invariant under Gal(K ( p )/K ) and so has entries in

K /t, t/. By the vector interpretation, it satisfies condition (a) of
Definition 13.5.2. (Another way to see this is to check the identity
E( ) (Sm,A ) = Sm,A Diag( A1 , . . . , An )

( pm )

and use the formula for change of basis in a difference module. See
Remark 14.1.3 below.)
We will check that Sm,A satisfies (b) of Definition 13.5.2 using
m+1 /( p1)
by
Lemma 13.5.4. For each pm , we have | 1| p p
Example 2.1.6. If we write
E( ) =

( 1)i t i Ni

i=0

then, under | | , the ith summand is bounded under | | by

c(m,i) for
max{1, |N1 |n1
}p
c(m, i) = (n1) log p i

i p m+1
i p m
= (n1)m+(n1) log p (i p m )
.
p1
p( p 1)

Here c(m, i) (n 1)m is a function of i p m which is continuous on (0, +)

and tends to as i p m tends to either 0 or +, so it is bounded above

13.5 Exponents for annuli

229

independently of m. Hence |E( )| p km for some k > 0, implying a similar

bound for Sm,A (but with a larger k).
We next choose A to satisfy (c) of Definition 13.5.2. Note that
vm,A,1 vm,A,n = det(Sm,A )e1 en
and that, for b {0, . . . , p 1}n ,
vm,A, j =

p1


vm+1,A+ pm b, j

( j = 1, . . . , n).

b=0

(In words this second equation states that each eigenspace for the action of
pm is the direct sum of p eigenspaces for the action of pm+1 .) Hence, for
any A Znp and any m 0, we have

det(Sm,A ) =
det(Sm+1,A+ pm b ).
(13.5.5.1)
b{0,..., p1}n

i
we can choose b such that
Write
Sm,A = iZ Sm,A,i
 t . By (13.5.5.1)


det(Sm+1,A+ pm b,0 ) det(Sm,A,0 ). Since S0,A is the identity matrix, this


det(Sm,A,0 )
satisfy
allows
us
to
choose
A
such
that
the
matrices
S
m,A






det(S0,A,0 ) = 1 for all m. Since det(Sm,A ) det(Sm,A,0 ) for all

[, ], this yields (c) of Definition 13.5.2.

We also have the following limited uniqueness result for the exponents; here
we see the first appearance of a non-Liouville condition. We also must begin
to assume that the annulus has positive width. (Compare [79, Theorem 4.4].)
Theorem 13.5.6. Assume that < . Let M be a finite differential module of
rank n over K /t, t/ satisfying the Robba condition. Then any two exponents for M are weakly equivalent. In particular, if M admits an exponent A
with non-Liouville differences then (by Lemma 13.4.3) any other exponent for
M is equivalent to A.
Proof. Fix a basis for M, let A, B be two exponents for M, and let Sm,A , Sm,B
be the corresponding sequences of matrices (for the same constant k > 0).

For j = 1, . . . , n, put vm, j,A = i (Sm,A )i ei and vm, j,B = i (Sm,B )i ei . Let
1
be the change-of-basis matrix between the bases vm, j,A and
Tm = Sm,A Sm,B

1
vm, j,B of M. Since |S 1 | |Sm,A |n1 det(Sm,A ) (from a description of
m,A

the inverse matrix using cofactors), we have



det(Sm,A )1 p nkm
|Tm | |Sm,B | |Sm,A |n1

( [, ]).
(13.5.6.1)

230

p-adic exponents

We will make progress by levering (in a metaphorical sense) the upper bound
(13.5.6.1) against various lower bounds involving Tm , using the log-convexity
of the Gauss norm as a function of the radius, i.e., the Hadamard three circles
theorem (Proposition 8.2.3(c)).
To begin with, we have
n

1 
|det(Tm )| det(Sm,A )  Sm,A  p nkm
( [, ]).

Put = . From the additive formula for the determinant of Tm , there must
be a permutation m of {1, . . . , n} such that
n


|Ti,m (i) | p nkm .

i=1

We now use (13.5.6.1) to isolate a single factor, yielding


n 2 km
|Ti,m (i) | p nkm
|T j,m ( j) |1
(i = 1, . . . , n).
p
j=i

(13.5.6.2)

Write Ti,m (i) = jZ Ti, j t j with Ti, j K ; we can then choose j = j (i, m)
so that |Ti, j (i) t j | = |Ti,m (i) | . We repeat the leverage process to limit the

size of j. Put = / > 1. In the case j 0, combine (13.5.6.2) with the

case = of (13.5.6.1) to get
 j
|Ti, j (i) t j |

2
j
=
p (n +n)km ;
=

|Ti, j (i) t j |
in the case j 0, combine (13.5.6.2) with the case = of (13.5.6.1) to get
j =

! " j

|Ti, j (i) t j |
2
p (n +n)km .
j
|Ti, j (i) t |

In either case we deduce that

| j| log p (n 2 + n)km.
Finally, note that
Diag(t B1 , . . . , t Bn )Tm Diag(t A1 , . . . , t An )
m

must have entries in K  p /t p , t p / p  by condition (a) of

Definition 13.5.2. Thus the integer j = j (i, m) is a representative of Bi
Am (i) modulo p m which is bounded in size by a constant times m. This implies
that A and B are weakly equivalent, as desired.

13.6 The p-adic Fuchs theorem for annuli

231

Corollary 13.5.7. Assume that < . Let M be a finite differential module

of rank n over K /t, t/ satisfying the Robba condition. Suppose that, for
some , with  <  , M K /t,t/ K   /t, t/   admits an
exponent A with p-adic non-Liouville differences. Then A is also an exponent
for M.
Proof. By Theorem 13.5.5 there exists an exponent B for M. By
Theorem 13.5.6, A and B are weakly equivalent and hence equivalent by
Lemma 13.4.3 since A has p-adic non-Liouville differences. Hence A is also
an exponent for M.
In general, it is quite difficult to compute the p-adic exponents of a differential module. However, one can at least check the following compatibility,
which will lead to an important instance in which the exponent of a differential
module can be controlled. See Corollary 13.6.2.
Lemma 13.5.8. Let M be a finite differential module of rank n over
K /t, t/ satisfying the Robba condition, and let : K /t, t/
K  1/q /t, t/ 1/q  be the substitution t  t q . If A is an exponent of M then
q A is an exponent of M.

13.6 The p-adic Fuchs theorem for annuli

We now come to the question of whether one can really invert the construction
of Remark 13.5.1, i.e., whether any differential module satisfying the Robba
condition is isomorphic to a differential module over a disc. We can hope to
treat this question only in the case where the module has an exponent with
p-adic non-Liouville differences: in this case the exponent is unique up to
equivalence, by Theorem 13.5.6, so there is no ambiguity about how to fill in
the hole in the annulus. The fact that no other conditions are necessary is the
content of the following remarkable theorem of Christol and Mebkhout.
Theorem 13.6.1 (ChristolMebkhout). Let M be a finite differential module
on an open annulus for the derivation td/dt satisfying the Robba condition
and admitting an exponent on some closed subannulus of positive width with
p-adic non-Liouville differences. Then M admits a basis on which the matrix
of action of D has entries in K and eigenvalues representing the exponent
of M (and hence belonging to Z p ). Consequently, M admits a canonical
decomposition
%
M=
M
Z p /Z

in which each M has exponent identically equal to .

232

p-adic exponents

The proof is loosely analogous to that of Theorem 13.5.6 but instead of

comparing two different sequences of matrices we compare one sequence with
itself.
Proof. Let , be the inner and outer radii of the original annulus. It suffices
to construct a basis of the desired form over the closed annulus  |t| 
for every pair  ,  with <  <  < . Choose  ,  with <  <
 <  <  < and  /  =  /  = > 1, so that M is represented by
a finite free module over K   /t, t/  .
By hypothesis, M admits an exponent A on some closed subannulus of positive width having p-adic non-Liouville differences; by replacing A with an
equivalent exponent we can force it also to have no nonzero integer differences.
By Corollary 13.5.7, A is also an exponent for M on any closed subannulus of positive width, including  |t|  . We may thus fix a basis of
M K   /t, t/   and then define a sequence Sm,A as in Definition 13.5.2.
1
Sm  ,A . As in (13.5.6.1) we have
For m  m, set Tm  ,m = Sm,A
nkm
|Tm  ,m | , |Tm1
 ,m | p

( [  ,  ]).

Our first goal is to show that Tm  ,m is close to its constant term. Write Tm  ,m =

h
hZ Tm  ,m,h t ; note that the (i, j)th entry of Tm  ,m,h can only be nonzero if
h Ai A j (mod p m ). Since A has p-adic non-Liouville differences, for
any c > 0 and for m sufficiently large the congruence h Ai A j (mod p m )
forces either h = Ai A j = 0 or |h| cm. If h > 0 we then have
!
"m
|Tm  ,m,h t h |  |Tm  ,m,h t h |  |Tm  ,m,h t h |  cm p nk c
and, if h < 0,

!
"m
|Tm  ,m,h t h |  |Tm  ,m,h t h |  |Tm  ,m,h t h |  cm p nk c .

Pick any (0, p nk ). Since > 1, we may choose c in the above argument
such that p nk c ; we then deduce that, for some m 0 ,
|Tm  ,m Tm  ,m,0 | m

( [  ,  ], m m 0 ).

This suggests renormalizing the Tm  ,m to make them convergent. Specifically,

we define Um  ,m = Tm  ,m Tm1
 ,m,0 , so that Um  ,m has constant term In . Then

 




  1 


In Tm  ,m,0 Tm1
 ,m  Tm  ,m Tm  ,m,0  Tm  ,m 

p nkm m < 1

( [  ,  ], m m 0 ),

and so |In Um  ,m | p nkm m < 1; in particular,

1
nkm
|Tm1
 ,m,0 | |Um  ,m | |Tm  ,m | p

( [  ,  ], m m 0 ).

13.6 The p-adic Fuchs theorem for annuli

233

Consider now three indices m  > m  > m. We have Tm  ,m Tm  ,m  = Tm  ,m , of

course; moreover, this multiplicativity is approximately preserved upon taking
constant terms. Namely, the identity

Tm  ,m,0 = Tm  ,m,0 Tm  ,m  ,0 +
Tm  ,m,h Tm  ,m  ,h
h=0

yields the bound

|Tm  ,m,0 Tm  ,m,0 Tm  ,m  ,0 | |Tm  ,m  Tm  ,m  ,0 | |Tm  ,m |


( [  ,  ], m m 0 ).

m p nkm
We now write

1
Um  ,m Um  ,m = Tm  ,m Tm1
 ,m,0 Tm  ,m Tm  ,m,0
1
= Tm  ,m Tm1
 ,m,0 (Tm  ,m,0 Tm  ,m  Tm  ,m,0 )Tm  ,m,0
1
= Tm  ,m Tm1
 ,m,0 (Tm  ,m,0 Tm  ,m  ,0 Tm  ,m,0 )Tm  ,m,0

+ Tm  ,m (Tm  ,m  Tm  ,m  ,0 )Tm1
 ,m,0 .
Using this last expression, we obtain the bound


|Um  ,m Um  ,m | m p 4nkm

( [  ,  ], m m 0 ).

We conclude that, for any fixed m m 0 , the sequence {Um  ,m }

m  =m is Cauchy.
It thus has limit U with the property that Sm,A U is the matrix that effects the
change to a basis of M K   /t, t/   of the desired form. This completes
the proof.
The hypothesis of Theorem 13.6.1 is difficult to verify in general, because
of the indirect nature of the definition of exponents. However, we do have the
following important case where the condition can be verified.
Corollary 13.6.2. Let M be a finite differential module on the open annulus
with inner radius and outer radius satisfying the Robba condition. Let q
be a power of p, and suppose that the intervals ( 1/q , 1/q ) and (, ) have
nonempty intersection. Suppose moreover that on some annulus there exists an
M M, where : K K is an isometry and is the
isomorphism K
=
K
q
substitution t  t . Then any exponent for M consists of rational numbers;
consequently, the conclusion of Theorem 13.6.1 applies.
Proof. Choose  ,  with  <  such that the intervals
((  )1/q , (  )1/q ) and (  ,  ) have nonempty intersection. On the intersection, M has exponent A by Theorem 13.5.5; A extends to be an exponent of
M on their entire domains of definition by Corollary 13.5.7.
both M and K

234

p-adic exponents

The rationality of A holds by Lemma 13.5.8 and Corollary 13.4.4. By

Theorem 13.5.6 and Lemma 13.4.3, any exponent for M is equivalent to A.
We may thus apply Theorem 13.6.1.
Remark 13.6.3. Corollary 13.6.2 is critical in what follows; it gives rise to
a quasiunipotence result (Lemma 21.2.1) which can be used to establish the
p-adic local monodromy theorem (Theorem 20.1.4).
One other consequence of Theorem 13.6.1 that can be stated without reference to exponents is the following. (In fact, it can also be proved without
exponents; see Remark 13.7.3.)
Corollary 13.6.4. Let M be a finite differential module on an open annulus for
the derivation td/dt satisfying the Robba condition. Suppose that the restriction M  of M to some smaller open annulus is either trivial or unipotent. Then
the same is true for M.
Proof. If M  is unipotent then it admits the exponent 0 by Remark 13.5.1.
However, any exponent A of M restricts to an exponent of M  and so is weakly
equivalent to 0 by Proposition 13.5.6. By Lemma 13.4.3, A is equivalent to 0
so Theorem 13.6.1 implies that M is unipotent. If now M  is trivial then M is
forced to be trivial also.

13.7 Transfer to a regular singularity

As an application of Theorem 13.6.1 we obtain a transfer theorem in the
presence of a regular singularity, in the spirit of Theorem 9.6.1 but with a
somewhat weaker estimate. For the necessity of weakening the estimate, see
Example 13.7.2 below.
Theorem 13.7.1. Let M be a finite differential module of rank n, on the open
unit disc for the derivation td/dt, with a regular singularity at t = 0 whose
exponents are in Z p and have p-adic non-Liouville differences. Then on the
open disc of radius lim1 R(M F )n , M admits a basis on which the
matrix of action of D has entries in Z p . In particular, if M is solvable at 1 then
this basis is defined over the whole open unit disc.
Proof. Using shearing transformations (as in Proposition 7.3.10), we may
reduce to the case where the exponents of M at the regular singularity
t = 0 are prepared. Note that, for any (0, 1), M K t/ admits a
basis. By Theorem 13.2.2 the fundamental solution matrix in terms of this
basis has entries defined over some disc of positive radius. From this and

13.7 Transfer to a regular singularity

235

Proposition 13.3.3, it follows that R(M F ) = for (0, 1) sufficiently

small.
Let be the supremum of (0, 1) for which R(M F ) = . Note that
the function f 1 (r ) = log R(M Fer ) is convex by Theorem 11.3.2(e), is
equal to r for r sufficiently large by the previous paragraph, and is also equal to
r for r = log by continuity (Theorem 11.3.2(a)). Consequently, f 1 (r ) = r
for all r log .
Choose , (0, ), with < , such that the fundamental solution matrix
of M (with respect to some basis on the closed disc of radius ) converges in
the open disc of radius . Let N0 be the matrix of action of D on the basis
B0 given by the columns of the fundamental solution matrix; by construction, N0 has entries in K whose eigenvalues are prepared and form a tuple
A Znp with p-adic non-Liouville differences. On the open annulus of inner
radius and outer radius , M admits the exponent A by Remark 13.5.1. By
Corollary 13.5.7, A is also an exponent for M on the open annulus of inner
radius and outer radius . By Theorem 13.6.1, on that annulus we obtain
another basis B1 of M on which D acts via a matrix N1 with entries in K and
eigenvalues equal to A.
Let U be the change-of-basis matrix from B0 to B1 ; it is an invertible n n
matrix over K {/t, t/} and satisfies the equation
d
(U ) = N1 ,
dt
or N0 U U N1 + t (d/dt)(U ) = 0. Since N0 , N1 have entries in K , if we write

U = i Ui t i then N0 Ui Ui N1 + iUi = 0 for each i Z. By Lemma 7.3.5

the map Ui  N0 Ui Ui N1 + iUi on n n matrices over K has eigenvalues
each of which is i plus a difference between two elements of A. In particular,
since A is prepared these eigenvalues can never vanish for i  = 0, so U = U0
has entries in K . (Compare the uniqueness argument in Proposition 7.3.6.) It
follows that B0 is in fact a basis of M on the open disc of radius , on which
D acts via a matrix over K . Since that matrix has eigenvalues in Z p we can
conjugate over K to put the matrix in Jordan normal form, in which case the
entries will belong to Z p .
It remains to give a lower bound for . By Theorem 11.3.2, for r
[0, log ] the function f 1 is continuous, convex, and piecewise affine, with
slopes belonging to
U 1 N0 U + U 1 t

1
1
Z Z.
1
n
Since the slope for r > log is equal to 1, the slopes for r log
cannot exceed 1. Moreover, there cannot be a slope equal to 1 in this range

236

p-adic exponents

as otherwise it would occur as the left-hand slope at r = log , so there

would exist > for which R(M F ) = , contrary to the definition
of . Consequently f 1 has all slopes less than or equal to (n 1)/n for r
[0, log ], yielding
log = f 1 ( log ) f 1 (0) +

n1
( log ).
n

From this we deduce that lim1 R(M F )n , as desired.

Example 13.7.2. The following example shows that the exponent n in
Theorem 13.7.1 cannot be replaced by 1. Pick K with || > 1. Let
M be the differential module of rank 2 over the open unit disc for the operator
td/dt, whose action of D on a basis is given by


0
.
t 0
Since this matrix modulo t is nilpotent, the exponents of M are both zero; in
particular, they are in Z p with p-adic non-Liouville differences.
For (||1 , 1) we may apply Theorem 6.5.3 to compute
1
|t D|sp,MF = ||1/2 1/2 and so R(M F ) = p 1/( p1) ||1/2 1/2 .
To compute R(M F ) for ||1 , we note that the function f 1 (M, r )
is affine of slope 1/2 for 0 r log || and of slope 1 for r large. By
Theorem 11.3.2(a), (b), (e), f 1 is piecewise affine and its slopes on [0, +)
must lie in [1/2, 1] (Z (1/2)Z). Consequently f 1 (M, r ) must have slope
1/2 until it reaches a point at which f 1 (M, r ) = r and slope 1 thereafter. In
other words,

p 1/( p1) ||1/2 1/2 > p 2/( p1) ||1 ,
R(M F ) =

p 2/( p1) ||1 .

In particular, the fundamental solution matrix of M in the given basis converges
on the open disc of radius p 2/( p1) ||1 = lim1 R(M F )2 but not on
any larger disc.
Remark 13.7.3. The final assertion of Theorem 13.7.1 (if M is solvable
at 1 then the fundamental solution matrix converges in the open unit disc)
can also be proved without relying on the p-adic Fuchs theorem for annuli
(Theorem 13.6.1); see the notes. This in turn can be used to give a proof of
Corollary 13.6.4 without using Theorem 13.6.1, as follows.
Let M be a finite differential module, on the open annulus < |t| < for
the derivation td/dt, satisfying the Robba condition. Suppose that the restriction M  of M to some smaller open annulus < |t| < is unipotent. Then, on

Notes

237

this smaller annulus, by Lemma 9.2.3 we obtain a basis of M on which td/dt

acts via a nilpotent matrix over K . This defines a differential module on the
disc |t| < with a nilpotent regular singularity at t = 0; by gluing the module
to M over the annulus < |t| < we obtain a differential module on the
disc |t| < with a nilpotent regular singularity at t = 0, which is solvable at
the boundary of the disc. By the final assertion of Theorem 13.7.1 this module
admits a basis on which td/dt acts via a nilpotent matrix over K . This gives us
such a basis of M over the annulus < |t| < ; by a similar argument, we get
a similar basis over the annulus < |t| < . These bases can only differ by
a K -linear transformation (as in the proof of Theorem 13.7.1), so each gives a
basis of M itself on which td/dt acts via a nilpotent matrix over K . Hence M
is unipotent.

Notes
The definition of a p-adic Liouville number was introduced by Clark [54]; our
presentation follows [80, VI.1].
The cited theorem of Clark [54, Theorem 3] is actually somewhat stronger
than Theorem 13.2.3 as it allows differential operators of possibly infinite
order.
The example in Remark 13.2.4 was loosely inspired by an example of
Monsky (a counterexample against a slightly different assertion); see [81, 7]
or [80, IV.8] for discussion.
Proposition 13.3.2 was originally due to Christol; compare [80,
Proposition IV.7.7].
The theory of exponents for differential modules on a p-adic annulus
satisfying the Robba condition was originally developed by Christol and
Mebkhout [50, 45]; in particular, Theorem 13.6.1 appears therein as [50,
Thorme 6.24]. A somewhat more streamlined development was later given
by Dwork [79], in which Theorem 13.6.1 appears as [79, Theorem 7.1].
(Dwork notes that he did not verify the equivalence between the two constructions; we do not recommend losing any sleep over this.) Our treatment is
essentially that of Dwork with a few technical simplifications. Another treatment will appears in the upcoming book of Christol [46]. Besides all these
there is an expository article [161] that summarizes the proofs (using Dworks
approach) and provides some further context, including the formulation of the
p-adic index theorem.
It is claimed in [161, 3.2] that the converse of Theorem 13.5.6 also holds,
i.e., if A is an exponent of a differential module M and B is weakly equivalent

238

p-adic exponents

to A then B is also an exponent of M. We do not know of a proof of this; in

[79, Remark 4.5] it is suggested that this may not be entirely trivial.
A somewhat more elementary treatment of Theorem 13.7.1 than the one
given here was given in [80, 6]; it does not rely on the p-adic Fuchs theorem
for annuli (Theorem 13.6.1). However, it gives a weaker result: it only establishes convergence of the fundamental solution matrix in the open disc of
2
radius lim1 R(M F )n . That weaker result is due to Christol, who
presented it himself in [42, Thorme 6.4.7] and [43].
The weaker result just mentioned is itself sufficient to imply that if
lim1 R(M F ) = 1 then the fundamental solution matrix converges
in the open unit disc. In the case of a nilpotent singularity one can show this
even more directly; see [134, Lemma 3.6.2]. As noted in Remark 13.7.3 this
can be used to derive Corollary 13.6.4 without using Theorem 13.6.1.
An intriguing archimedean analogue to the theory of p-adic exponents
appears in the local analytic theory of q-difference equations in the case
|q| = 1. This seems to extend to an analogy between q-difference equations
with |q|  = 1 and p-adic differential modules with intrinsic subsidiary radii
strictly less than 1. See [73] for part of this story.

Exercises
(1) Prove that rational numbers are p-adic non-Liouville numbers.
(2) Give another proof of Lemma 13.1.6 (as in [80, Lemma VI.1.2]) by first
verifying that both sides of the desired equation have the same coefficients of x 0 and x 1 and are killed by the second-order differential operator
"
d !d
x .
dx dx
(3) Show that Theorem 13.2.2 can be deduced from Theorem 13.2.3. (Hint:
show that if H 0 (M)  = 0 then 0 must occur as an eigenvalue of N0 .)
(4) Prove that there exists a Z p satisfying
type(a) = 1,

type(a) < 1.

(5) Prove Proposition 13.3.3.

(6) Verify Example 13.4.6.
p
(7) (a) Prove that the set of A Zn which are weakly equivalent to 0 is a
p
subgroup of Zn .
(b) Give a counterexample against the claim that if Ai is weakly equivalent to Bi for i = 1, 2 then A1 + A2 is weakly equivalent to B1 + B2 .
(Hint: give an example using rational numbers, in which case weak
equivalence implies equivalence, by an earlier exercise.)

Exercises

239

(8) Complete the proof of Proposition 13.5.3.

(9) Prove that if a differential module M over K /t, t/ admits a basis
on which td/dt acts via a nilpotent matrix over K then the K -span of
the basis is fixed by the Taylor series action of p , but not necessarily
by the full action of 1 + m K . (Hint: this can be reduced to the fact that
p is in the kernel of the logarithm map defined using the power series
log(1 z), which converges for |z| < 1.)
(10) Let M be a differential module on an open annulus satisfying the Robba
condition. Use Theorem 13.6.1 to prove that H 0 (M) and H 1 (M) are
both finite-dimensional and of the same dimension, i.e., that the index
of M is 0.

Part IV
Difference Algebra and Frobenius Modules

14
Formalism of difference algebra

We now step away from differential modules for a little while to study the
related subject of difference algebra. This is the theory of algebraic structures
enriched not with a derivation but with an endomorphism of rings. Our treatment of difference algebra will run largely in parallel with that for differential
algebra but in a somewhat abridged fashion; our goal is to be able to use difference algebra to say nontrivial things about p-adic differential equations. We
will begin to do that in Part V.
In this chapter we introduce the formalism of difference rings, fields, and
modules and the associated notion of twisted polynomials. Then we study
briefly the analogue of algebraic closure for a difference field. Finally, we
make a detailed study of difference modules over a complete nonarchimedean
field culminating in a classification of difference modules for the Frobenius
automorphism of a complete unramified p-adic field with algebraically closed
residue field (the DieudonnManin classification).
Hypothesis 14.0.1. Throughout this part of the book and the next, we retain
Notation 8.0.1 with p > 0 (so that K is a complete nonarchimedean field of
characteristic 0 and residual characteristic p > 0) but, unless otherwise specified, we will require K to be discretely valued. This is necessary to avoid a
number of technical complications, some of which we will point out as we go
along. (We will make almost no reference to K in this particular chapter; we
only include this hypothesis here in order to place it at the beginning of Part V.)

14.1 Difference algebra

We start with the central definition. See the notes at the end of the chapter for
some explanation of the term difference in this context.
243

244

Formalism of difference algebra

Definition 14.1.1. A difference ring or field is a ring or field R equipped

with an endomorphism . A difference module over R is an R-module M
equipped with a map  : M M which is additive and -semilinear; the
latter means that
(r m) = (r )(m)

(r R, m M).

This data determines an R-linear map (M) = M R, R M sending

m 1 to (m); conversely, any R-linear map M M induces the structure
of a difference module on M. A free difference module is trivial if it admits a
-invariant basis; when we refer to the trivial module we mean the module
R itself, with  = . A difference submodule of R is also called a difference
ideal.
Morphisms of difference rings or modules are required to be -equivariant.
Tensor products behave as expected: if M, N are difference modules over the
same difference ring R then M R N acquires the structure of a difference
module with
(m n) = (m) (n)

(m M, n N ).

In particular, if R S is a morphism of difference rings then we can perform

a base change from difference modules over R to difference modules over S.
As in differential algebra, difference modules can often be described in
terms of matrices of action.
Definition 14.1.2. If M is a finite difference module over R freely generated
by e1 , . . . , en then we can recover the action of  from the n n matrix A
defined by

(e j ) =
Ai j ei .
i

Namely, if we use the basis to identify M with the space of column vectors of
length n over R then
(v) = A(v).
In parallel with the differential case, we call A the matrix of action of  on the
basis e1 , . . . , en .
We now record the effect of a change of basis on a matrix of action.
Remark 14.1.3. Let M be a finite free difference module over R, and let
e1 , . . . , en and e1 , . . . , en be two bases of M. Recall that the change-of-basis
matrix U from the first basis to the second is the n n matrix U satisfying

14.1 Difference algebra

245

ej = i Ui j ei . If A is the matrix of action of  on the ei then the matrix of

action of  on the ei is
A = U 1 A(U ).
It is important to consider dual modules, but the description in difference
algebra is somewhat more complicated than in differential algebra. We will
introduce it only in the free case. (One can extend the definition to locally free
modules; compare Remark 5.3.4.)
Definition 14.1.4. Let M be a finite free difference module over R. We say
that M is dualizable if the map (M) M induced by  is an isomorphism.
Given a basis of M, it is equivalent to check that the matrix of action A of  on
that basis is invertible (the same then holds for any other basis). However, even
in this case the action of  on M itself is not invertible unless the action of
on R is invertible (i.e., unless R is inversive in the sense of Definition 14.1.5
below).
If M is dualizable, there is a unique way to equip the module-theoretic dual

M = Hom R (M, R) with a -action such that

( f )((m)) = f (m)

( f M , m M).

In terms of a given basis, the matrix of action on the dual basis is given by the
inverse transpose AT . We call the resulting difference module the dual of M.
There is also a notion of an opposite difference ring but only under an extra
hypothesis.
Definition 14.1.5. We say that the difference ring R is inversive if is an automorphism. In this case, we can define the opposite difference ring R opp to be R
again, but now equipped with the endomorphism 1 . If R is inversive and M
is a finite free difference module over R, we define the opposite module M opp
of M to be M equipped with the inverse of ; that is, if A is the matrix of action
of  on some basis then the matrix of action of 1 is given by 1 (A1 ).
Definition 14.1.6. For M a difference module, write
H 0 (M) = ker(id ),

H 1 (M) = coker(id ).

If M1 , M2 are difference modules with M1 dualizable then H 0 (M1 M2 )

computes -equivariant morphisms from M1 to M2 and H 1 (M1 M2 ) computes extensions 0 M2 M M1 0 of difference modules (exercise).
That is, we have natural identifications
H 0 (M1 M2 ) = Hom(M1 , M2 ),

H 1 (M1 M2 ) = Ext(M1 , M2 )

of groups (see the proof of Lemma 5.3.3 for the group structure on extensions).

246

Formalism of difference algebra

14.2 Twisted polynomials

As in differential algebra, there is a relevant notion of twisted polynomials.
Definition 14.2.1. For R a difference ring, we define the twisted polynomial

ring R{T } as the set of finite formal sums i=0

ri T i , with multiplication obeying the rule T r = (r )T . For any P R{T }, the quotient R{T }/R{T }P is a
difference module; if M is a difference module, we say that m M is a cyclic
vector if there is an isomorphism M
= R{T }/R{T }P carrying m to 1.
Definition 14.2.2. If R is inversive, we again have a formal adjoint construction: given P R{T }, its formal adjoint is obtained by taking the coefficients
to the right-hand side of T . This may then be viewed as an element of the
opposite ring of R{T }, which we may identify with R opp {T }.
It is not completely straightforward to analogize the cyclic vector theorem
to difference modules; see the exercises for one attempt to do so. Instead, we
will use only the following trivial observation.
Lemma 14.2.3. Any irreducible finite difference module over a difference field
contains a cyclic vector.
Proof. If F is a difference field, M is a finite difference module over F, and
m M is nonzero then m, (m), . . . generate a nonzero difference submodule
of M. If M is irreducible, this submodule must be the whole of M.
Definition 14.2.4. If is isometric for a norm | | on F then we have the
usual definition of Newton polygons and slopes for twisted polynomials, with
a natural analogue of the additivity of slopes (i.e., Proposition 2.1.2). If R
is inversive then a twisted polynomial and its adjoint have the same Newton
polygon.
As in the differential case, we may apply the master factorization theorem
(Theorem 2.2.2), as in the proof of Theorem 2.2.1, to get a factorization result.
However, the result is inherently asymmetric; see Remark 14.2.6 below.
Theorem 14.2.5. Let F be a difference field complete for a norm | | under
which is isometric. Then any monic twisted polynomial P F{T } admits a
unique factorization
P = Pr1 Prm
for some r1 < < rm , where each Pri is monic with all slopes equal to ri .
(The same holds with the factors in the opposite order if F is inversive, but not
always otherwise; see Remark 14.2.6 and the exercises.)
Remark 14.2.6. It is worth clarifying why the conditions of Theorem 2.2.2 can
be satisfied by F{T } but not necessarily by its opposite ring. The key condition
in this theorem is (b), which states that with

14.3 Difference-closed fields

247

U = {P F[T ] : deg(P) deg(S) m},

V = {P F[T ] : deg(P) m 1},
W = {P F[T ] : deg(P) deg(S)},
a = 1,
b = T m,
the map (u, v)  av + ub surjects U V onto W . This holds because, on the
one hand, any element of F{T } whose coefficients of 1, T, . . . , T m1 vanish
/ (F) then r T m is not
is divisible by T m on the right. On the other hand, if r
divisible on the left by T , let alone by T m .

14.3 Difference-closed fields

Now we study briefly the difference-theoretic analogue of the algebraic closure
property of ordinary fields.
Definition 14.3.1. We will say that a difference field F is weakly differenceclosed if every dualizable finite difference module over F is trivial. We say F
is strongly difference-closed if F is inversive and weakly difference-closed.
Remark 14.3.2. Note that the property that F is weakly difference-closed
includes the fact that short exact sequences of dualizable finite difference
modules over F always split. By contrast, if for instance is the identity
map then this is never true even if F is algebraically closed, because linear
transformations need not be semisimple.
Here is a useful criterion for checking for difference closure.
Lemma 14.3.3. The difference field F is weakly difference-closed if and only
if the following conditions all hold.
(a) Every nonconstant monic twisted polynomial P F{T } factors as a
product of linear factors.
(b) For every c F , there exists x F with (x) = cx.
(c) For every c F , there exists x F with (x) x = c.
Proof. We first suppose that F is weakly difference-closed. To prove (a), it
suffices to check that if P F{T } is nonconstant monic with nonzero constant
term then P factors as P1 P2 with P2 linear. The nonzero constant term implies
that M = F{T }/F{T }P is a dualizable finite difference module over F and
so must be trivial by the hypothesis that F must be weakly difference-closed.
In particular, there exists a short exact sequence 0 M1 M M2 0
with M2 trivial of dimension 1; this corresponds to a factorization P = P1 P2
with P2 linear.

248

Formalism of difference algebra

To prove (b), note that F{T }/F{T }(T c1 ) must be trivial. Since each
element of this difference module is represented by some x F, we must have
x F such that T x x is divisible on the right by T c1 . By comparing
degrees, we see that T x x = y(T c1 ) for some y F. Then y = (x)
and yc1 = x, proving the claim.


1 c
To prove (c), form the -module V corresponding to the matrix
. By
0 1
construction, we have a short exact sequence 0 V1 V V2 0 with
V1 , V2 trivial; since V must also be trivial, this extension must split. In other
words we can find x F such that





1 x
1 c
1 (x)
1 0
=
,
0 1
0 1
0
1
0 1
that is, (x) x = c.
Conversely, suppose that (a)(c) hold. Every nonzero dualizable finite difference module over F admits an irreducible quotient. This quotient admits a
cyclic vector, by Lemma 14.2.3, and so admits a quotient of dimension 1 by
(a). That quotient in turn is trivial by (b). By induction, we deduce that every
dualizable finite difference module over F admits a filtration whose successive
quotients are trivial of dimension 1. This filtration splits, by (c).
The notion of difference closure is particularly simple for the absolute
Frobenius endomorphism (the p-power map) on a field of characteristic p > 0.
Proposition 14.3.4. Let F be a separably (resp. algebraically) closed field
of characteristic p > 0 equipped with a power of the absolute Frobenius
endomorphism. Then F is weakly (resp. strongly) difference-closed.

m
Pi T m F{T }, with m > 0, Pm = 1, and P0  = 0, the
Proof. For P = i=0

m
i
polynomial Q(x) = i=0 Pi x q has degree q m 2, and x = 0 occurs as a
root only with multiplicity 1. Moreover, the formal derivative of P is a constant
polynomial and so has no common roots with P; hence P is a separable polynomial. Since F is separably closed, there must exist a nonzero root x of Q;
this implies criteria (a) and (b) of Lemma 14.3.3. To deduce (c) note that, for
c F , the polynomial x q x c is again separable and so has a root in F.

14.4 Difference algebra over a complete field

Hypothesis 14.4.1. For the rest of this chapter, let F be a difference field complete for a nonarchimedean norm | |, with respect to which is isometric. (For
short, we will say that F is an isometric complete nonarchimedean difference

14.4 Difference algebra over a complete field

249

field.) We do not assume that F is inversive; if it is not then we can embed F

into an inversive difference field by forming the completion F  of the direct
limit of the system

F F .
We sometimes call F  the perfection, or more precisely the -perfection, of F.
As in the differential case, we would like to classify finite difference modules over F by the spectral radius of ; this turns out to be somewhat easier in
the difference case because there is now no analogue of the distinction between
the visible and full spectra. A related fact is that the use of matrix inequalities
here is much closer to that in Chapter 4 than that in Chapter 6; the main difference is that we do not begin with a satisfactory theory of eigenvalues or
eigenvectors.
We first note the following analogue of Lemma 6.2.5.
Lemma 14.4.2. Let V be a nonzero finite difference module over F. Let
e1 , . . . , en be a basis of V . For each nonnegative integer s, let As be the matrix
of action of s on e1 , . . . , en . Then
||sp,V = lim |As |1/s .
s

Proof. It suffices to observe that, using the supremum norm defined by

e1 , . . . , en , the operator norm of s is precisely | As |. Thus the limit on the
right-hand side exists and matches the definition of the left-hand side.
The following basic properties will help with our classification, as long as
we are mindful of the discrepancies between the differential and difference
cases. For instance, the difference case has a simpler rule for tensor products but a less simple rule for duals. (It may be helpful to keep in mind the
case where is the identity map, and relate the following observations to
eigenvalues of linear transformations.)
Lemma 14.4.3. Let V, V1 , V2 be nonzero finite difference modules over F.
(a) For a short exact sequence 0 V1 V V2 0,
||sp,V = max{||sp,V1 , ||sp,V2 }.
(b) We have
||sp,V1 V2 = ||sp,V1 ||sp,V2 .
(c) For any complete extension E of F to which extends isometrically,
||sp,V = ||sp,V E .

250

Formalism of difference algebra

(d) For V dualizable,

||sp,V ||sp,V 1.
Proof. Exercise.
Lemma 14.4.4. If V
= F{T }/F{T }P and P has only one slope r < + in
its Newton polygon, then V admits a basis on which
||V = er ,

|1 |V F  = er .

Consequently
||sp,V = er ,

| |sp,V = er

and, if F is inversive, then also

|1 |sp,V = er .
Proof. Put n = deg(P), and define a norm on V by
|a0 + + an1 T n1 | = max{|ai |eri };
i

then clearly
||V = er ,

|1 |V F  = ||V = er .

We deduce that
||sp,V er ,

|1 |sp,V F  , ||V er ;

since
1 ||sp,V |1 |sp,V F  er er ,

1 ||sp,V ||sp,V er er ,

we obtain the desired equalities.

Corollary 14.4.5. For any nonzero finite difference module V over F, either
||sp,V = 0 or there exists an integer m {1, . . . , dim F V } such that

||m
sp,V |F |.
Definition 14.4.6. Let V be a nonzero finite difference module over F. We say
that V is pure (and) of norm s if all the Jordan-Hlder constituents of V have
spectral radius s. Note that V is pure of norm 0 if and only if dim F V = 0. If
V is pure of norm 1, we also say that V is tale or unit-root.
Remark 14.4.7. It is more common to classify pure modules using additive
notation, i.e., in terms of the slope ( log s) instead of the norm s. (Correspondingly, pure modules are also called isoclinic modules.) For better or

14.4 Difference algebra over a complete field

251

worse, we have decided here to keep the notation consistent with the multiplicative terminology used in the differential case. We will switch to the
additive language only in the next section, in order to talk about Hodge and
Newton polygons.
Proposition 14.4.8. Let V be a nonzero finite difference module over F. Then
V is pure of norm s > 0 if and only if
||sp,V F  = s,

|1 |sp,V F  = s 1 ;

(14.4.8.1)

in this case, V is pure of norm s 1 .

Proof. If V is pure of norm s then (14.4.8.1) holds by Lemmas 14.4.3(a), (c)
and 14.4.4. Conversely, if (14.4.8.1) holds and W is an irreducible subquotient
of V then, by Lemma 14.4.3(a),
||sp,W F  ||sp,V F  ,

|1 |sp,W F  |1 |sp,V F  .

We thus have
1 ||sp,W F  |1 |sp,W F  ss 1 = 1,
which forces the equalities ||sp,W = ||sp,W F  = s and |1 |sp,W F  =
s 1 . In particular, W is pure of norm s. Moreover, by Lemma 14.2.3
there is an isomorphism W
= F{T }/F{T }P for some twisted polynomial
P; by Theorem 14.2.5, P has only one slope in its Newton polygon. By
Lemma 14.4.4 that slope must equal log s, and so W must be pure of norm
s 1 , as then is V .
Corollary 14.4.9. Let V1 , V2 be nonzero finite difference modules over F
which are pure and of respective norms s1 , s2 . Then V1 V2 is pure of
norm s1 s2 .
Note that this does not follow immediately from Lemma 14.4.3(b) because
the tensor product of two irreducibles need not be irreducible.
Proof. If s1 s2 = 0 then it is easy to check that V1 V2 is pure of norm 0.
Otherwise, Lemma 14.4.3(b) plus Proposition 14.4.8 yields
||sp,V1 V2 F  = ||sp,V1 F  ||sp,V2 F  = s1 s2 ,
|1 |sp,V1 V2 F  = |1 |sp,V1 F  |1 |sp,V2 F  = s11 s21 ;
thus Proposition 14.4.8 again implies that V1 V2 is pure of norm s1 s2 .

252

Formalism of difference algebra

Corollary 14.4.10. Let V be a nonzero finite difference module over F. Then,

for any positive integer d, V is pure of norm s if and only if V becomes pure
of norm s d when viewed as a difference module over (F, d ).
Proposition 14.4.11. Let V be a nonzero finite difference module over F.
Suppose that either
(a) ||sp,V < 1 or
(b) F is inversive and |1 |sp,V < 1.
Then H 1 (V ) = 0.
Proof. In case (a), given v V , the series
w=

i (v)

i=0

converges to a solution of w (w) = v. In case (b), the series

w=

i1 (v)

i=0

does likewise.
Corollary 14.4.12. If V1 , V2 are nonzero finite differential modules over F
which are pure of respective norms s1 , s2 and if either
(a) s1 < s2 or
(b) F is inversive and s1 > s2 ,
then any exact sequence 0 V1 V V2 0 splits.
Proof. If s2 > 0 then, by Proposition 14.4.8 and Corollary 14.4.9, V2 V1 is
pure of norm s1 /s2 , so Proposition 14.4.11 gives the desired splitting. Otherwise we must be in case (b), so we can pass to the opposite ring to make the
same conclusion.
If F is inversive, we again get a decomposition theorem.
Theorem 14.4.13. Suppose that F is inversive. Let V be a finite difference
module over F. Then there exists a unique direct sum decomposition
%
V =
Vs
s0

of difference modules, in which each nonzero Vs is pure of norm s. (Note that

V is dualizable if and only if V0 = 0.)
Proof. This follows at once from Corollary 14.4.12.

14.4 Difference algebra over a complete field

253

Remark 14.4.14. Note that if is the identity map on F, Theorem 14.4.13

simply reproduces the decomposition of V in which Vs consists of the
generalized eigenspaces for all eigenvalues of norm s.
If F is not inversive, we get a filtration instead of a decomposition.
Theorem 14.4.15. Let V be a finite difference module over F. Then there
exists a unique filtration
0 = V0 V1 Vl = V
of difference modules, such that each successive quotient Vi /Vi1 is pure of
some norm si and s1 > > sl . (Note that V is dualizable if and only if
V = 0 or sl > 0.)
Proof. Start with any filtration 0 = V0 V1 Vl = V with irreducible
successive quotients. Let si be the radius of Vi /Vi1 . In the case si < si+1 we
may apply Corollary 14.4.12 to choose another difference module Vi , containing Vi1 and contained in Vi+1 , such that Vi /Vi1 , Vi+1 /Vi are pure and have
slopes si+1 , si , respectively.
By repeating this process, we eventually reach the case where s1
sl . We obtain the existence of a suitable filtration from the given filtration by
simply omitting Vi whenever si = si1 . Uniqueness follows by tensoring with
F  and invoking the uniqueness in Theorem 14.4.13.
The following alternative characterization of pureness may be useful in
some situations.
Proposition 14.4.16. Let V be a finite difference module over F, and choose
F . Then V is pure of norm || if and only if there exists a basis of V on
which  acts via times an element of GLn (o F ).
Proof. If such a basis exists then Proposition 14.4.8 implies that V is pure
of norm ||. Conversely, if V is irreducible of spectral radius || then
Lemma 14.4.4 provides a basis of the desired form. Otherwise we proceed
by induction on dim F V . Suppose that we are given a short exact sequence
0 V1 V V2 0 in which V1 , V2 admit bases of the desired form.
Let e1 , . . . , em V form such a basis for V1 , and let em+1 , . . . , en V lift
such a basis for V2 . Then, for F of sufficiently small norm,
e1 , . . . , em , em+1 , . . . , en
will form a basis of V of the desired form.

254

Formalism of difference algebra

Remark 14.4.17. Note that, whenever V is pure and of positive norm, we can
apply Proposition 14.4.16 after replacing  by some power of itself, thanks to
Corollary 14.4.5.

14.5 Hodge and Newton polygons

The theory of Hodge and Newton polygons, which we introduced when studying nonarchimedean matrix inequalities in Chapter 4, admits a close analogue
when considering a difference algebra over a complete nonarchimedean field.
Throughout this section, we continue to retain Hypothesis 14.4.1.
Definition 14.5.1. Let V be a finite difference module over F equipped with
the supremum norm with respect to some basis. Let A be the matrix of action
of  on this basis; define the Hodge polygon of V as the Hodge polygon of
the matrix A (see Definition 4.3.3). Given the choice of the norm on V , this
definition is independent of the choice of the basis: we can change basis only
by using a matrix U GLn (o F ) which replaces A by U 1 A(U ); since is
an isometry, this ensures that (U ) GLn (o F ) also. As in the linear case we
list the Hodge slopes s H,i , . . . , s H,n in increasing order.
Definition 14.5.2. Let V be a finite difference module over F. Define the
Newton polygon of V to have slopes s N ,1 , . . . , s N ,n such that r appears
with multiplicity equal to the dimension of the quotient of the norm er in
Theorem 14.4.15.
Lemma 14.5.3. Let V be a finite difference module over F. We have
s H,1 + + s H,i = log ||i V
s N ,1 + + s N ,i = log ||sp,i V

(i = 1, . . . , n),
(i = 1, . . . , n).

Proof. The first assertion follows from the corresponding fact in the linear
case (which is analogous to Lemma 4.1.9). The second assertion reduces to
the fact that if V is irreducible of dimension n and spectral radius s then i V
has spectral radius s i for i = 1, . . ., n; this follows from the fact that, in the
basis given by Lemma 14.4.4, i i V = s i .
Corollary 14.5.4 (Newton above Hodge). We have
s N ,1 + + s N ,i s H,1 + + s H,i

(i = 1, . . . , n)

with equality for i = n.

As in the linear case (Theorem 4.3.11), we have a HodgeNewton decomposition theorem.

14.5 Hodge and Newton polygons

255

Theorem 14.5.5. Let V be a finite difference module over F equipped with a

basis. If for some i {1, . . . , n 1} we have
s N ,i < s N ,i+1 ,

s N ,1 + + s N ,i = s H,1 + + s H,i

then we can change basis, using a matrix in GLn (o F ), so that the matrix of
action of  becomes block upper triangular; the top left block accounts for
the first i Hodge and Newton slopes of V . Moreover, if F is inversive and
s H,i > s H,i+1 , we can ensure that the matrix of action of  becomes block
diagonal.
Proof. First we change basis by a matrix in GLn (o F ) such that the F-span
of the first i basis vectors equals the step of the filtration of Theorem 14.4.15
consisting of norms greater than or equal to es N ,i . We may then proceed as in
Theorem 4.3.11.
Remark 14.5.6. Be aware that the Newton polygon, unlike the Hodge polygon, cannot be read off directly from the matrix of action of ; see the
exercises for an example. However, this can be done if the matrix of  is a
companion matrix. A restatement follows.
Proposition 14.5.7. If V
= F{T }/F{T }P then the Newton polygon of V
coincides with that of P.
Proof. This reduces to Lemma 14.4.4 using Theorem 14.2.5.
We also obtain the following analogue of Proposition 4.4.10.
Proposition 14.5.8. Let V be a finite difference module over F, equipped with
the supremum norm for some basis. For i = 1, 2, . . . , let s H,1,i , . . . , s H,n,i be
the Hodge slopes of V as a difference module over (F, i ). Then, as i ,
the quantities i 1 s H,1,i , . . . , i 1 s H,n,i converge to the Newton slopes of V .
Proof. The claim is independent of the choice of basis, so by Theorem 14.4.15
we may reduce to considering a pure module. In that case Lemma 14.4.2
implies that i 1 s H,1,i s N ,1 as i .
Note that s N ,1 = = s N ,n by purity and that i 1 s H,1,i i 1 s H, j,i for
j 1. Also, the Newton slopes of V as a difference module over (F, i ) are
is N ,1 , . . . , is N ,n ; by the equality case of Corollary 14.5.4 we obtain i 1 s H,1,i +
+ i 1 s H,n,i = s N ,1 + + s N ,n . All this plus the fact that i 1 s H,1,i s N ,1
as i yields the desired convergence.
Proposition 14.5.9. Suppose that F = K (i.e., F is discretely valued) and that
R = Ft0 carries the structure of an isometric complete nonarchimedean
difference ring for the norm | |1 . Let M be a finite free difference module over

256

Formalism of difference algebra

R with least Newton slope c. Then there exists a basis of M with respect to
which, for each positive integer m, the least Hodge slope of m is the least
element of v(F ) greater than or equal to cm.
Proof. Recall that if K is discrete then E is the completed fraction field of R
(see Definition 9.4.3). Construct a suitable basis of M E by imitating the
proof of Lemma 4.3.13 and then apply Lemma 8.6.1.

14.6 The DieudonnManin classification theorem

We also have an analogue in difference algebra of the fact that linear transformations over an algebraically closed field can be put into Jordan normal form.
In fact, the situation is even better: in this setting all objects are semisimple.
We continue to retain Hypothesis 14.4.1.
We first define some standard difference modules of a particularly simple
form.
Definition 14.6.1. For F and d a positive integer, let V,d be the difference
module over F with basis e1 , . . . , ed such that
(e1 ) = e2 ,

...,

(ed1 ) = ed ,

(ed ) = e1 .

Lemma 14.6.2. Suppose that

and the positive integer d are such that
there is no i {1, . . . , d 1} such that ||i/d |F |. Then V,d is irreducible.
Proof. Note that
d (ei ) = i1 ()ei

(i = 1, . . . , d).

Hence, by Proposition 14.4.16, V,d is pure of norm 1/d , as then is any submodule. But if the submodule were proper and nonzero then we would have a
violation of Corollary 14.4.5.
Next we show that, if F has a sufficiently large residue field, one can classify
all dualizable finite difference modules in terms of the standard modules V,d .
One can always enlarge F to reach this case; see the exercises.
Theorem 14.6.3. Let F be a complete discretely valued field, equipped with
an isometric endomorphism , such that F is strongly difference-closed. Then
every dualizable finite difference module over F can be split (non-uniquely) as
a direct sum of submodules, each of the form V,d for some , d. Moreover,
given any generator of m F , we can force each to be a power of .
Proof. We first check that if V is pure of norm 1 then V is trivial; for this
step, we only need F to be weakly difference-closed. We must show that, for

14.6 The DieudonnManin classification theorem

257

any A GLn (o F ), there exists a convergent sequence U1 , U2 , GLn (o F )

such that
Um1 A(Um ) In

(mod m ).

Specifically, we will insist that Um+1 Um (mod m ). Finding U1 amounts

to trivializing a dualizable difference module of dimension n over F , which
is possible because F is assumed to be weakly difference-closed. For m > 1,
given Um we wish to set Um+1 = Um (In + m X m ), for some X m , in such a
way that
(In + m X m )1 (Um1 A(Um ))(In + ()m (X m )) In

(mod m+1 ).

Since we already have Um1 A(Um ) In (mod m ), this amounts to solving

X m + m (Um1 A(Um ) In ) + ()m m (X m ) 0

(mod ),

which we can do by applying criterion (c) from Lemma 14.3.3.

Next we check that is surjective on o F , which implies that F is inversive.
Given x o F , it suffices to exhibit a sequence y0 , y1 , y2 , . . . , with ym ym+1
(mod m ), such that (ym ) x (mod m ). We start with y0 = 0; given ym ,
solving the equation
( m (ym+1 ym )) ()m (x (ym ))

(mod )

is possible because F is inversive.

We now check that if V is trivial then H 1 (V ) = 0. Given x o F , it suffices
to exhibit a sequence y0 , y1 , y2 , . . . , with ym ym+1 (mod m ), such that
(ym ) ym x (mod m ). Again, we start with y0 = 0; given ym , solving
the equation
m ( m )( m (ym+1 ym )) m (ym+1 ym ) m (x (ym ) + ym ) (mod )

is possibly by criteria (b) and (c) from Lemma 14.3.3. (We must use (b) to
remove the leading coefficient in the first term before applying (c).)
At this point, we may apply Theorem 14.4.13 to reduce the desired result to
the case where V is pure of norm s > 0. Let d be the smallest positive integer
such that s d = | m | for some integer m. Then the first paragraph of the proof
implies that m d fixes some nonzero element of V ; this gives us a nonzero
map from V m ,d to V . By Lemma 14.6.2 this map must be injective. Repeating
this argument, we can write V as a successive extension of copies of V m ,d .
However, Vm ,d V m ,d is pure of norm 1 and so has trivial H 1 as above.
Thus V splits as a direct sum of copies of V m ,d , as desired.
By Proposition 14.3.4, Theorem 14.6.3 has the following immediate
corollary.

258

Formalism of difference algebra

Corollary 14.6.4. Let F be a complete discretely valued field such that F

is algebraically closed and of characteristic p > 0. Let : F F be an
isometric automorphism lifting a power of the absolute Frobenius on F . Then
every dualizable finite difference module over F can be split (non-uniquely) as
a direct sum of difference submodules, each of the form V,d for some F
and some positive integer d coprime to the valuation of . Moreover, given any
generator of m F , we can force each to be a power of .
Remark 14.6.5. Let k be an algebraically closed field of characteristic p >
0, let W (k) be the ring of p-typical Witt vectors (i.e., the unique complete
discrete valuation ring with residue field k and maximal ideal generated by
p), and put F = Frac(W (k)). Then, for each power q of p, there is a unique
Frobenius lift on W (k), namely the Witt vector Frobenius lift. For this data,
Corollary 14.6.4 with = p is precisely the DieudonnManin theorem, i.e.,
the classification theorem of rational Dieudonn modules over k. (For more on
Witt vectors, see the notes and the exercises.)
Corollary 14.6.6. Let F be a complete discretely valued field, equipped with
an isometric endomorphism , such that F is strongly difference-closed. Then,
for any finite difference module V over F, H 1 (V ) = 0.
Proof. By Theorem 14.4.13 this reduces to the case where V is pure and of
some norm s 0. If s > 0 then V is dualizable and Theorem 14.6.3 implies
the claim. If s = 0 then the action of  on V is nilpotent, so we may check the
claim for V by checking the kernel and image of V . We may thus reduce to the
case where the action of  on V is zero, when F becomes inversive; this was
also checked in the proof of Theorem 14.6.3.

Notes
The parallels between difference algebra and differential algebra are close
enough that a survey of references for difference algebra strongly resembles
its differential counterpart. A well-established, perhaps rather dry, reference
is [56]; a somewhat more lively and more modern reference, which develops difference Galois theory under somewhat restrictive conditions, is [201].
We again mention [4] as a useful unifying framework for difference and
differential algebra.
The choice of the modifier difference in the phrase difference algebra is
motivated by the following basic example. Consider the automorphism on the
ring of entire functions f : C C given by ( f )(z) = f (z + q) for some q.
The difference operator

Notes
( f ) =

259

( f ) f
q

obeys the twisted Leibniz rule

( f g) = ( f )(g) + ( f )(g),
and its limit as q 1 is the usual derivation with respect to z.
Proposition 14.3.4 can be found in SGA7 [70, Expos XXII, Corollaire 1.1.10], wherein Katz attributes it to Lang. Indeed, it is a special case
of the nonabelian ArtinSchreier theory associated with an algebraic group
over a field of positive characteristic (in our case GLn ) via the Lang torsor;
see [153]. It is also the basis for the theory of (, )-modules, which we will
mention briefly in Chapter 24. (See also Section 19.1.)
Suppose that k is a perfect field of characteristic p > 0 and that Ck is a complete discretely valued field with residue field k and maximal ideal generated
by p (e.g., k = F p and Ck = Z p ). It was originally noticed by Teichmller
that each element x k has a unique lift [x] Ck admitting p n th roots for
all n; this is called the Teichmller lift of x. (These were first considered for
unramified extensions of Z p , in which context we have already seen them, in
Chapter 0.) One can use Teichmller lifts as digits to form a canonical base- p
expansion of any element of Ck ; the arithmetic operations can be expressed
using certain polynomials in the p-power roots of these digits. These polynomials were used by Witt to prove that Ck exists, is unique, and is functorial
in k. More generally, Witt gave a functor for each p accepting arbitrary rings
(the functor of p-typical Witt vectors) and associating k as above to Ck ; he also
gave a natural functor through which all the functors, for different choices of p,
factorize (the functor of big Witt vectors). These have applications far beyond
their origins, in fields as diverse as arithmetic geometry, algebraic topology,
and combinatorics. See [110] for a comprehensive summary.
In the special case of the difference field Frac(W (k)), with k perfect and
of characteristic p > 0 and equal to the unique lift of the absolute ppower Frobenius morphism, a number of the results in this chapter appear (in
marginally less generality) in [120], such as the following.
Corollary 14.5.4 reproduces Mazurs [120, Theorem 1.4.1].
Theorem 14.5.5 is [120, Theorem 1.6.1].
Proposition 14.5.8 reproduces [120, Corollary 1.4.4] without requiring
nonnegative Hodge slopes (as Katz does in his basic slope estimate
[120, 1.4.3]).
Proposition 14.5.9 reproduces (and slightly generalizes) one case of
[120, Theorem 2.6.1].

260

Formalism of difference algebra

For the original classification of rational Dieudonn modules over an

algebraically closed field, see Manins original paper [166] or the book of
Demazure [71]. We do not have a prior reference for Theorem 14.6.3, but
nevertheless we do not believe it to be original.
An equal-characteristic analogue of Remark 14.6.5 is to take F = k((z))
for k an algebraically closed field of characteristic p > 0, with acting
as the Frobenius morphism on k and trivially on z. This special case of
Corollary 14.6.4 is due to Laumon [155, Theorem 2.4.5].
As in Chapter 4, one can interpret what has been done here as the special
case for GLn of a construction for any reductive algebraic group. This point
of view was originally introduced by Kottwitz [148, 149], but a full development of the analogy is the subject of ongoing work of Kottwitz [150] and
Csima [64].

Exercises
(1) For M1 , M2 difference modules over a difference ring R, with M1 dualizable, give a canonical identification of H 1 (M1 M2 ) with the Yoneda
extension group Ext(M1 , M2 ). (Hint: as in Lemma 5.3.3 you may wish to
first reduce to the case M1 = R.)
(2) Prove Lemma 14.4.3. (Hint: Lemma 14.4.2 may be helpful, particularly
for (c) and (d).)
(3) Let F be a difference field (of arbitrary characteristic) containing an element x such that (x) = x for some fixed by which is not a
root of unity. Prove that every finite difference module for M admits a
cyclic vector. (Hint: imitate the proof of Theorem 5.4.2. At the key step,
instead of getting a polynomial in s you should get a polynomial in s ; the
root-of-unity condition forces s to take infinitely many values.)
(4) Let F be the completion of Q p (t) for the 1-Gauss norm, viewed as a difference field for equal to the substitution t  t p . Let V be the difference
module corresponding to the matrix


1 t
A=
.
0 p
Prove that there is a nonsplit short exact sequence 0 V1 V V2
0 with V1 , V2 pure and of norms s1 , s2 with s1 < s2 .
(5) Here is a beautiful example from [120, 1.3] (attributed to B. Gross). Let
p be a prime congruent to 3 modulo 4, put F = Q p (i) with i 2 = 1,
and let be the automorphism i  i of F over Q p . Define a difference
module V of rank 2 over F, using the matrix

Exercises


1 p
A=
( p + 1)i

261

( p + 1)i
.
p1

Compute the Newton polygons of A and V and verify that they do not
coincide. (Hint: find another basis of V , on which  acts diagonally.)
(6) (a) Prove that every difference field can be embedded into a strongly
difference-closed field. (Hint: use your favorite equivalent of the
axiom of choice, e.g., Zorns lemma or transfinite induction.)
(b) Prove that every complete isometric discretely valued difference field
can be embedded into a complete isometric difference field with the
same value group but having a strongly difference-closed residue field.
(7) Prove that Theorem 14.6.3 continues to hold if the hypothesis that F is
discretely valued is replaced by one of the following hypotheses.
(a) There exists a constant  > 0 with the following property: for c
{0, 1} and x o F there exists y o F with |(y) cy x| < |x|.
(Hint: use  to replace the -adically convergent sequences in the
proof of Theorem 14.6.3.)
(b) The field F is spherically complete. (Hint: attempt to form convergent sequences as in (a) and then invoke spherical completeness if the
sequences fail to converge.)
(8) This exercise is related to the construction of Witt vectors. Let k,  be
perfect fields of characteristic p > 0. Suppose that Ck , C are complete discrete valuation rings, having residue fields identified with k, ,
respectively, whose maximal ideals are generated by p. Prove that any
homomorphism k  lifts uniquely to a homomorphism Ck C .
(Hint: read about Teichmller lifts in the notes.)

15
Frobenius modules

Having introduced the general formalism of difference algebra and made a

more careful study over a complete nonarchimedean field, we specialize to the
sort of power series rings over which we studied differential algebra. We have
seen most of these rings before, but we will encounter a couple of new variations, notably the Robba ring. This ring consists of power series convergent on
some annulus of outer radius 1 but with unspecified inner radius, which may
vary with the choice of power series. This may seem to be a strange construction at first, but it is rather natural from the point of view of difference algebra:
the endomorphisms that we will consider (Frobenius lifts) do not preserve the
region of convergence of an individual series but do act on the Robba ring as a
whole.
This chapter serves mostly to set the definitions and notation for what follows. One nontrivial result here is the behavior of the Newton polygon under
specialization. Remember that Hypothesis 14.0.1 is still in force, so the field
K will always be discretely valued.

15.1 A multitude of rings

One can talk about Frobenius structures on a variety of rings; for convenience,
we review here the definitions of some rings introduced in Chapter 8 and then
add some special notation that will be useful later.
Remark 15.1.1. The following rings were defined in Chapter 8:

K /t, t/ =
K t 0 =

iZ


i=0

262


ci t : ci K , lim |ci | = 0, lim |ci | = 0 ,
i

i+


ci t i : ci K , sup{|ci |} < ,
i

15.1 A multitude of rings

K {t} =




K /t, t 0 =


K /t, t} =


ci t i : ci K , lim |ci | i = 0 ( (0, 1)) ,
i

i=0


ci t : ci K , lim |ci | = 0, sup{|ci |} < ,
i

iZ

iZ

263

ci t : ci K , lim |ci | = 0, lim |ci | = 0

i

i+

( (0, 1)) .

Definition 15.1.2. For later use, we give a special notation to certain rings
appearing in this framework. We have already defined E to be the completion
of o K ((t)) oK K = K t0 [t 1 ] for the 1-Gauss norm; that is, E consists of

formal sums i ci t i which have bounded coefficients and satisfy |ci | 0 as

i . Since we are assuming that K is discretely valued, E is a complete
nonarchimedean field with residue field K ((t)). We next set
$
E =
K /t, t0 ;
(0,1)

i
ci t which have bounded coefficients
that is, E consists of formal sums
and converge in some range |t| < 1. This ring is sometimes called the
bounded Robba ring, since it consists of the bounded elements of the Robba
ring (see Definition 15.1.4 below). Note that it can also be written as
$
K /t, t0 .
E =
(0,1)

Lemma 15.1.3.
(a) The ring E is a field.
(b) Under the 1-Gauss norm | |1 , the valuation ring oE is a local ring with
maximal ideal m K oE .
(c) The field E , equipped with | |1 , is henselian (see Remark 3.0.2).
This last property implies that finite separable extensions of E = K ((t))
lift functorially to finite tale extensions of oE (and to unramified extensions
of E ). In particular, the maximal unramified extension E ,unr carries an action
of the absolute Galois group of K ((t)).
Proof. To check (a), note that the Newton polygon of any nonzero x E
has finite width (since K is discretely valued). We can thus choose such
that x K /t, t0 has no slope between 0 and log . In this case x is a
unit in K /t, t0 by Lemma 8.2.6(c), yielding (a). We may deduce (b) as an
immediate corollary of (a).
To prove (c) it suffices to check the criterion of Remark 3.0.2: for any monic
polynomial P(x) oE [x] and any simple root r E of P E [x], there

264

Frobenius modules

exists a unique root r oE of P lifting r . In fact, there exists a unique such

root r oE by Hensels lemma (see Remark 2.2.3), so it suffices to produce
such a root in oE .

We may rescale P by a unit in oE to force P (r ) = 1. Choose r0 oE
lifting r . We can then choose (0, 1) such that K /t, t0 contains both r0
and the coefficients of P and, in addition, |P(r0 )| , |1 P  (r0 )| < 1. Since
|P(r0 )|1 , |1 P  (r0 )|1 < 1, this implies that |P(r0 )| , |1 P  (r0 )| < 1 for
[, 1) by Proposition 8.2.3(c).
We may now finish using a standard application of Newtons method. Given
r0 as defined above, for i 0 set ri+1 = ri P(ri )/P  (ri ). We find by
induction on i (exercise) that |P(ri+1 )| |P(ri )|2 and |1 P  (ri )| < 1
for [, 1). Consequently, the ri form a Cauchy sequence under | |
for each [, 1); since K /t, t0 is Frchet complete for these norms
(Proposition 8.2.5), the sequence has a limit r of the desired form. This
yields (c).
We next confect another novel but important ring.
Definition 15.1.4. Put
R=

K /t, t};

(0,1)

that is, R consists of formal sums i ci t i which converge in some range

|t| < 1 but need not be bounded. The ring R is commonly known as the
Robba ring having coefficients in K .
Remark 15.1.5. Be aware that, since R consists of series with unbounded
coefficients, the 1-Gauss norm | |1 is not defined on the whole of R. We will,
conventionally, write |x|1 = + if x R has unbounded coefficients. A
related issue is that R is not a field; indeed, it is not even noetherian, by a similar argument as that for K {t} (see the exercises for Chapter 8). We will return
to this and related ring-theoretic issues concerning R later (Section 16.2).

15.2 Frobenius lifts

Next, we equip these rings with a particular sort of endomorphism.
Definition 15.2.1. Let q be a power of p. Let R be one of the following rings:
K t, K t0 , or K {t};
the union of K /t, t, K /t, t0 , or K /t, t} over all (0, 1);
F1 , the completion of K (t) for the 1-Gauss norm;
E, the completion of K t0 [t 1 ] for the 1-Gauss norm.

15.2 Frobenius lifts

265

By a q-power Frobenius lift on R we will mean a map : R R of the form



ci t i 
K (ci )u i ,
i

where:
the map K : K K is an isometry;
the element u R satisfies |u t q |1 < 1. (If R = R, this forces
u E .)
(These are not the most inclusive conditions possible; see the notes.) The most
important case is when is absolute, i.e., when K provides the q-power
Frobenius lift on K .
Remark 15.2.2. Note that, unless is absolute, the property that it is a
Frobenius lift depends on the implicit choice of the series parameter t; that
is, is not invariant under isometric automorphisms of R.
Remark 15.2.3. Note that, in Definition 15.2.1, one cannot define a Frobenius
lift on an individual ring like K /t, t0 ; for instance, the simple substitution
t t q carries K /t, t0 to K  1/q /t, t0 . One can make this remark more
quantitative, as in the following lemma.
Lemma 15.2.4. Let be a Frobenius lift on E . Then there exists 
(0, 1) such that, for , [, 1) with , carries K /t, t/  to
K  1/q /t, t/ 1/q  and
| f | = |( f )| 1/q .
Proof. Since |(t)t q 1|1 < 1, by continuity we can choose  (0, 1) and
(0, 1) so that |(t)t q 1| 1/q for [, 1]. This inequality implies
that |(t i )t qi 1| 1/q for all i Z. For such an , the claim is easily

verified: if f = i f i t i then, for [, 1),











qi 
qi
i 
K ( f i )t 

K ( f i )(t (t ))
.
| f | = 




i

1/q

1/q

Definition 15.2.5. Let be a Frobenius lift on K t0 . We say that is centered

if there exists m K such that
(t ) 0 (mod t ).
We call such a a center of ; it follows from Lemma 15.2.6 below that
is unique if it exists. Furthermore, it always exists if K is inversive (but not

266

Frobenius modules

always otherwise; see the exercises). We say that is zero-centered if its center
is equal to 0, i.e., if (t) 0 (mod t).
Lemma 15.2.6. Suppose that is a Frobenius lift on K t0 and that K :
K K is inversive. Then is centered and its center is unique.
Proof. Exercise.

15.3 Generic versus special Frobenius lifts

For difference modules over K t0 there are two natural Newton polygons,
and there is an important relationship between them.
Definition 15.3.1. Let M be a finite free difference module over K t0 , for a
centered Frobenius lift . Define the generic Newton polygon of M to be the
Newton polygon of M E. Define the special Newton polygon of M to be the
Newton polygon of M/(t )M, for the center of .
The following result is sometimes called the semicontinuity theorem for
Newton polygons.
Theorem 15.3.2 (Grothendieck, Katz). Let M be a finite free difference module over K t0 , for a centered Frobenius lift . Then the special
Newton polygon lies on or above the generic Newton polygon with the same
endpoints.
Proof. Choose a basis of M, and use it to define supremum norms on M E
and M/(t )M. Then it is evident that, for any positive integer n, the Hodge
polygon of n acting on M/(t )M lies on or above the Hodge polygon
of n acting on M E with the same endpoints. If we divide all slopes by n
and take limits as n then Proposition 14.5.8 implies that the generic or
special Hodge slopes converge to the generic or special Newton slopes.
As in the comparison of Hodge and Newton polygons, one obtains a decomposition result in the case when the special and generic Newton polygons touch
somewhere.
Theorem 15.3.3. Let M be a finite free difference module of rank n over
K t0 , for a centered Frobenius lift . Let sg,1 sg,n and ss,1
ss,n be the generic and special Newton slopes, respectively. Suppose that, for
some i {1, . . . , n 1}, we have
ss,i < ss,i+1 ,

sg,1 + + sg,i = ss,1 + + ss,i .

15.3 Generic versus special Frobenius lifts

267

Then there exists a difference submodule N of M, with M/N free, whose

generic and special Newton slopes are sg,1 , . . . , sg,i and ss,1 , . . . , ss,i , respectively. Moreover, if sg,i < sg,i+1 then N is unique.
We will relax the hypothesis of uniqueness later (Corollary 16.4.7).
Proof. We may assume that is zero-centered. Uniqueness in the case
sg,i < sg,i+1 follows from the uniqueness in M E, as in Theorem 14.4.15.
For the existence, we first replace by a suitable power to ensure that all the
slopes are in the additive value group of K ; we then apply Proposition 14.4.16
followed by Lemma 8.6.1 to change basis in M, in order to ensure that the
generic Hodge slopes of M are also equal to sg,1 , . . . , sg,n .
If ss,H,1 , . . . , ss,H,n denote the special Hodge slopes in this basis then we
have
ss,1 + + ss,i ss,H,1 + + ss,H,i
by Corollary 14.5.4, but also
ss,H,1 + + ss,H,i sg,1 + + sg,i
as in the proof of Theorem 15.3.2 (since sg,1 , . . . , sg,i match the generic
Hodge slopes of M). Consequently, ss,1 + + ss,i = ss,H,1 + + ss,H,i ;
that is, for this basis, the condition of Theorem 14.5.5 is also satisfied by
M/t M.
We can thus change the basis over o K t to obtain a new basis of M on
which the action of  is via the block matrix


B0 C0
,
A0 =
D0 E 0
in which, modulo t, B0 accounts for the first i Hodge and Newton slopes of
M/t M, E 0 accounts for the remaining Hodge and Newton slopes of M/t M,
and D0 vanishes. In particular, det(B0 (mod t)) has the valuation ss,1 + +
ss,i . The valuation of det(B0 ) itself cannot be any greater, but it must be at least
the sum of the first i generic Hodge slopes of M, which we also know to be
ss,1 + + ss,i . Consequently det(B0 ) is a unit in K t0 , and it has minimal
valuation among all i i minors of A0 . Then, by Cramers rule (as in the proof
of Theorem 4.3.11), D0 B01 and B01 C0 have entries in o K t. Note also that
the least Hodge slope of B01 is sg,i , since det(B0 ) is a unit and the entries of
det(B0 )B01 are the cofactors of B0 .
Conjugating by the block lower triangular unipotent matrix U0 with offdiagonal block D0 B01 , we obtain

268

Frobenius modules


A1 = U01 A0 (U0 ) =

B0 + C0 (D0 B01 )
E 0 (D0 B01 ) D0 B01 C0 (D0 B01 )


C0
.
E 0 D0 B01 C0

Since D0 B01 has entries in to K t, we have A1 A0 (mod t). We can thus
iterate the construction to obtain a sequence of block matrices


Bl Cl
Al =
(l = 0, 1, . . . )
Dl El
for which Al+1 = Ul1 Al (Ul ), with Ul GLn (o K t) equal to the block
lower triangular unipotent matrix with off-diagonal block Dl Bl1 . Note that
Bl+1 = Bl (I + Bl1 Cl (Dl Bl1 )), in which Bl1 Cl (Dl Bl1 ) has entries in
(t, m K ), so the Hodge slopes of Bl are preserved. In particular, the least Hodge
slope of Bl1 equals sg,i .
To complete the proof we must show that the Ul converge to the identity for
the (t, m K )-adic topology. Since the least Hodge slope of Bl1 equals sg,i , it is
enough to check that the Dl converge to zero for the same topology. Moreover,
it suffices to check that, for each positive integer m, Dl (mod t m ) converges
to zero for the m K -adic topology.
We do this by induction on m, the case m = 1 being clear because Dl 0
(mod t). Assume that the convergence is known modulo t m . Write
Dl+1 = El (Dl )(Bl1 ) Dl (Bl1 Cl )(Dl Bl1 ).

Setting Dl = h Dl,h t h , we then have

Dl+1,m t m El (Dl,m t m )(Bl1 ) Dl,m t m (Bl1 Cl )(Dl Bl1 )
+

(mod t m+1 ),

(15.3.3.1)

where the ellipses represent terms depending on Dl,0 , . . . , Dl,m1 which are
already known to converge to 0 as l .
On the one hand the sum of the least Hodge slopes of the reductions of
El and (Bl1 ) modulo t equals sg,i+1 sg,i 0. Thus the reduction of
El (Dl,m t m )(Bl1 ) modulo t m+1 has p-adic valuation no less than that of the
reduction of (Dl,m t m ), which in turn has p-adic valuation strictly greater than
that of the reduction of Dl,m t m . On the other hand, Bl1 Cl has entries in o K t,
and (Dl Bl1 ) has entries in the ideal (t, m K ) since Dl is divisible by t. Thus
the reduction of Dl,m t m (Bl1 Cl )(Dl Bl1 ) modulo t m+1 has valuation strictly
greater than that of the reduction of Dl,m t m . We may conclude that, for any
c > 0, for l sufficiently large (so that the terms represented by the ellipses in
(15.3.3.1) all have norm less than c) we have |Dl+1,m | < max{|Dl,m |, c}. This
yields the desired convergence.

15.4 A reverse filtration

269

Theorem 15.3.4. Let M be a finite free difference module of rank n over

K t0 , for a zero-centered Frobenius lift. Suppose that the generic and special Frobenius slopes of M are all equal to a single value r . Then there is a
canonical isomorphism M
= (M/t M) K K t0 of differential modules.
Proof. First suppose that r = 0. By Lemma 8.6.1 we can choose a basis for
which the generic Hodge slopes are all equal to 0. Let A be the matrix of action

i
of  on this basis. We wish to construct an n n matrix U =
i=0 Ui t
over o K t, with U0 = In , such that U 1 A(U ) = A0 or, equivalently,
1
U = A(U )A1
0 . Since the map U  A(U )A0 is contractive for the
(t, m K )-adic topology on In + t Mnn (o K t), it has a unique fixed point, by
the contraction mapping theorem, which gives the desired isomorphism.
If r v(K ), we may apply the above argument after twisting by a scalar.
Otherwise we may replace by a power and then twist and apply the above
argument.

15.4 A reverse filtration

E

Since
is not complete, we cannot apply Theorem 14.4.15 to filter a finite
difference module over E with pure quotients of decreasing norms. It was
originally observed by de Jong that one can get a filtration with pure quotients
of increasing norms, but at the expense of replacing E with its -perfection.
Definition 15.4.1. Let be a Frobenius lift on E . Let E denote the
-perfection of E, that is, the completion of the direct limit R0 R1 ,
with Ri = E and the transition map Ri Ri+1 being . Choose  as in
Lemma 15.2.4. For [, 1) we may define | | on E as the function
| | 1/q i on Ri ; this is consistent because Lemma 15.2.4 guarantees that the
transition maps are isometries. Let E be the subring of x E such that,
for some [, 1) depending on x, |x| < + for [, 1]. Note that,
within E ,
E E = E .
We obtain the following analogue of Lemma 15.1.3.
Lemma 15.4.2.
(a) The ring E is a field.
(b) Under the 1-Gauss norm | |1 , the valuation ring oE is a local ring

with maximal ideal m K oE .

(c) The field E equipped with | |1 is henselian.

270

Frobenius modules

Proof. Exercise.
Example 15.4.3. Suppose that K is the identity and (t) = t q . Then E may

be identified with the set of formal sums iZ[1/ p] ci t i having the property
that, for each  > 0, the set of i Z[1/ p] for which |ci |  is both bounded
below and has bounded denominators. In this interpretation, E consists of

those sums i ci t i for which there exists (0, 1) for which |ci | i 0
as i . (Equivalently, there exists (0, 1) for which |ci | i remains
bounded as i .)
Theorem 15.4.4. Let V be a finite free dualizable difference module over E .
Then there exists a unique filtration
0 = V0 V1 Vl = V
of difference modules, such that each successive quotient Vi /Vi1 is pure and
of some norm si and s1 < < sl .
Proof. By Theorem 14.4.15 we obtain a decomposition of V  = V E E

as a direct sum s Vs of difference modules in which each nonzero Vs is pure
of norm s. By replacing with a power, we can ensure that each s for which
Vs  = 0 appears in |F |. It suffices to check that the summand of V  of least
norm descends to a submodule of V , as we may then repeat the argument after
quotienting by this submodule.
By twisting and then approximating a suitable basis of V  , we obtain a basis
of V on which the matrix of action of 1 is a block matrix


A B
,
C D
in which A is invertible over oE and B, C, D have entries in m K oE . Put

X = C A1 and then change basis using the block lower triangular unipotent
matrix with lower left block X , to obtain the new matrix of action


 
A + B 1 (X )
B
A B
=
.
C  D
(D X B) 1 (X ) D X B
Repeating this operation yields a p-adically convergent sequence of changes
of basis; in the limit, we get a block upper triangular matrix in which the first
block corresponds to the summand of V  of least norm. To prove that this
summand descends to a submodule of V , it suffices to check that the changeof-basis matrices are bounded in norm by 1 under | | for some (0, 1).
At the first stage, for sufficiently close to 1 we have
|A1 | max{|B| , |C| , |D| } < 1,

Notes

271

which implies that |X | < 1. Since we also have |X |1 < 1, by

Proposition 8.2.3(c) we have |X | 1/q < 1; by Lemma 15.2.4, | 1 (X )| =
|X | 1/q < 1. Hence
|A1 A In | < 1,


max{|B | , |C  | , |D| } max{|B| , |C| , |D| }.

We may thus use the same for the next stage, so |X | < 1 at all stages.
In order to use the reverse filtration later, we will need the following
projection construction.
Lemma 15.4.5. There exists a E-linear map : E E sending E to E ,
such that (x) = x for all x E.
Proof. Suppose first that K is inversive. In this case, the residue field of E
((t 1/q i )). Each element of this union can be written
can be written as i=0
K

1
(xi ), with xi K ((t)) such that the
uniquely as the finite sum i=0
j
coefficient of t in xi vanishes whenever i > 0 and j is divisible by q.
This implies by an easy induction that each x E can be written uniquely

i
(xi ), with xi E
as a convergent (for the m K -adic topology) sum i=0
such that the coefficient of t j in xi vanishes whenever i > 0 and j is divisible
by q. In this presentation, for  as in Lemma 15.2.4 we have | i (xi )| |x|
for [, 1) (exercise). We may thus take (x) = x0 .
In the general case, let K  be the -perfection of K and put E = K  1/t, t0 .
(That is, E is the analogue of E with base field K  instead of K .) Argue as above
Then choose any continuous linear map K  K
to construct a map E E.
whose composition with the inclusion K K  is the identity, and use it to
define a projection E E. The composition E E E has the desired
effect.
Corollary 15.4.6. The multiplication map : E E E E is injective.
Proof. Exercise.

Notes
Much existing literature makes the restriction that Frobenius lifts must be
absolute. The generalization that we have considered here is relevant for
some applications (e.g., to families of Galois representations in p-adic Hodge
theory), so it is prudent to allow it as much as possible.
However, we have not opted to consider cases where does not carry
K into K . (At that level of generality, one might consider that our definition

272

Frobenius modules

restricts to scalar-preserving Frobenius lifts.) Also, we have not allowed K to

be nondiscrete; this avoids having to worry about whether to allow Frobenius
lifts for which (t) t q has 1-Gauss norm 1 but each of its coefficients has
norm less than 1. (This cannot occur for K discrete.)
As in the previous chapter, a number of the results are based on [120], at
least when restricted to an absolute Frobenius lift.
Theorem 15.3.2 is a local formulation of a geometric result of
Grothendieck. The proof given is from [120, Theorem 2.3.1].
Theorem 15.3.3 is a variant of [120, Theorem 2.4.2].
Theorem 15.3.4 is an adaptation of [120, Theorem 2.7.1].
For stronger global results along these lines (in the theory of F-crystals), see
[67], [226], and [176].
Theorem 15.4.4 in the case of an absolute Frobenius lift is due to de Jong
[65, Proposition 5.8]. Corollary 15.4.6 and its proof are a variant on [65,
Proposition 8.1].

Exercises
(1) Verify the claims about Newton iteration from the proof of
Lemma 15.1.3(c).
(2) Suppose that K is perfect and that : E E is a continuous isometric
endomorphism inducing the absolute q-power Frobenius lift on K ((t)).
Prove that is an absolute Frobenius lift in the sense of Definition 15.2.1.
(Hint: use Witt vector functoriality to show that carries K into K . In
n
n
other words, for x K , show that (x p ) = (x) p is p-adically close to
an element of K .)
1
( (t)|t= ) on
(3) Prove Lemma 15.2.6. (Hint: show that the map  K
m K is contractive; here (t)|t= denotes the substitution t  .)
(4) Let K be the completion of Q p (x) for the 1-Gauss norm. Prove that the
Frobenius lift defined by (t) = t p + px does not have a center.
(5) Prove Lemma 15.4.2.
(6) Check the omitted details in the proof of Lemma 15.4.5. (Hint: you can
in fact show that |x| = supi {| 1 (xi )| }. To check this, it suffices to
m (E ) of E .)
consider x in the dense subset

m=0
(7) Prove Corollary 15.4.6. (Hint: if the claim fails, we can choose a nonzero

n
xi yi of the kernel with n as small as possible. Now
element z = i=1
use Lemma 15.4.5 to construct an element with an even shorter representation, as in the proof of Lemma 1.3.11. A related argument in Galois theory
is Artins proof of the linear independence of characters.)

16
Frobenius modules over the Robba ring

In Chapter 14 we discussed some structure theory for finite difference modules

over a complete isometric nonarchimedean difference field. This theory can be
applied to the field E, which is the p-adic completion of the bounded Robba
ring E ; however, the information it gives is somewhat limited.
For the purposes of studying Frobenius structures on differential modules
(see Part V), it would be useful to have a structure theory over E itself. This is
a bit too much to ask for; what we can provide is a structure theory that applies
over the Robba ring R, which is somewhat analogous to what we obtain over
E. In particular, with an appropriate definition of pure modules, we obtain a
slope filtration theorem analogous to Theorem 14.4.15 but valid over R.
Given a difference module over E , one obtains slope filtrations and Newton
polygons over both E and R. For a module over K t0 these turn out to
match the generic and special Newton polygons, and so in particular they need
not coincide. However, they do admit a specialization property analogous to
Theorem 15.3.2.
Unfortunately, a proof of the slope filtration theorem over R would take us
rather far afield, so we do not include one here. Instead, we limit ourselves to
a brief overview of the proof and consign further discussion and references to
the notes.
Hypothesis 16.0.1. Throughout this chapter, let be a Frobenius lift on the
Robba ring R. (For a possible relaxation of this hypothesis, see the notes.) All
difference modules over R will be taken with respect to unless otherwise
specified.

16.1 Frobenius modules on open discs

We start with the fact that, over the ring K {t}, the classification of finite free
difference modules reduces (mostly) to classification over K .
273

274

Frobenius modules over the Robba ring

Theorem 16.1.1. Suppose that K is strongly difference-closed (e.g., is an

absolute Frobenius lift and K is algebraically closed). Let M be a finite
free dualizable difference module over K {t} for a zero-centered Frobenius
lift. Then there exists a noncanonical isomorphism of difference modules
M
= (M/t M) K K {t}.
Proof. Let c be the reduction of (t)/t modulo t, so that c m K . Let
e1 , . . . , en be a basis for M. Let A be the matrix of action of  on this basis,
which is invertible because M is dualizable. Let A0 be the constant term of A,
which is an invertible matrix over K . We wish to find a matrix U over K {t}
with U In (mod t) such that U 1 A(U ) = A0 .
We first construct a sequence of matrices U0 , U1 , U2 , . . . over K t, with
U0 = In and Ui+1 = Ui (In + X i t i ) for some matrix X i over K such that
i+1 ). Namely, given U , we find X by solving
Ui1 A(Ui )A1
i
i
0 In (mod t
the equation
1
1
i
ci A0 (X i )A1
0 X i +t (Ui A(Ui )A0 In ) 0

(mod t), (16.1.1.1)

which amounts to trivializing a class in H 1 (V ) for some finite difference

module V over K . This is possible by Corollary 14.6.6, although it can
not necessarily be done uniquely. However, let us choose h 0 such
that |ch ||A0 ||A1
0 | < 1. Then, for i h and for any choice of X i ,
|ci A0 (X i )A1
0 | < |X i |, so that X i is uniquely determined by (16.1.1.1);
moreover, it has the same norm as the reduction of t i Ui1 A(Ui )A1
0 In
modulo t.
Choose so that, for i = h,
|Ui In | , |Ui1 In | , |Ui1 A(Ui )A1
0 In | < 1.

(16.1.1.2)

Then, by the previous paragraph, |X i t i | |Ui1 A(Ui )A1

0 In | . By our
i | ; since
|
<
|X
t
choice of h, |A0 (X i t i )A1

0
1
1
1
i 1 1
i
Ui+1
A(Ui+1 )A1
0 = (In + X i t ) Ui A(Ui )A0 (In + A0 (X i t )A0 ),

we may conclude that (16.1.1.2) holds also with i replaced by i + 1. By

induction, we see that (16.1.1.2) holds for all i h.
This means that, for any < , the matrices U and U 1 have entries in
K t/. Using the relation U = A(U )A1
0 , we may conclude that U and
U 1 also have entries in K t/  , for  equal to the minimum norm of a root
of (t) c over all c K alg of norm . Since is zero-centered, by considering the Newton polygon of (t) c we see that  min{ 1/q , /| |} for
a generator of m K . Iterating the function  min{ 1/q , /| |}, we see that
U and U 1 have entries in K t/ for all (0, 1), proving the claim.

16.2 More on the Robba ring

275

Remark 16.1.2. We will see a similar result later in the presence of a differential structure. This is introduced as Dworks trick (Corollary 17.2.2). The main
differences are that in the setting of Dworks trick there will be a canonical
isomorphism M
= (M/t M) K K {t}, and no restriction on K is needed.
Remark 16.1.3. Note that the function  min{ 1/q , /| |} becomes
exactly the function  min{1/ p , p} from Chapter 10 in the case where
q = p and K is absolutely unramified. See also Theorem 17.2.1.
One can apply Theorem 16.1.1 to coherent locally free modules on the open
unit disc, by virtue of the following observation. (Compare Definition 8.4.3,
remembering that in this part of the book K is assumed to be discretely valued.)
Proposition 16.1.4. Any coherent locally free module on the open unit disc
over K is freely generated by finitely many global sections.
Proof. Let M be such a module. Choose a sequence 0 < 1 < 2 < with
limit 1, and put Mi = M K t/i . Choose any basis B1 of M1 . Given a
basis Bi of Mi , there must exist X i GLn (K t/i ) such that changing basis
from Bi via X i produces a basis of Mi+1 . By Lemma 8.3.4 we can factor X i as
Vi Wi , with Vi GLn (K t/i ) such that |Vi In |i < 1 and Wi GLn (K [t]).

j
j
Write Vi =
j=0 Vi, j t , let Yi be the sum of Vi, j t over only those j for
which |Vi, j |1 1, and put Ui = Vi Yi1 . Changing basis from Bi via Ui also
produces a basis Bi+1 of Mi+1 .
Since K is discretely valued, for each positive integer j and for i sufficiently
large (depending on j), we have |Vi, j t j |1 1 if and only if |Vi, j t j |i
1. Consequently, the Ui converge t-adically to the identity matrix. Given any
(0, 1), if we choose (, 1) and combine the previous observation
with the fact that the Ui are bounded under | | , we may deduce that the
Ui converge under | | to the identity matrix. Hence the product U1 U2
is the change-of-basis matrix from B1 to a simultaneous basis of each Mi , as
desired.

16.2 More on the Robba ring

To discuss the classification of difference modules on annuli, we must recall
some properties of the Robba ring beyond simply its definition.
Remark 16.2.1. Recall (Definition 15.1.4) that we have defined the Robba
ring to be
R = (0,1) K /t, t};

276

Frobenius modules over the Robba ring

that is, R consists of formal sums i ci t i that converge in some range

|t| < 1 but need not have bounded coefficients. Unlike its subring E , R is not
a field; for instance the element
log(1 + t) =


(1)i1
i=1

ti

is not invertible (because its Newton polygon has infinitely many slopes). More
generally, we have the following easy fact.
Lemma 16.2.2. We have R = (E ) = E \ {0}.
Proof. On the one hand, a unit in K /t, t} must have empty Newton polygon,
so it must belong to E . On the other hand, E is a field by Lemma 15.1.3(a).
This proves the claim.
Definition 16.2.3. Because R consists of series with possibly unbounded
coefficients, it does not carry a natural p-adic norm or topology. The most useful topology on R is the LF topology, which is the direct limit of the Frchet
topology on each K /t, t} defined by the | | for [, 1). That is, a
sequence converges if on the one hand it is contained in some K /t, t} and
on the other hand it converges for the Frchet topology on that ring. Note
that it does not matter which is chosen since, for > , the inclusion
K /t, t} K  /t, t} is strict (i.e., it is a homeomorphism to its image
equipped with the subspace topology).
In fact, the ring R is not even noetherian (by an argument similar to that
for K {t}; see the exercises for Chapter 8), but the following useful facts,
mentioned in the notes, are true.
Proposition 16.2.4. For an ideal I of R, the following are equivalent.
(a) The ideal I is closed in the LF topology.
(b) The ideal I is finitely generated.
(c) The ideal I is principal.
Remark 16.2.5. The equivalence of (b) and (c) in Proposition 16.2.4 implies
that R is a Bzout domain, i.e., an integral domain in which every finitely
generated ideal is principal. Such rings enjoy many properties analogous to
principal ideal domains; see the exercises.
We also have the following analogue of Proposition 16.1.4.
Proposition 16.2.6. Any coherent locally free module on the half-open annulus with closed inner radius and open outer radius 1 is represented by a

16.3 Pure difference modules

277

finite free module over K /t, t} and so corresponds to a finite free module
over R.
Proof. Let M be such a module. Then M K /t, t/ is finite free since
K /t, t/ is a principal ideal domain by Proposition 8.3.2. By choosing a
basis of M K /t, t/ and then invoking Lemma 8.3.6, we may extend M
to a coherent locally free module on the whole unit disc. We may then deduce
the claim from Proposition 16.1.4.

16.3 Pure difference modules

Over a complete nonarchimedean difference field such as E , we already have
a notion of a pure difference module. Since R does not come with a Frobeniusinvariant norm we cannot use the same definition. The appropriate definition
in this case is as follows.
Definition 16.3.1. A finite free difference module M over R is pure of norm
s if there exists a finite free difference module M over E which is pure of
norm s (in the sense of Definition 14.4.6) and such that M
= M E R. We

will see shortly (Corollary 16.3.7) that the module M is uniquely determined
by this requirement.
Lemma 16.3.2. Let A be an n n matrix over oE . Then the map v  v
A(v) induces a bijection on (R/E )n .
Proof. Exercise, or see [136, Proposition 1.2.6].
Corollary 16.3.3. Let M be a finite free difference module over E such that
||sp,ME 1. Then the map H 1 (M) H 1 (M E R) is bijective.
Proof. This will follow from Lemma 16.3.2 once we produce a basis of M
on which the matrix of action of  has entries in oE . Such a basis exists, as
explained in the proof of Proposition 14.5.9.
Corollary 16.3.4. Let M1 , M2 be two finite free dualizable difference modules
over E . Suppose that every Newton slope of M1 E E is less than or equal to
every Newton slope of M2 E E. Then
Hom(M1 , M2 ) = Hom(M1 E R, M2 E R).
Proof. Note that (M1 ) M2 has norm less than or equal to 1 (because its
Newton slopes are all nonnegative). It thus suffices to check that, for M of
norm s 1, H 0 (M ) = H 0 (M E R). As in the previous proof, we
can find a basis e1 , . . . , en of M such that the matrix A of action of  has

278

Frobenius modules over the Robba ring

entries in oE . Any element of H 0 (M E R), when written in terms of the

basis, corresponds to a column vector v Rn satisfying A(v) = v. By
Lemma 16.3.2 this forces v (E )n , so v corresponds to an element of M
itself.
Corollary 16.3.5. Let M1 , M2 be finite free difference modules over R that
are pure of norms s1 , s2 , respectively. If s1 > s2 > 0 then Hom(M1 , M2 ) = 0.
Proof. By Corollary 16.3.4 it suffices to check that if M1 , M2 are finite free
difference modules over E that are pure of norms s1 , s2 , respectively, and s1 >
s2 then Hom(M1 , M2 ) = 0. This holds because the image of any morphism
M1 M2 , if nonzero, would have to be both pure of norm s1 and pure of
norm s2 , which is impossible.
Remark 16.3.6. In the proof of Corollary 16.3.5 we used the fact that, when
working over a difference field, purity of the middle term in a short exact
sequence 0 M1 M M2 0 implies purity (of the same norm)
at the extremes. This is not true over R, which means that one cannot deduce
Corollary 16.3.5 in the case s1 < s2 . For example, if is a Frobenius lift for
which (1 + t) = (1 + t) p , and M = Rv with (v) = p 1 v, then
log(1 + t)v H 0 (M);
this constitutes a key example in p-adic Hodge theory (see Chapter 24).
Corollary 16.3.7. Let M be a finite free difference module over R that is pure
of norm s > 0. Then there is a unique finite free difference module M over
E , which is pure of norm s, such that M
= M E R.
Remark 16.3.8. Even if M is pure of norm s, the uniqueness in
Corollary 16.3.7 can fail if we do not require M to be pure.
Proposition 16.3.9. Let M1 , M2 be finite free difference modules over E that
are pure of the same norm s > 0. Then any finite free difference module M
over R fitting into a short exact sequence of the form
0 M1 E R M M2 E R 0
is also pure of norm s.
Proof. It is equivalent to check that
Ext(M1 , M2 ) = Ext(M1 E R, M2 E R).
If we set M = (M1 ) M2 , it is also equivalent to check that the natural map
H 1 (M ) H 1 (M E R)
is a bijection (see Definition 14.1.6); this holds by Corollary 16.3.3.

16.4 The slope filtration theorem

279

16.4 The slope filtration theorem

The fundamental theorem in the theory of difference modules over the Robba
ring is the following result, which is analogous to Theorem 14.4.15. We do not
give a complete proof in this book, but we will sketch the argument in the next
section. (As noted in Remark 14.4.7, this theorem is usually stated in additive
rather than multiplicative terms; this explains our choice of the name slope
filtration theorem rather than norm filtration theorem.)
Theorem 16.4.1 (Slope filtration theorem). Let M be a finite free dualizable difference module over R. Then there exists a unique filtration 0 =
M0 Ml = M by dualizable difference submodules with the following
properties.
(a) Each successive quotient Mi /Mi1 is finite free and is pure of some
norm si (in the sense of Definition 16.3.1).
(b) We have s1 > > sl .
In this book the main application of the slope filtration will be to the p-adic
local monodromy theorem for differential modules over R (Theorem 20.1.4).
For the moment, let us record one or two additional corollaries. (For further
applications, see the notes.)
Definition 16.4.2. Let M be a finite free dualizable difference module over R.
Set the notation as in Theorem 16.4.1. Define the Newton polygon of M to be
the polygon associated with the multiset containing log si with multiplicity
rank(Mi /Mi1 ).
By the compatibility of purity of difference modules over E with tensor
products (Corollary 14.4.9) and duals (Proposition 14.4.8), we obtain the usual
behavior of Newton slopes under tensor products, exterior powers, and duals.
Lemma 16.4.3. Let M, N be finite free dualizable difference modules over
R. Let s M,1 , . . . , s M,m and s N ,1 , . . . , s N ,n be the Newton slopes of M and N ,
respectively.
(a) The Newton slopes of M N are s M,i + s N , j for i = 1, . . . , m and
j = 1, . . . , n.
(b) For j = 1, . . . , rank(M), the Newton slopes of j M are s M,i1 + +
s M,i j for 1 i 1 < < i j m.
(c) The Newton slopes of M are s M,1 , . . . , s M,m .
We have the following generalization of Corollary 16.3.5.
Proposition 16.4.4. Let M1 , M2 be finite free dualizable difference modules
over R. Suppose that every Newton slope of M1 is less than every Newton
slope of M2 . Then Hom(M1 , M2 ) = 0.

280

Frobenius modules over the Robba ring

Proof. We induct on rank(M1 ) + rank(M2 ). If M1 and M2 are both pure then

Corollary 16.3.5 yields the claim. If M1 is not pure, by Theorem 16.4.1 there
exists a proper nonzero difference submodule N of M1 such that N and M1 /N
are both finite free, and we deduce that Hom(M1 , M2 ) = 0 from the fact that
Hom(N , M2 ) = Hom(M1 /N , M2 ) = 0. (Namely, any morphism from M1
to M2 must vanish on N , yielding a morphism from M1 /N to M2 that also
vanishes.) We argue similarly if M2 fails to be pure.
We now mention a result that extends the semicontinuity theorem for
Newton polygons (Theorem 15.3.2). See Example 20.2.1 for an explicit
example.
Definition 16.4.5. For M a finite free dualizable difference module over E ,
we can construct two different Newton polygons, namely those associated with
M E E and M E R. We call these the generic Newton polygon and the special Newton polygon, respectively. The terminology is justified by the fact that
if K is strongly difference-closed, is induced by a zero-centered Frobenius
lift on K t0 , and M is obtained by base extension from a finite free dualizable
difference module M0 over K t0 , then by Theorem 16.1.1 the special Newton
polygon of M coincides with the special Newton polygon of M0 .
Theorem 16.4.6. Let M be a finite free dualizable difference module over E .
Then the special Newton polygon lies on or above the generic Newton polygon
with the same endpoints.
Proof. It suffices to show that the least special slope of M is greater than or
equal to the least generic slope, as then applying this inequality to the exterior
powers of M yields the comparison of Newton polygons by Lemmas 14.5.3
and 16.4.3.
Suppose on the contrary that the least special slope is less than the least
generic slope. Let M1 be the first step in the filtration of M E R given by
Theorem 14.4.15. Since M1 is pure, it descends to a difference module N1
over E of the same norm. By hypothesis, the generic slopes of N1 M are
all positive. It follows that H 0 (N1 M) = 0 are and then by Corollary 16.3.4
that H 0 ((N1 M) E R) = 0 also. However, the latter contradicts the fact
that N1 is a submodule of M E R. We thus deduce the claim.
We can similarly use slope filtrations to finish the uniqueness part of
Theorem 15.3.3.
Corollary 16.4.7. With the notation of Theorem 15.3.3, the submodule N is
always unique.
Proof. It suffices to check that N K t 0 R is uniquely determined; this follows
because it coincides with one step of the slope filtration of M K t 0 R.

16.5 Proof of the slope filtration theorem

281

Remark 16.4.8. Note that in Corollary 16.4.7 we need only the uniqueness of
the slope filtration in Theorem 16.4.1. The existence, which lies much deeper,
can be supplied using Theorem 16.1.1 because we start with a difference
module over K t0 .

16.5 Proof of the slope filtration theorem

The proof of Theorem 16.4.1 is unfortunately in some respects orthogonal to
much of the material presented in this book. We will thus limit ourselves to an
indication of the key steps. See the notes for further discussion.
Definition 16.5.1. Let M be a finite free dualizable difference module over R
of rank n > 0. Define the average norm of M to be
1/n

(M) = ||sp,n M ;
this quantity is multiplicative in the sense that (M1 M2 ) = (M1 )(M2 ).
We say M is semistable if we have (M) (N ) for any nonzero difference
submodule N of M.
Lemma 16.5.2. Let M be a nonzero finite free dualizable difference module
over R that is pure of some norm. Then M is semistable.
Proof. Suppose on the contrary that there exists a difference submodule N of
M with (N ) > (M). Put e = rank(N ); then e N is of rank 1 and hence
pure. Consequently, the inclusion e N e M violates Corollary 16.3.5; this
contradiction yields the claim.
Lemma 16.5.3. Let M be a nonzero finite free dualizable difference module over R. There exists a unique filtration 0 = M0 Ml =
M by difference submodules, such that each quotient Mi /Mi1 is free and
semistable and
(M1 /M0 ) > > (Ml /Ml1 ).
Proof. First note that the claim holds for any module of rank 1 by
Lemma 16.5.2. From this point, the rest of the proof is formal; we leave it
as an exercise.
Definition 16.5.4. The filtration in Lemma 16.5.3 is called the Harder
Narasimhan or HN filtration of the module M. To prove Theorem 16.4.1 it
will suffice to check that semistable modules over R are pure.
Remark 16.5.5. The reader may wonder why it is necessary to prove
Theorem 16.4.1 if defining the HN filtration is a mere formality. The answer is

282

Frobenius modules over the Robba ring

that one cannot prove much about the norms occurring in the HN filtration; for
instance, it has not been formally proved that the tensor product of semistable
modules is semistable but this is indeed a consequence of Theorem 16.4.1. A
similar situation arises in the theory of vector bundles, on which the usage
of the word semistable is modeled; see the discussion preceding [136,
Theorem 1.7.1]. Another loosely analogous situation is Delignes use of determinantal weights in his second proof of the Weil conjectures [69]; there one
has matrices over a p-adic field whose eigenvalues are supposed to be algebraic numbers with certain archimedean norms, but one has direct control only
over the determinants.
We need to introduce something like a residual difference closure of the
ring R. (This characterization is justified by Corollary 16.5.8.) This construction, based on Malcev-Neumann series (Example 1.5.8), is called the extended
Robba ring in [136].

i
be the ring of formal sums x =
Definition 16.5.6. Let R
iQ x i t with
xi K that satisfy the following properties. (Here is a number in (0, 1)
which may vary with x.)
(a) For (, 1) and we have limi |xi | i = 0.
(b) For (, 1) and for c > 0, the set of indices i for which |xi | i c
is a well-ordered subset of R (i.e., it contains no infinite decreasing
subsequence).
as follows. Recall that the field K ((t Q )) of MalcevWe may also construct R

Neumann series over K consists of formal sums iQ xi t i for which {i Q :

xi  = 0} is well-ordered. On the subring of K ((t Q )) consisting of series with
bounded coefficients, for (0, 1) define the Gauss norm




 i 
i


x
t
i  = sup{|x i | }.

i

iQ

For each (0, 1) take the Frchet completion for the norms | | , for

(, 1); the union of these completions over all gives R.

turns out to be a Bzout domain (comLike the usual Robba ring, the ring R
consisting of formal sums
pare Proposition 16.2.4). Let E be the subring of R
with bounded coefficients. By imitating the proofs for E (Lemma 15.1.3) it
can be shown that E is a henselian discretely valued field with residue field
K ((t Q )). Various definitions (e.g., purity, semistability) carry over from R to
we will not write these out explicitly.
R;

16.5 Proof of the slope filtration theorem

283

as a difference ring with the Frobenius morphism given by

We view R



xi t i =
K (xi )t qi .

iQ

iQ

The following lemma is where most of the hard work is concentrated.

Lemma 16.5.7. Suppose that K is strongly difference-closed. Then every
is pure of some norm.
semistable dualizable finite difference module over R
Sketch of proof. One first constructs (using an explicit calculation) a filtration
in which the successive quotients are each pure of some norm, but these norms
do not necessarily increase in the right direction as we proceed up the filtration.
One then argues (using a second explicit calculation) that if there are two steps
in the wrong order that cannot be switched, because the extension between
them is not split, then one can change the filtration to move the norms of these
steps closer to each other. (Note the similarity to Grothendiecks classification
of vector bundles on the projective line.) See [136, Theorem 2.1.8] for full
details.
Although we will not need it explicitly, we mention the following analogue
of the generalized DieudonnManin classification (Theorem 14.6.3). It is not
stated explicitly in [136] but is an easy consequence of results from there.
Corollary 16.5.8. Suppose that K is strongly difference-closed. Then every
can be split (non-uniquely) as a
dualizable finite difference module over R
direct sum of difference submodules, each of the form V,d for some , d (as
in Definition 14.6.1).
Proof. By [136, Proposition 2.1.6], the categories of pure dualizable differ are equivalent. Thus, by
ence modules of a given norm over K and over R
Lemma 16.5.7 and Theorem 14.6.3, it suffices to split any short exact sequence
0 M1 M M2 0 in which M1 , M2 are base extensions of pure difference modules over K with (M1 ) > (M2 ). This can be deduced from the
proof of [136, Proposition 2.1.5] or by a simple direct calculation.
It remains to descend in our understanding of semistability and purity from

R to R. We must first make the following observation.

Remark 16.5.9. Let L be a complete extension of K as a difference field, such
that L is strongly difference-closed. (Such an L can always be constructed; see
L denote the
the exercises for Chapter 14 or [136, Proposition 3.2.4].) Let R
extended Robba ring constructed using L as the coefficient field. Then there
L that is isometric;
is always an embedding of difference rings : R R

284

Frobenius modules over the Robba ring

i.e., for some (0, 1) and for (, 1) we have |(x)| = |x| for all
x R (in particular, the two quantities are either both finite or both infinite).
This is clear if for instance (t) = t q , as we can then set (t) = t; it was
verified in general by [136, Proposition 2.2.6].
We then obtain the following results; the key tool used is descent for modules along a faithfully flat ring homomorphism. (We also make considerable
use of an analogue of the projection E E of Lemma 15.4.5.)
Lemma 16.5.10. Let M be a semistable dualizable finite difference module
L is semistable.
over R. Then M R R
Proof. See [136, Theorem 3.1.2].
Lemma 16.5.11. Let M be a dualizable finite difference module over R such
L is pure. Then M is pure.
that M R R
Proof. See [136, Theorem 3.1.3].
Putting everything together, we may prove Theorem 16.4.1 as follows.
Proof of Theorem 16.4.1. We wish to show that the HN filtration of M is pure.
By Lemma 16.5.10, if we start with the HN filtration of M and tensor with
L then the resulting filtration still has semistable successive quotients. ConR
L . By Lemma 16.5.7, each
sequently, it must be the HN filtration of M R R
successive quotient of the HN filtration of M becomes pure when tensored
L . By Lemma 16.5.11 each successive quotient is itself pure.
with R

Notes
The restriction that must be a Frobenius lift is probably stronger than necessary. In fact, as suggested in [136], if the action of on K ((t)) satisfies
v((x)) = mv(x) for some integer m > 1 this should suffice.
Our proof of Theorem 16.1.1 is essentially that of [130, Proposition 4.3],
except that in the latter only the absolute case is treated, and it is further
assumed that M descends to K t0 .
Proposition 16.2.4 is the essential content of a paper of Lazard [156].
Note that it depends on K being spherically complete and is false otherwise;
however, we have assumed in this part that K is discretely valued, so we
are safe.
The treatment of pure modules over R is an abridged version of [136, 1].
Corollary 16.3.4 has appeared in several guises previously. It figures
in the work of Cherbonnier and Colmez [38], which we will discuss in

Notes

285

Chapter 24 (see Remark 24.2.6); it also appears in work of Tsuzuki [206,

Proposition 4.1.1].
The existence of slope filtration for Frobenius modules over the Robba ring
was anticipated by Tsuzuki [209], who introduced the definition in the case
of a Frobenius structure on a differential module [209, Definition 5.1.1]. For
more on why he did this, see the notes to Chapter 20.
The slope filtration theorem itself (Theorem 16.4.1), in the case of an absolute Frobenius lift with K of characteristic 0, was originally proved by Kedlaya
in [125, Theorem 6.10]. That proof was significantly more complicated than
the one given here. A second proof in the absolute case (again with K of characteristic 0), which also gives an important generalization (see below), was
given in [129]; this proof introduced the formalism of semistability, inspired by
some parallel work of Hartl and Pink [104] (on which more below). Our sketch
was modeled on a third proof, this time for an arbitrary Frobenius lift, given
in [136, Theorem 1.7.1]; the latter proof adds several technical simplifications
to the previous proofs. One is the direct characterization of pure modules; the
previous proofs used a characterization in terms of an appropriate analogue
of the DieudonnManin classification (Corollary 16.5.8). Another is the use
of faithful flat descent, replacing a more complicated Galois descent argument
used in the previous proofs. A third simplification is the use of the extended
as described here. In the previous proofs the role of R
was
Robba ring R
played by a somewhat smaller ring: its bounded elements (plus 0) form a field
whose residue field is the algebraic closure of K ((t)) rather than the much
larger field of generalized power series.
The proof of Theorem 16.4.1 in [129] includes a nontrivial generalization
of the original slope filtration theorem that was not covered by [136]. In this
generalization, the ring E is replaced by a Cohen ring (or a ramified extension thereof), not for a field of power series but for a more general complete
nonarchimedean field of characteristic p; the rings E and R are replaced
by appropriate analogues. The resulting theorem plays an important role in
the study of isocrystals; see the notes for Chapter 23. It also is important for
applications to RapoportZink spaces; see below.
In the case of an absolute Frobenius lift, Theorem 16.4.6 becomes [129,
Proposition 5.5.1]. (This reference is also the source of the terminology of
special and generic Newton polygons over E ). The proof in [129] uses the
reverse filtration of de Jong (see Theorem 15.4.4).
Besides the p-adic local monodromy theorem (Theorem 20.1.4), several
additional applications of the slope filtration theorem have been found. One
of these is in p-adic Hodge theory; this is discussed further in the notes to
Chapter 24. Another is a q-difference analogue of the p-adic local monodromy

286

Frobenius modules over the Robba ring

theorem, due to Andr and Di Vizio [8]. A third application has been pursued by Hartl in the context of period morphisms for RapoportZink spaces.
These are moduli spaces for deformations of certain p-divisible groups into
mixed characteristic; the main conjecture of Rapoport and Zink is that a
period morphism between this space and a space of linear-algebraic objects
can be constructed, via which the Galois representations for to the p-divisible
groups correspond to some coherent data. In equal positive characteristic, Hartl
[101] established the appropriate analogue of this conjecture, using a form
of Theorem 16.4.1 in which K is of positive characteristic. (Hartl prove that
case of Theorem 16.4.1 directly, building on work of Hartl and Pink [104]
giving a form of Corollary 16.5.8 for K of characteristic p.) In the original
mixed-characteristic setting, Hartl has given some partial results [102, 103];
a complementary but related partial result has been established by Faltings
(though only the brief research announcement [86] was available at the time
of writing), and the combination of ideas may lead to a proof of the full
RapoportZink conjecture.

Exercises
(1) Show that if is a Frobenius lift such that (t) 0 (mod t 2 ) then the
conclusion of Theorem 16.1.1 holds without any restriction on K .
(2) Let R be a Bzout domain (an integral domain in which every finitely
generated ideal is principal).
(a) Prove that any x1 , . . . , xn R that generate the unit ideal appear as
the first row in some n n invertible matrix over R.
(b) Prove that every finitely generated torsion-free R-module is free.
(3) Prove Lemma 16.3.2. (Hint: reduce to the case where | A| 1 for
[, 1). For the injectivity, given that v A(v) (E )n show that
|v| is bounded for [, 1) by comparing |v| with |v| 1/q , using
Lemma 15.2.4. For the surjectivity, to find the class of w Rn in the
image of the map, separate off the positive terms w+ of w, replace w with
A(w+ ) + (w w+ ), and then repeat this procedure.)
(4) Prove that any A GLn (R) can be factored as U V with U GLn (K {t})
and V GLn (E ). (Hint: imitate the proof of Proposition 8.3.5.)
(5) Let M be a finite free difference module over K {t} such that M K {t} R
is pure of some norm s. Let M be the pure module over E satisfying
M K {t} R
= M E R. Prove that there exists a basis e1 , . . . , en of M
that is also a basis of M . (Hint: imitate the proof of Lemma 8.3.6 using
the previous exercise.)

Exercises

287

(6) Prove Lemma 16.5.3. (Hint: first check that (N ) (M) whenever N
is a difference submodule of M of full rank, by comparing the top exterior
powers of M and N . See also [136, Proposition 1.4.15].)
(7) Let M be a nonzero finite dualizable difference module over R. Let 0 =
M0 Ml = M be any filtration of M by difference submodules,
such that each quotient Mi /Mi1 is free and semistable. Form the convex
polygon of length n associated (as in Definition 4.3.2) with the multiset of
slopes given by log (Mi /Mi1 ), with multiplicity rank(Mi /Mi1 ), for
i = 1, . . . , l. Prove that this polygon lies on or below the corresponding
polygon for the HN filtration with the same endpoint. (Again, this is a
purely formal consequence of the definitions of semistability and the HN
filtration, so the proof is as in the theory of vector bundles.)

Part V
Frobenius Structures

17
Frobenius structures on differential modules

In this part of the book, we bring together the streams of differential algebra
(from Part III) and difference algebra (from Part IV), realizing Dworks fundamental insight that the study of differential modules on discs and annuli is
greatly enhanced by the introduction of Frobenius structures.
This chapter sets the foundations for this study. First, we introduce the
notion of a Frobenius structure on a differential module, with some examples. Then we consider the effect of Frobenius structures on the generic radius
of convergence and obtain the fact that a differential module on a disc has
a full basis of horizontal sections (Dworks trick). We also show that the
existence of a Frobenius structure does not depend on the particular choice of
Frobenius lift; this independence plays an important role in rigid cohomology
(Chapter 23).
Throughout Part V, Hypothesis 14.0.1 remains in force unless explicitly
contravened. In particular, K will by default be a discretely valued complete
nonarchimedean field.

17.1 Frobenius structures

We start with the basic compatibility between differential and difference
structures.
Definition 17.1.1. Let R be a ring as in Definition 15.2.1. For M a finite
free differential module over R, a Frobenius structure on M with respect to
a Frobenius lift on R is an isomorphism  : M
= M of differential modules. In more explicit terms, we must equip M with the structure of a dualizable
difference module over (R, ), such that
D((m)) =

d(t)
(D(m))
dt

(m M).
291

292

Frobenius structures on differential modules

In even more explicit terms, if A, N are the matrices of action of , D on some

basis then A is invertible, and we have the compatibility
NA+

dA
d(t)
=
A(N ).
dt
dt

(17.1.1.1)

Remark 17.1.2. We may also speak of Frobenius structures on finite free

differential modules for the derivation td/dt; the analogue of (17.1.1.1) is
NA+t

dA
t d(t)
=
A(N ).
dt
(t) dt

(17.1.2.1)

However, if R is a subring of K t then (17.1.2.1) only makes sense if (t) =

t q u for u R , in which case taking constant terms in (17.1.2.1) yields
1
N0 A0 = qu 0 A0 (N0 ). This gives |N0 |sp = |A1
0 N0 A0 |sp = q |(N0 )|sp =
q 1 |N0 |sp , so N0 must have spectral radius 0 and hence must be nilpotent.
We now describe some examples where Frobenius structures can be constructed explicitly.
Definition 17.1.3. Suppose that K satisfies p1 = p. Then the power
series E(t) = exp(t t p ) (sometimes called the Dwork exponential series)
2
has radius of convergence p ( p1)/ p (exercise), even though the series exp(t)
has radius of convergence 1.
Example 17.1.4 (Dwork). Assume that K contains an element with
p1 = p. Pick any f oE , and let M f be the differential module over
E of rank 1 with a generator v satisfying D(v) = (d f /dt)v. (Note that this
is the pullback along f of Example 9.3.5.) This module is typically nontrivial
because the exponential series exp( f ) does not represent an element of E
(e.g., consider f = t 1 ). However, the exponential series gives an important
clue about how to associate a Frobenius structure with M f ; namely, for an
absolute Frobenius lift we would like to define
(v) = exp( f ( f ))v.
To verify this formula, note that it suffices to do so in the case (t) = t q , since
exp(( f q ( f ))) is already well-defined in E . We can then write
a

exp( f f p ) = E( f )E( f p ) E( f p

a1

),

where E is the Dwork exponential series from Definition 17.1.3.

Example 17.1.5. We can similarly construct a Frobenius structure for Example 9.9.3. (The case h = 0 will reproduce Example 17.1.4; see Remark 17.1.7
below.) For h a nonnegative integer and K a primitive p h+1 th root of

17.1 Frobenius structures

293

unity, let M be the differential module of rank 1 over E with the action of D
on a generator v given by
h

i
i
D(v) =
( p 1)t p 1 v.
i=0

(Note that we have replaced t by t 1 in the formula from Example 9.9.3. ) Let
: E E be an absolute Frobenius lift fixing . On the disc |t 1 | < 1 we
have already constructed the horizontal section
 h

 1 pi
E p (t 1 )
i
v = exp
t p v,
pi
E p ( t 1 )
i=0

where


E p (t) = exp

pi

t
i=0

pi

is the ArtinHasse exponential. This suggests that one should define a

Frobenius action fixing this horizontal section, which would then be given by
the formula
(v) =

E p (t 1 )E p ( t p )
v.
E p (t p )E p ( t 1 )

(17.1.5.1)

In fact, this gives a Frobenius structure defined over E , because the coefficient of v in (17.1.5.1), as a power series in t 1 , has radius of convergence
strictly greater than 1. We will not verify this here; it was shown for p > 2
by Matsuda [169] by an explicit calculation, and for all p by Pulita [185] as
part of a much broader result. We note here that the existence of this Frobenius
structure implies the following strengthening of Theorem 12.7.2.
Theorem 17.1.6. Let b be a positive integer. Let K be a complete nonarchimedean field (not necessarily discretely valued) containing the p h th roots of
unity for all h log p b and having a perfect residue field. Let M denote a
finite differential module of rank 1 on a half-open annulus with open outer
radius 1, which is solvable at 1 with differential slope b. Then there exist
c1 , . . . , cb {0} o
K and nonnegative integers j1 , . . . , jb such that
M M1,c1 Mb,cb
has differential slope 0, for Mi,ci defined as in Theorem 12.7.2.
Proof. Since K is perfect, K contains a subfield K 0 isomorphic to the fraction field of the Witt vectors of K . Let 0 be the Witt vector Frobenius

294

Frobenius structures on differential modules

morphism on K 0 . Let be the substitution t  t p , and set = 0

as an endomorphism of E0 = (0,1) K 0 /t, t0 .
On the one hand, by Example 17.1.5, for any cb o K 0 we have Mb,cb
=

(Mb,cb ) = (Mb,0 (cb ) ). On the other hand, for cb , db o K with |cb

M
db | < 1, Mb,c
b,db is trivial because we can use (9.9.3.1) to write down a
b
horizontal section. Since K is perfect, given cb o K we can choose cb o K 0
with |(cb ) p cb | < 1, and so |0 (cb ) cb | < 1. We thus deduce that
Mb,cb
= (Mb,cb ).
With this fact, we may deduce the claim from Theorem 12.7.2.
Remark 17.1.7. We will have special need later for the case of
Theorem 17.1.6, in which M p has differential slope 0. In this case, we
only deal with objects of the form Mi,ci with i not divisible by p, thanks to
Corollary 12.7.3.
This implies first that we only need K to contain the pth roots of unity, not
the p h th roots of unity, for all h log p b. It also implies that we can use the
Frobenius structure from Example 17.1.4 instead of that from Example 17.1.5.
Namely, for a primitive pth root of unity, there exists a unique Q p ( )
with p1 = p and | ( p 1)| < p 1/( p1) (exercise) and, for this
choice, if ci o K and f = ci t i then M f Mi,ci has differential slope 0 by
an explicit calculation (exercise).
Remark 17.1.8. In addition to the above examples, Dwork managed to construct explicit Frobenius structures in several classical cases, using explicit
formal solutions of the corresponding differential equations; see for instance
Example 20.2.1. These examples are uniformly explained by the fact that
PicardFuchs modules carry Frobenius structures in rather broad generality.
See Chapter 22 for further discussion.

17.2 Frobenius structures and the generic radius

of convergence
One of Dworks early discoveries was that the presence of a Frobenius structure forces solvability at the boundary. (There is also a converse for modules
of rank 1; see the notes.)
Theorem 17.2.1. Let M be a finite differential module on the half-open annulus with closed inner radius and open outer radius 1 equipped with a
Frobenius structure. Then

17.2 Frobenius structures and the generic radius of convergence

295

lim I R(M F ) = 1,

that is, M is solvable at 1. More precisely, for (0, 1) sufficiently close to 1,

I R(M F 1/q ) I R(M F )1/q .
Proof. By imitating the proof of Lemma 10.3.2 (using Lemma 15.2.4), we may
show that, for (0, 1) sufficiently close to 1,
I R(M F 1/q ) min{I R(M F )1/q , q I R(M F )}.
The function f (s) = min{s 1/q , qs} on (0, 1] is strictly increasing, and any
sequence of the form s, f (s), f ( f (s)), . . . converges to 1 (as in the proof of
Theorem 16.1.1). This proves the first claim; for the second claim, note that
f (s) = s 1/q when s is sufficiently close to 1.
The following corollary is sometimes called Dworks trick. It may be viewed
as a true nonarchimedean analogue of the fundamental theorem of ordinary
differential equations.
Corollary 17.2.2 (Dwork). Let M be a finite differential module on the open
unit disc, such that the restriction of M to some half-open annulus with
open outer radius 1 admits a Frobenius structure. Then M admits a basis of
horizontal sections.
Proof. By Theorem 17.2.1, for each < 1 there exists (, 1) such that
R(M F ) > . By Dworks transfer theorem (Theorem 9.6.1), M K t/
admits a basis of horizontal sections. Taking arbitrarily close to 1 yields the
claim.
Remark 17.2.3. The proof of Corollary 17.2.2 admits the following geometric
interpretation. By Proposition 9.3.3 the horizontal sections converge on some
disc of positive radius . Pulling back by a Frobenius lift gives a new space
of horizontal sections on the disc of radius min{ 1/q , q}, but this space must
coincide with the original space. Repeating the construction, we can eventually
stretch the horizontal sections over the entire open unit disc.
One also has a nilpotent analogue of Dworks trick by using Theorem 13.7.1
in place of Theorem 9.6.1.
Corollary 17.2.4. Let M be a finite differential module, on the open unit disc
for the derivation td/dt with a nilpotent singularity at t = 0, such that the
restriction of M to some annulus with open outer radius 1 admits a Frobenius
structure. Then M has radius of convergence 1; that is, for any < 1 and
for any basis of M K t/, the fundamental solution matrix for that basis

296

Frobenius structures on differential modules

has entries in K t/. (Note that the nilpotency of the singularity is automatic
if the Frobenius lift is of the form described in Remark 17.1.2 and if the
Frobenius structure is defined on the entire disc.)
A nice application of Dworks trick is the following.
Proposition 17.2.5. Let M be a finite differential module over K t0 with
R(M) = 1. (For instance, this holds if M admits a Frobenius structure, by
Dworks trick.) Then H 0 (M) = H 0 (M E ).
Proof. By Theorem 9.6.1 there exists a horizontal basis e1 , . . . , en of M K t 0

n
K {t}. If v H 0 (M K t 0 E ) then when we write v = i=1
vi ei with vi R
we must have d(vi ) = 0 for i = 1, . . . , n. This forces vi K for i = 1, . . . , n,
and so
v (M K t 0 E ) (M K t 0 K {t}) = M K t 0 (E K {t}) = M.

17.3 Independence from the Frobenius lift

Another key property of Frobenius structures is that their existence does not
depend on the exact shape of the Frobenius lift.
Proposition 17.3.1. Let 1 , 2 be two Frobenius lifts on R that agree on K .
Let M be a finite free differential module over R equipped with a Frobenius
structure for 1 . Then there is a functorial way to equip M with a Frobenius
structure for 2 .
Proof. The Frobenius structure for 2 is defined by the following Taylor
series:
2 (m) =


(2 (t) 1 (t))i
i=0

i!

1 (D i (m)).

(17.3.1.1)

By Theorem 17.2.1 and the fact that |2 (t) 1 (t)|1 < 1, this series converges
under | | for (0, 1) sufficiently close to 1 (if this is well-defined for R),
and also under | |1 (if this is well-defined for R).
Corollary 17.3.2. Let 1 , 2 be two Frobenius lifts on R that agree on K .
Then there is a canonical equivalence between the categories of finite free
differential modules over R equipped with Frobenius structures with respect
to i for i = 1, 2; this equivalence is the identity functor on the underlying
difference modules.

17.3 Independence from the Frobenius lift

297

Definition 17.3.3. Corollary 17.3.2 allows us to switch from one Frobenius

lift on R to another that may be more convenient. One useful choice is what
we call the standard q-power Frobenius lift for a given choice of K , namely
the Frobenius lift for which (t) = t q .
We will also see that changing Frobenius lifts preserves purity.
Lemma 17.3.4. Let M be a finite free differential module over E equipped
with a unit-root Frobenius structure. Let M0 be a finite free dualizable difference module over oE such that M is isomorphic as a difference module to
M0 oE E . Then, under any such isomorphism, M0 is stable under D i /i! for
each nonnegative integer i.
Proof. We proceed by induction on i. Note that oE is stable under d i /i! for
each i; hence, by the Leibniz rule, given the claim for all j < i it suffices to
exhibit a single basis of M0 on which D i /i! acts via a basis over oE .
Choose a matrix of M0 , and let A, N be the matrices of action of , D
on this basis. If we apply  to this basis, the matrices of action of , D on
the resulting basis are (A), (d(t)/dt)(N ), respectively. By repeating this
construction, for any  < 0 we can find a basis of M0 on which the matrix of
action of D has norm at most .
In particular, we will deduce the base case i = 1. For i > 1, use the previous paragraph to choose a basis e1 , . . . , en of M0 such that (D/i)(e j ) M0
for j = 1, . . . , n and then invoke the induction hypothesis to deduce that
(D i /i!)(e j ) = (D i1 /(i 1)!)((D/i)(e j )) M0 .
Proposition 17.3.5. Suppose that R = E or R = R. Let 1 , 2 be two
Frobenius lifts on R that agree on K . Let M be a finite free differential module
over R equipped with a Frobenius structure for 1 that is pure of some norm.
Then M is also pure of the same norm with respect to the induced Frobenius
structure for 2 .
Proof. It suffices to check the case R = E , because if M is pure over R, and
M is the unique pure module over E with M
= M E R, then  acts on M
by Corollary 16.3.4. We may also reduce to the case where M is pure of norm
1, by replacing 1 , 2 by powers of themselves and invoking Corollary 14.4.5.
By Proposition 14.4.16 we can write M = M0 oE E for some finite free
oE -module M0 such that M0 and M0 are stable under the action of 1 . By
Lemma 17.3.4, M0 is stable under D i /i! for each nonnegative integer i. By
(17.3.1.1) the Frobenius structure with respect to 2 carries M0 into itself, and
similarly for M0 . This yields the claim.

298

Frobenius structures on differential modules

17.4 Slope filtrations and differential structures

In order to apply the slope filtration theorem (Theorem 16.4.1) to differential modules with a Frobenius structure, we must check some compatibilities
between slope filtrations and differential modules.
Lemma 17.4.1. Let M be a finite differential module over R equipped with
a Frobenius structure. Then the steps of the filtration of Theorem 16.4.1 are
differential modules, not just difference modules.
Proof. It suffices to check this for M1 , as then we may quotient by M1 and
D

repeat the argument. Note that the composition M1 M M/M1 is

R-linear, since D(r v) = r D(v) + d(r )v and the second term becomes zero
in the quotient. Hence it suffices to check that M1 M/M1 is the zero map.
However, M1 is pure of some norm that is greater than every norm appearing in the slope filtration of M/M1 . Consequently, Hom(M1 , M/M1 ) = 0 by
Proposition 16.4.4.
Lemma 17.4.2. Let M be a finite free unit-root difference module over E such
that M E R admits a compatible differential structure. Then this structure is
induced by a corresponding differential structure on M itself.
Proof. Let N , A be the matrices via which D and  act on a basis of M.
Write the commutation relation (17.1.2.1) between N and A in the form
N (t/(t))(d(t)/dt)A(N )A1 = (d/dt)(A)A1 . We deduce from
Lemma 16.3.2 that N has entries in E .
From Lemmas 17.4.1 and 17.4.2, we deduce the following refinement of the
slope filtration theorem in the presence of a differential structure.
Theorem 17.4.3. Let M be a differential module over R equipped with a
Frobenius structure. Then there exists a unique filtration 0 = M0
Ml = M, by differential submodules preserved by the Frobenius structure,
with the following properties.
(a) Each successive quotient Mi /Mi1 is finite free and descends uniquely
to a differential module over E with an induced Frobenius structure
that is pure of some norm si (in the sense of Definition 16.3.1).
(b) We have s1 > > sl .

17.5 Extension of Frobenius structures

The following result allows us to extend certain Frobenius actions. An
important application will be to PicardFuchs modules (Chapter 22).

Notes

299

Proposition 17.5.1. Let M be a finite differential module on the open unit disc
for the derivation td/dt with a nilpotent singularity at t = 0. Assume that
either:
(a) the Frobenius lift on K t0 is arbitrary, and M has no singularity at
t = 0; or
(b) the Frobenius lift on K t0 satisfies (t) = t q u for some u
o K t .
Then any Frobenius structure with respect to on the restriction of M to some
annulus with open outer radius 1 is induced by a Frobenius structure on M
itself.
Proof. It suffices to check this for a standard Frobenius lift, as then under either
(a) or (b) we may switch back and forth between the given Frobenius lift and
a standard Frobenius lift using Proposition 17.3.1. Thus we assume hereafter
that (t) = t q .
Let  be a Frobenius structure with respect to on the restriction of M
to some annulus with open outer radius 1. By Corollary 17.2.4, M admits a
basis on which the matrix of action of D is a nilpotent matrix N over K . Let

A = iZ Ai t i be the matrix of action of  on the same basis; then the commutation relation (17.1.2.1) between and D states that N A + td A/dt =
q A K (N ). Consequently, for each i Z we have N Ai + i Ai = q Ai K (N ).
By Lemma 7.3.5 the operator X  N X + i X q X K (N ) on n n matrices
over K has all its eigenvalues equal to i, because N and K (N ) are both nilpotent. Hence Ai = 0 for i  = 0, so A GLn (K ). In particular the Frobenius
structure  can be defined on the entire open unit disc.

Notes
The statement of Theorem 17.1.6 (whose proof includes that of Theorem
12.7.2) is a slight weakening of [171, Corollaire 2.02], with a similar proof;
the technique goes back to Robba [190]. A complete classification of rank 1
solvable modules on an open annulus with outer radius 1 (and unspecified inner radius) has been given by Pulita [185] and can be formulated in
terms of any LubinTate group; using the formal multiplicative group returns
Theorem 17.1.6.
Theorem 17.2.1 admits the following partial converse: if M is a finite free
differential module over R of rank 1 that is solvable at 1 then there exists
Z p such that M V admits a Frobenius structure for some absolute
Frobenius lift. (Here V is defined as in Example 9.5.2.) For M defined over
F1 the field of analytic elements, this was shown by Chiarellotto and Christol

300

Frobenius structures on differential modules

[39]; it is straightforward to extend their argument to E , but not to R. The

general case was shown by Pulita [184]; it constitutes a refinement of our
Theorem 17.1.6.
We cannot resist viewing Dworks trick (Corollary 17.2.2) as an instance of
a general principle articulated beautifully by Coleman [57, III]:
Rigid analysis was created to provide some coherence in an otherwise totally disconnected p-adic realm. Still, it is often left to Frobenius to quell the rebellious outer
provinces.

A direct proof of Corollary 17.2.4, which does not employ the transfer
theorem for a nilpotent regular singularity (Theorem 13.7.1), appeared in [134,
3.6].
The proof of Proposition 17.3.1 was taken from [209, Theorem 3.4.10].
Theorem 17.4.3, in the case of an absolute Frobenius lift, is the original form
of the slope filtration theorem suggested, though not explicitly conjectured,
by Tsuzuki in [209]. Its derivation from Theorem 16.4.1 is the same as the
derivation of [125, Theorem 6.12] from [125, Theorem 6.10]. See also [129,
7.1].

Exercises
(1) Verify the unproved assertion in Definition 17.1.3. (Hint: see [191, 7.2]).
(2) Verify the unproved assertions in Remark 17.1.7.

18
Effective convergence bounds

In this chapter, we discuss some effective bounds on the solutions of p-adic

differential equations with nilpotent singularities. These come in two forms.
We start by discussing bounds that make no reference to a Frobenius structure;
these are due to Christol, Dwork, and Robba. They could have been presented
earlier, and indeed one was invoked in Chapter 13; we chose to postpone them
until this point so that we could better contrast them with the bounds available
in the presence of a Frobenius structure. The latter are original, though strongly
inspired by some recent results of Chiarellotto and Tsuzuki.
These results carry both theoretical and practical interest. Besides their
application in the study of p-adic exponents mentioned above (and in the
proof of the unit-root p-adic local monodromy theorem to follow; see
Theorem 19.3.1), another theoretical point of interest is their use in the study
of the logarithmic growth of horizontal sections at a boundary. We will discuss
some recent advances in this subject due to Andr, Chiarellotto, and Tsuzuki.
(An area of application that we will not discuss is the theory of G-functions,
as found in [80].)
A point of practical interest is that effective convergence bounds are useful
for carrying out rigorous numerical calculations, e.g., in the machine computation of zeta functions of varieties over finite fields. See the notes for Chapter 23
for further discussion.
Hypothesis 18.0.1. In this chapter, we will drop the running restriction that
K is discretely valued, imposing it only when we discuss Frobenius structures.
We retain the condition p > 0, however.

18.1 A first bound

We now go back to the p-adic Fuchs theorem for discs (Theorem 13.2.2) and
extract an effective convergence bound in the case of a nilpotent singularity.
301

302

Effective convergence bounds

One can also give effective bounds for more general regular singularities (as
long as one takes into account the condition of p-adic non-Liouville differences), but the nilpotent case is sufficient for most applications in algebraic
geometry and the bounds are much easier to describe in this case. (The case of
no singularity reproduces the p-adic Cauchy theorem, Proposition 9.3.3.)

Proposition 18.1.1. Let N = i=0

Ni t i be an n n matrix over K t/0
corresponding to the differential system D(v) = N v + d(v), where d = td/dt.
Assume that N0 is nilpotent with nilpotency index m, that is, N0m = 0 but
N0m1  = 0. Assume also that |N | 1. Then the fundamental solution matrix

Ui t i over K t (as in Proposition 7.3.6) satisfies

U = i=0
|Ui | i |i!|2m+1

(i = 1, 2, . . . ).

(18.1.1.1)

Consequently, U has entries in K t/( p (2m1)/( p1) )0 , as does its inverse.
Proof. Recall that U is determined by the recursion (7.3.6.1):
N0 Ui Ui N0 + iUi =

i


N j Ui j

(i > 0).

j=1

By Lemma 7.3.5 the map f (X ) = N0 X X N0 on n n matrices is nilpotent,

with nilpotency index 2m 1. Hence the map X  i X + f (X ) has inverse
X 

2m2


(1) j i j1 f j (X ).

j=0

This gives the claim by induction on i.

18.2 Effective bounds for solvable modules

We now give an improved version of Proposition 18.1.1 under the hypothesis
that U has entries in K {t/0 }. The hypothesis is only qualitative, in that it
implies that |Ui | i 0 as i , but it does not give a specific bound
on |Ui | for any particular i. Somewhat surprisingly, this hypothesis plus an
explicit bound on N together imply a rather strong explicit bound on |Ui |. (We
will continue to restrict to the case of nilpotent singularities; see Section 18.5
for what happens when the exponents are allowed to range over Z p .)

Theorem 18.2.1. Let N = i=0

Ni t i and U = i=0
Ui t i be n n matrices
over K t satisfying the following conditions.
(a) The matrix N has entries in K t/0 .
(b) We have U0 = In .

18.2 Effective bounds for solvable modules

303

(c) We have U 1 N U + U 1 t (d/dt)(U ) = N0 .

(d) The matrix N0 is nilpotent.
(e) The matrices U and U 1 have entries in K {t/}.
Then, for every positive integer i,
|Ui | i p (n1)

log p i"

max{1, |N |n1
}.

The first step in the proof of Theorem 18.2.1 is to change basis to reduce
|N | ; this comes at the expense of enlarging K slightly and decreasing
slightly.
Lemma 18.2.2. With notation as in Theorem 18.2.1, suppose that K has value
group R. Then, for any < 1, there exists an invertible n n matrix X over
K t/()0 such that


 1

 X N X + X 1 t d (X ) 1


dt

|X 1 | 1
|X | |N |n1
.
Proof. Let M be the differential module over K t/0 for the operator td/dt,
with a basis on which D acts via N , and let | | be the supremum norm defined
by this basis. Over the closed disc of radius , M becomes isomorphic to a
successive extension of trivial differential modules. Consequently the generic
radius of convergence of M F is equal to . In particular,
 
d
= .
p 1/( p1) = |t 1 D|sp,MF  
dt F
By Proposition 6.5.6 plus Lemma 8.6.1, we obtain the desired matrix X .
Using Lemma 18.2.2, we wish to prove Theorem 18.2.1 by using Frobenius
antecedents to reduce the index from i to i/ p". One can improve upon this
argument if one has a Frobenius structure on the differential module; see
Lemma 18.3.2 below.
Lemma 18.2.3. With notation as in Theorem 18.2.1, suppose that |N | 1.
Then there exist n n matrices N  , U  over K t/ p , satisfying the hypotheses
of Theorem 18.2.1, such that
|N  | p p,
max{|U j | j : 0 j i} max{|U j | pj : 0 j i/ p}.

304

Effective convergence bounds

i
Proof. Define the invertible n n matrix V =
i=0 Vi t over K t/
as follows. Start with V0 = In . Given V0 , . . . , Vi1 , if i 0 (mod p)

i1
j
1 N W +
then put Vi = 0. Otherwise, put W =
j=0 V j t and N W = W
1
W t (d/dt)(W ), and let Vi be the unique solution of the matrix equation
N0 Vi Vi N0 + i Vi = (N W )i .
By induction on i we have |Vi | i 1 for all i, so V is invertible over K t/0 .
Let : K t K t denote the substitution t  t p . Put N  = V 1 N V +
V 1 t (d/dt)(V ); then N  has entries in K t p  and | 1 (N  )| p 1. Put
U  = V 1 U ; then
(U  )1 N  U  + (U  )1 t

d
(U  ) = N0 = N0 ,
dt

which forces U  also to have entries in K t p . We may then take N  =

p 1 1 (N  ) and U  = 1 (U  ).
We now put everything together.
Proof of Theorem 18.2.1. There is no harm in enlarging K , so we may assume
that K has value group R. We will then prove the claim by induction on
i, in three stages. First, if i < p and |N | 1 then the desired estimate
follows from Proposition 18.1.1. Second, for any given i, the desired estimate for general N follows from the estimate for the same i in the case
|N | 1, by Lemma 18.2.2. (More precisely, for any < 1, replace the
pair N , U by X 1 N X + X 1 t (d/dt)(X ), X 1 U X 0 and then take the limit as
1.) Third, if |N | 1 then the desired estimate for any given i follows
from the corresponding estimate for general N with i replaced by i/ p", by
Lemma 18.2.3.
Example 18.2.4. It is easy to give an example that shows that one cannot
significantly improve the bound of Theorem 18.2.1 without extra hypotheses. (There is a tiny improvement possible; see the notes.) For instance, the
functions
fi =

1
(log(1 + t))i
i!

(i = 0, . . . , n 1)

satisfy the differential system

d
f 0 = 0,
dt

d
1
fi =
f i1
dt
1+t

(i = 1, . . . , n 1),

in which the coefficients have 1-Gauss norm at most 1.

18.2 Effective bounds for solvable modules

305

One important special case of these results is that of a generic disc.

Considering this case will fulfill an earlier promise to prove Lemma 13.5.4.
Corollary 18.2.5. Let V be a finite differential module over F (for the derivation d/dt) for some > 0 such that I R(V ) = 1. Choose a basis of V and, for
i 0, let Di be the matrix of action of D i on this basis. Then
 
 Di  i
  p (n1) log p i" max{1, |D1 |n1 }
(i > 0).

 i! 

Proof. Let L be the completion of K (t ) for the -Gauss norm, so that t L

is a generic point of norm . As in Section 9.4, we may base-extend V to the
open disc of radius with series parameter t t . The fundamental solution
matrix at t can be computed using Remark 5.8.4; it is


(t t)i
Di .
i!
i=0

If we write this matrix as i=0

Ti (t t )i , where Ti has entries in L, we obtain
from Theorem 18.2.1 the bound
|Ti | i p (n1)

log p i"

max{1, |D1 |n1

}.

We now deduce the claim by induction on i. Write Di as a power series

j
j
j=0 Di, j (t t ) whose coefficients have entries in L, so that |Di, j |
|Di | for j 0 with equality for j = 0. (See Section 9.4 to recall where these
inequalities come from.) We then have
Ti =

i


(1) j

j=0

D j,i j
.
j!

By the induction hypothesis, for j < i we have

|D j,i j /j!| i |D j /j!| j
p (n1)

log p j"

max{1, |D1 |n1

}

p (n1)

log p i"

max{1, |D1 |n1

}.

Combined with the bound on Ti , this yields

|Di /i!| i = |Di,0 /i!| i p (n1)
as desired.

log p i"

max{1, |D1 |n1

},

306

Effective convergence bounds

18.3 Better bounds using Frobenius structures

Although Theorem 18.2.1 is close to optimal under its hypotheses, it can be
improved if the differential module in question admits a Frobenius structure.
For simplicity, we restrict to standard Frobenius structures.
Hypothesis 18.3.1. In this section, we restore the hypothesis that K is discretely valued. Fix a power q of p, and let be the standard qth-power
Frobenius lift on K t0 with respect to some isometry K : K K .
The key here is to imitate the proof of Theorem 18.2.1 but with the
differential equation replaced by a certain difference equation.

Ui t i , A = i=0
Ai t i be n n matrices over
Lemma 18.3.2. Let U = i=0
K t satisfying the following conditions.
(a) The matrix A has entries in K t0 .
(b) We have U0 = In , and the matrix A0 is invertible.
(c) We have U 1 A(U ) = A0 .
Then
max{|U j | : 0 j i} |A1
0 ||A|1 max{|U j | : 0 j i/q"}.
Consequently, for every positive integer i,
logq i
.
|Ui | (|A1
0 ||A|1 )

Proof. Note that (c) can be rewritten as

U = A(U )A1
0 .
This gives the first inequality. To deduce the second inequality, we proceed
as in the proof of Theorem 18.2.1 except that we iterate logq i times to
reach the case i = 0 (rather than iterating logq i" times to reach the case
0 < i < p).

Theorem 18.3.3. Let N = i=0

Ni t i , U = i=0
Ui t i , and A = i=0
Ai t i
be n n matrices over K t satisfying the following conditions.
(a) The matrix A has entries in K t0 .
(b) We have U0 = In , and the matrix A0 is invertible.
(c) We have U 1 N U + U 1 t (d/dt)(U ) = N0 .
(d) We have N A + t (d/dt)(A) = q A(N ).
Then U 1 A(U ) = A0 and, for every positive integer i,
logq i
.
|Ui | (|A1
0 ||A|1 )

18.3 Better bounds using Frobenius structures

307

Proof. As noted in Remark 17.1.2, the commutation relation (d) implies that
N A = q A0 (N0 ), which forces N0 to be nilpotent. Put B = U 1 A(U ) =

0 0
i
i=0 Bi t . Then B0 = A0 and N0 B + t (d/dt)(B) = q B(N0 ). Hence
N0 Bi + i Bi = q Bi (N0 ) = Bi A1
0 N0 A0
or
1
1
N0 (Bi A1
0 ) + i(Bi A0 ) = (Bi A0 )N0 .

(18.3.3.1)

By Lemma 7.3.5 the operator X  N0 X X N0 + i X on n n matrices

is invertible for i  = 0, so (18.3.3.1) implies that Bi = 0 for i > 0. Hence
U 1 A(U ) = A0 does hold, so we may conclude by applying Lemma 18.3.2,
to reduce to the case i < q, Theorem 18.2.1.
Remark 18.3.4. By combining Theorem 18.3.3 with Theorem 18.2.1 (applying the latter for i < q), we can obtain the bound
|Ui | |N |n1
p (n1)
1

log p i(log p q) logq i""

(| A1
0 ||A|1 )

logq i"

Remark 18.3.5. In applications to PicardFuchs modules, the difference

between the bounds given by Theorems 18.2.1 and 18.3.3 can be quite significant. For instance, given a PicardFuchs module arising from a family
of curves of genus g, the bound of Theorem 18.2.1 contains the factor
p (2g1) log p i" but the bound of Theorem 18.3.3 replaces the factor 2g 1
by 1. In general, it should be possible to use Theorem 18.3.3 (and perhaps also
Theorem 18.3.6) to explain various instances in which a calculation of n terms
of a power series involves a precision loss p O(log(n)) even though the accumulated factors p by which one divides throughout the calculation amount to
p O(n) . (A typical example of this is given in [124, Lemma 3].)
We record also a sharper form of Theorem 18.3.3 for use in the discussion
of logarithmic growth in the next section.
Theorem 18.3.6. Let v be a column vector of length n over K t, let A =

i
i=0 Ai t be an n n matrix over K t, and let K be chosen to satisfy
the following conditions.
(a) The matrix A has entries in K t0 .
(b) The matrix A0 is invertible.
(c) We have A(v) = v.
Then
max{|v j | : 0 j i} |1 ||A|1 max{|v j | : 0 j i/q"}.

308

Effective convergence bounds

Consequently, for every positive integer i,

|vi | |v0 |(|1 ||A|1 )logq i .
Proof. Rewrite (c) as v = 1 A (v) and proceed as in Lemma 18.3.2.

18.4 Logarithmic growth

In general, the fundamental solution matrix of a differential system with a
nilpotent regular singularity at 0 need not be bounded on a closed disc, even if
the matrix defining the system is bounded and the solution matrix converges on
the open disc. However, one gets a fairly mild growth condition at the boundary; better still, one can extract some interesting information by distinguishing
different solutions on the basis of their order of growth. This is loosely inspired
by a comparable archimedean situation, for which see the notes.
Definition 18.4.1. For 0, let K t be the subset of K t consisting of

f i t i for which
those f = i=0

| fi |
< +;
| f | = sup
(i + 1)
i
note that K t is complete under the norm | | . For = 0 we recover the ring
K t0 of bounded power series. However, K t is not a ring for > 0; rather,
we have
K t1 K t2 K t1 +2 .
Also, K t is stable under d/dt but antidifferentiation carries it into
K t+1 . Put
)
K t+ =
K t  .
 >

For another useful characterization of K t , see the exercises.

Definition 18.4.2. For f K t, we say that f has order of log-growth if
/ K t  for any  < . We say f has order of log-growth
f K t but f
+ if f
/ K t but f K t  for any  > . We have similar definitions
for vectors or matrices over K t and for elements of M K t 0 K t if M is
a finite free module over K t0 (by computing in terms of a basis, the choice
of which will not affect the answer).
We then deduce the following from Theorem 18.2.1.

18.4 Logarithmic growth

309

Proposition 18.4.3. Let M be a differential module of rank n over K t0 for

the operator td/dt, such that the action of D on M/t M is nilpotent. Then any
element of H 0 (M K t 0 K t) has order of log-growth at most n 1.
Remark 18.4.4. One can say something slightly stronger than
Proposition 18.4.3, by observing that M K t 0 K t[log t] admits a basis of
horizontal sections, each of which has degree at most n 1 in log t (exercise). If we write some m H 0 (M K t 0 K t[log t]) as a formal sum

n1
i
i=0 m i (log t) , with m i M K t 0 K t, then Theorem 18.2.1 implies
that, for i = 0, . . . , n 1, m i has order of log-growth at most n 1. However,
we suspect that this can be improved to n 1 i.
In the presence of a Frobenius structure, one obtains a much sharper bound,
due to Chiarellotto and Tsuzuki [41, Theorem 6.17]. (One can also formulate
a refinement in the manner of Remark 18.4.4.)
Theorem 18.4.5. Assume that K is discretely valued. Let M be a differential
module of rank n over K t0 for the operator td/dt, equipped with a Frobenius
structure for a qth-power Frobenius lift as in Remark 17.1.2. Then any element
v H 0 (M K t 0 K t) satisfying (v) = v for some K has order
of log-growth at most ( log || s0 )/(log q), where s0 is the least generic
Newton slope of M.
Proof. We may assume that the Frobenius lift is standard, thanks to
Proposition 17.3.1. By then replacing the Frobenius lift by some power, we can
reduce to the case where s0 is a multiple of log p. We can then twist into the
case s0 = 0. By Proposition 14.5.9, we can choose a basis of M such that the
least generic Hodge slope of M is also 0. Then the claim follows immediately
from Theorem 18.3.6.
Remark 18.4.6. Refining a conjecture of Dwork, it was conjectured
by Chiarellotto and Tsuzuki [41] that if M is indecomposable then
Theorem 18.4.5 is optimal. That is, in the notation of Theorem 18.4.5, v should
have order of log-growth exactly ( log || s0 )/(log q); Chiarellotto and
Tsuzuki have proved this for rank(M) 2 [41, Theorem 7.2]. It should be
possible to extend their proof to all cases where log || is less than or equal
to the least Newton slope of M strictly greater than s0 , but it is less clear how
to extend to the general case.
Remark 18.4.7. By contrast, if M does not carry a Frobenius structure then
the order of log-growth of a horizontal section behaves much less predictably.

310

Effective convergence bounds

For instance, it need not be rational, and it could have the form + instead of
[41, 5.2].

18.5 Nonzero exponents

So far, we have considered only regular differential systems with all exponents
equal to zero. We now allow for arbitrary exponents in Z p ; the bound we get
is very slightly weaker in this case.

Ni t i and U = i=0
Ui t i be n n matrices
Theorem 18.5.1. Let N = i=0
over K t satisfying the following conditions for some > 0.
(a) The matrix N has entries in K t/0 .
(b) We have U0 = In .
(c) We have U 1 N U + U 1 t (d/dt)(U ) = N0 .
(d) The matrix N0 has prepared eigenvalues in Z p .
(e) The matrices U and U 1 have entries in K {t/}.
Then, for every positive integer i,
|Ui | i p n+(n1)log p i max{1, |N |n1
}.
Proof. The statement of Lemma 18.2.2 carries over to this situation without change. However, we can carry out the construction in the proof of
Lemma 18.2.3 only if the exponents are divisible by p. Otherwise we must
first enforce this condition by performing up to p 1 shearing transformations,
yielding the weaker bound
max{|U j | j : 0 j i} max{|U j | pj : 0 j (i + p 1)/ p}.
In the case i p (and |N | = 1), we instead perform the shearing transformations and then calculate explicitly as in Proposition 18.1.1 to obtain the
bound
|Ui | i p 2n1 .
If i p h+1 for some h 0 then, after h iterations of the map i  (i + p
1)/ p", we end up with a quantity that is at most p. We thus obtain the claimed
bound.

Notes
In the case of no singularities (N0 = 0), the effective bound of Theorem 18.2.1
is due to Dwork and Robba [82], but is slightly stronger: one may replace
p (n1) log p i" with the maximum of | j1 jn1 |1 over j1 , . . . , jn1 Z with
1 j1 < < jn1 i. See also [80, Theorem IV.3.1].

Exercises

311

The general case of Theorem 18.2.1 is due to Christol and Dwork [47],
except that their bound is significantly weaker: it is roughly p c(n1) log p i"
with c = 2 + 1/( p 1). The discrepancy comes from the fact that the role
of Proposition 6.5.6 is played in [47] by an effective version of the cyclic vector theorem, which does not give optimal bounds. As usual, the use of cyclic
vectors also introduces singularities, which must then be removed, leading to
some technical difficulties. See also [80, Theorem V.2.1].
Theorems 18.3.3 and 18.3.6 are original, but they owe a great debt to the
proof of [41, Theorem 7.2]. The main difference is that we prefer to argue in
terms of matrices rather than cyclic vectors.
In the case of no singularities, Proposition 18.4.3 was first proved by Dwork;
it appears in [75] and [76]. (See also [42].) The extension suggested by
Remark 18.4.4 is original; as noted above, the effective bounds in [47] are
not strong enough to imply this.
The archimedean motivation for the study of logarithmic growth comes from
Delignes study of regular singularities [68]. He showed that a differential
module over the ring of germs of meromorphic functions at a point has a regular singularity if and only if the horizontal sections have at worst logarithmic
growth at the singular point.
In the p-adic setting, the theory of logarithmic growth emerged from some
close analysis made by Dwork [75, 76] of the finer convergence behavior
of solutions of certain p-adic differential equations. The subject languished
until the recent work of Chiarellotto and Tsuzuki [41]; inspired by this, Andr
[6] proved a conjecture of Dwork [76, Conjecture 2] analogizing the semicontinuity theorem for Newton polygons (Theorem 15.3.2) to logarithmic
growth.
Theorem 18.5.1 is an improvement on [80, Theorem V.9.1], in which the
2
bound takes the form p (n +cn) log p i" for some constant c. Again, part (though
not all) of the improvement is due to the avoidance of cyclic vectors.

Exercises
(1) Prove that, for 0,
K t = { f K t : lim sup
1

| f |
< }.
( log )

(Hint: the inequality

i

sup{(i + 1) }
i

may be helpful.)

( log )
e

312

Effective convergence bounds

(2) Let M be a differential module of rank n over K t for the operator
td/dt, such that the action of D on M/t M is nilpotent. Show that if we
extend the action of td/dt to K t[log t], by setting (td/dt)(log t) = 1,
then M K t  K t[log t] admits a basis of horizontal sections, and each
horizontal section has degree in log t bounded by n 1.

19
Galois representations and
differential modules

In this chapter we construct a class of examples of differential modules on open

annuli which carry Frobenius structures and hence are solvable at a boundary. These modules are derived from continuous linear representations of the
absolute Galois group of a positive-characteristic local field.
We first construct a correspondence between Galois representations and differential modules over E carrying a unit-root Frobenius structure. The basic
mechanism for producing these modules is to tensor with a large ring carrying
a Galois action and then take Galois invariants. This mechanism will reappear
when we turn to p-adic Hodge theory, at which point we will attempt to simulate this situation using the Galois group of a mixed-characteristic local field.
See Chapter 24.
Then we refine the construction to compare Galois representations having
finite image of inertia with differential modules over E carrying a unit-root
Frobenius structure; the main result here is an equivalence of categories due to
Tsuzuki. It is generalized by the absolute case of the p-adic local monodromy
theorem (Theorem 20.1.4 below) and indeed can be used together with the
slope filtration theorem (Theorem 17.4.3) to prove the monodromy theorem in
the absolute case. This result also has an analogue in p-adic Hodge theory; see
Theorem 24.2.5.
We finally describe (without proof) a numerical relationship between the
wild ramification of a Galois representation and the convergence of solutions of p-adic differential equations. Besides making explicit the analogy
between the wild ramification of Galois representations and the irregularity
of meromorphic differential systems, it also suggests an approach to higherdimensional ramification theory. We reserve most discussion of the latter to
the notes.
313

314

Galois representations and differential modules

Remark 19.0.1. The reader encountering this material for the first time is
strongly encouraged to assume that K is perfect. For an explanation of how
things get complicated otherwise, see Section 20.5.
Notation 19.0.2. In this chapter we write d and instead of D and  for
the actions on a differential or difference module. This is to avoid confusion with another standard usage of the letter D, which will first appear in
Definition 19.1.2.

19.1 Representations and differential modules

We first describe a simple correspondence between Galois representations and
differential modules due to Fontaine. This will serve as a model later, when
we introduce (, )-modules associated with Galois representations of local
fields in mixed characteristic (Chapter 24).
Definition 19.1.1. For L a finite separable extension of K ((t)), let E L be
the finite unramified extension of E with residue field L (see Corollary 3.2.4).
Let E be the completion of the maximal unramified extension of E, which in
particular contains the completion K of the maximal unramified extension of
K . Then G K ((t)) acts on E with fixed field E.
Definition 19.1.2. Let V be a finite-dimensional vector space over K , and
let : G K ((t)) GL(V ) be a continuous homomorphism for the p-adic
topology on GL(V ). Let us view V K E as a left G K ((t)) -module, with the
action on the first factor coming from and the natural action on the second
factor. Put
G K ((t)) .
D(V ) = (V K E)
Lemma 19.1.3. The space D(V ) is an E-vector space of dimension dim K (V ).
Equivalently, the natural map D(V ) E E V K E is an isomorphism.
Proof. We first check this in the case where has finite image; then factors
through G E/ K ((t)) for some finite separable extension E of K ((t)). In this
case the claim is a consequence of Noethers nonabelian version of Hilberts
Theorem 90: for any finite Galois extension E/F of fields, the nonabelian
cohomology set H 1 (G E/F , GLn (E)) is trivial.
In the general case we must argue a bit more carefully. Since G K ((t)) is
compact as a topological group, its image under is also compact. This implies
that we can find an oE -lattice T in V that is stable under the Galois action,
e.g., by starting with any lattice and taking the span of its images under the

19.1 Representations and differential modules

315

Galois action. (The compactness ensures that the resulting oE -submodule of V

is indeed finitely generated.) For any positive integer i the induced topology
on T /miK T is discrete, so the Galois action factors through the Galois group of
a finite separable extension of K ((t)). We may thus argue for each i as above
and then take the inverse limit.
Definition 19.1.4. Note that d/dt extends uniquely to E L , and hence to D(V )
by taking the action on V to be trivial. Since the action of d/dt commutes
with the Galois action, we also obtain an action on D(V ). That is, D(V )
is a differential module over E. By the same token, if we equip V with the
structure of a difference module with respect to K , then D(V ) inherits a
Frobenius structure for any Frobenius lift on E acting on K via K . Since
we can always start with a unit-root Frobenius structure on V (e.g., by forcing
a basis of V to be fixed), we deduce that D(V ) admits a unit-root Frobenius
structure.
We obtain an instance of nonabelian ArtinSchreier theory. See the notes to
Chapter 14 for some background.
Proposition 19.1.5. Suppose that is an absolute qth-power Frobenius lift
and that the fixed field K 0 of on K has residue field Fq and the same
value group as K . Given a continuous representation of G K ((t)) on a finitedimensional K 0 -vector space V0 , equip V = V0 K 0 K with the Frobenius
action induced by the trivial action on V0 . Then V0  D(V ) is an equivalence
of categories between the category of continuous representations of G K ((t))
on finite-dimensional K 0 -vector spaces and the category of finite differential
modules over E equipped with a unit-root Frobenius structure.
Proof. We will show that the reverse equivalence is provided by the functor
V0 on finite differential modules over E equipped with unit-root Frobenius
structures and defined by
=1 .
V0 (D) = (D E E)
Since D(V0 ) E E V0 K 0 E is an isomorphism by Lemma 19.1.3, we may
=1 = V0 . (Here the conditions
naturally identify V0 (D(V0 )) with (V K E)
=1
= K 0 . See the exercises at the end of the
on K 0 are needed to ensure that E
chapter.)
It remains to give a canonical isomorphism of D(V0 (D)) with D; by the
same token, it is sufficient to check that the natural map V0 (D) K 0 E D E
E is a bijection. This is exactly the statement that D E E is a trivial difference
module. This holds by the first part of the proof of Theorem 14.6.3, since the

316

Galois representations and differential modules

residue field of E is separably closed and hence weakly difference-closed for

absolute Frobenius lifts.
Corollary 19.1.6. Suppose that is an absolute qth-power Frobenius lift and
that K = Qq is the unramified extension of Q p with residue field Fq . Then the
functor D, from continuous representations of G Fq ((t)) on finite-dimensional
Qq -vector spaces to finite differential modules over E equipped with unit-root
Frobenius structure, is an equivalence of categories.
Remark 19.1.7. In Proposition 19.1.5 the differential structure is not necessary; see the exercises. What we do need is the symmetry between two actions
that of the Galois group and that of the monoid given by the nonnegaon E,
tive powers of . The proof works because E is large enough to trivialize both
the Galois action (by the HilbertNoether theorem) and the -action (by the
DieudonnManin theorem). Starting from this point, Fontaine realized that
one could set up an analogous situation when G Fq ((t)) is replaced by the Galois
group of a finite extension of Q p ; the result is a central construction in p-adic
Hodge theory. See Chapter 24.
Remark 19.1.8. If is a Frobenius lift that is not absolute, then
Definition 19.1.4 remains valid. The reason is that even though an individual
finite separable extension of K ((t)) may not carry an action of , the separable closure of K ((t)) does carry such an action. However, Proposition 19.1.5
does not remain valid, because the residue field of E need not be weakly
difference-closed.

19.2 Finite representations and overconvergent

differential modules
Following an idea of Crew, we give a refinement of the previous construction
for some special representations.
Definition 19.2.1. Since E is henselian (Lemma 15.1.3), for each finite separable extension L of K ((t)) there exists a unique finite unramified extension
E L with residue field L. In fact

EL
= E E E L ,

and E L is the integral closure of E in E L . In particular, G K ((t)) acts on E L

with fixed field E .

19.2 Finite representations and overconvergent differential modules 317

Definition 19.2.2. Let V be a finite-dimensional vector space over K , and let
: G K ((t)) GL(V ) be a continuous homomorphism for the discrete topology on GL(V ). That is, factors through G L/ K ((t)) for some finite separable
extension L of K ((t)). Let us view V K E L as a G K ((t)) -module in which
the action on the first factor comes from and the action on the second factor
is as above. Put
D (V ) = (V K E L )G K ((t)) .
Lemma 19.2.3. The space D (V ) is an E -vector space of dimension
dim K (V ). Equivalently, the natural map D (V ) E E L V K E L is an
isomorphism. (In particular, D (V ) is canonically independent of the choice
of L.)
Proof. This again follows from the HilbertNoether theorem.
Definition 19.2.4. As in Definition 19.1.4, D (V ) is a differential module over
E admitting a unit-root Frobenius structure for any Frobenius lift on E .
However, now that we have a module over E it makes sense to compute the
subsidiary radii of D (V ) F for (0, 1) sufficiently close to 1. Namely,
realize D (V ) as a differential module over K /t, t0 for some and compute there. Be aware that any two such realizations for a given need only
become isomorphic over K /t, t0 for some [, 1). However, statements
about the germ at 1 of the function  R(D (V ) F ) are unambiguous.
Proposition 19.2.5. The generic radius of convergence of D(V ) is equal to 1.
Consequently (by the continuity of the generic radius of convergence, as in
Theorem 11.3.2(a)), D (V ) is solvable at 1.
Proof. This follows from the existence of a Frobenius structure on D (V ),
using Theorem 17.2.1.
Example 19.2.6. Assume that K contains an element with p1 = p;
then K contains a unique pth root of unity p satisfying 1 p (mod 2 )
(see Remark 17.1.7). Let L = K ((t))[z]/(z p z f ) be an ArtinSchreier
extension, and let V be the Galois representation corresponding to the character of G L/ K ((t)) taking the automorphism z  z + 1 to p . We can then
explicitly describe D (V ): it is the module M f of Example 17.1.4 (exercise).
Similarly, one can realize the construction of Example 9.9.3 as D (V ) for
a certain explicit character of order p h . This observation has been thoroughly
generalized by Pulita [185].

318

Galois representations and differential modules

Remark 19.2.7. Note that the kernel of d on E L is the integral closure K  of K

in E L (exercise). Consequently, the space of horizontal sections of D (V ) E
E L is equal to V K K  . This suggests that we cannot recover the whole of
V from D (V ), at least if we use only the differential structure; instead, we
recover the restriction of V to the inertia subgroup of G K ((t)) , which we can
identify with G sep ((t)) .
K

The previous remark suggests the following construction.

Definition 19.2.8. Let V be a finite-dimensional vector space over K , and
let : G K ((t)) GL(V ) be a continuous homomorphism for the p-adic
topology. We say that has finite local monodromy if the image of the inertia subgroup of G K ((t)) is finite. (That inertia subgroup is isomorphic to
be the ring defined in the same fashion
G sep ((t)) .) In this case, let E sep
K

K ((t))

unr , the completion of the maximal unramified extension of

as E but using K+
K , for the coefficients; let G K ((t)) act on this ring via the quotient by its inertia
subgroup. We can then define

D (V ) = (V K (E sep

K ((t))

)unr )G K ((t))

and this will be a differential module over E of the correct dimension, again
admitting a unit-root Frobenius structure for any Frobenius lift.

19.3 The unit-root p-adic local monodromy theorem

We have the following refinement of Proposition 19.1.5. See the notes for
a detailed attribution. (For an analogous fact in p-adic Hodge theory, see
Theorem 24.2.5.)
Theorem 19.3.1 (Tsuzuki). Suppose that is an absolute qth-power
Frobenius lift, and that the fixed field K 0 of on K has residue field Fq and
the same value group as K . Given a continuous representation of G K ((t))
with finite local monodromy on a finite-dimensional K 0 -vector space V0 , equip
V = V0 K 0 K with the Frobenius action induced by the trivial action on V0 .
Then V0  D (V ) is an equivalence of categories with the category of finite
differential modules over E equipped with unit-root Frobenius structure.
Although it is possible to deduce this theorem from the p-adic local
monodromy theorem (see Remark 20.1.5 below), it is both instructive and historically appropriate to give a proof using the tools we have available at this
point. We proceed to this task now.

19.3 The unit-root p-adic local monodromy theorem

319

Definition 19.3.2. Let M be a finite differential module over E equipped with

a unit-root Frobenius structure. For c (0, 1), we say that M is c-constant if
there exists a basis of M on which  acts via a matrix A with | A In |1 c;
we call such a basis a c-constant basis.
If is absolute, we can see directly that the property of M of being
c-constant is invariant under a change in Frobenius lift (Corollary 17.3.2).
Namely, by Proposition 19.1.5, M is c-constant if and only if it occurs as
D(V ) for some representation : G K ((t)) GL(V ) such that V admits a
supremum norm | | for which | (g)(x) x| c|x| for all g G K ((t)) and
x V.
For arbitrary , we may reach the same conclusion by imitating the proof of
Proposition 17.3.5. See the exercises.
Lemma 19.3.3. Let M be a finite differential module over E admitting a
unit-root Frobenius structure for a standard (but not necessarily absolute)
Frobenius lift. Then there exists a positive integer m coprime to p such that
M E E [t 1/m ] admits a basis on which the matrices A, N of action of
, t D have entries in E [t 1/m ] o K t 1/m , as does A1 . Moreover, if M is
c-constant for some c < 1 then we can ensure that m = 1 and | A In |1 c.
Proof. Since M admits a unit-root Frobenius structure, we can choose a basis
on which the matrix of action A of  belongs to GLn (oE ). By reordering the
basis vectors, we can ensure that the minimum t-adic valuation of the reduction
of Ai j occurs for i = j = 1. By adjoining t 1/(q1) and then rescaling, we can
force this minimum valuation to equal 0. We can then conjugate to ensure that
Ai1 = 0 for i = 2, . . . , n. Proceeding in this manner, we can force A to be
upper triangular and invertible over K t 1/m .
We next put A into the desired form, by repeating the following operation.
(If M is c-constant for some c < 1, we may start the argument here.) Write

A = i Ai t i/m , so that A0 is upper triangular and invertible modulo m K and

Ai vanishes modulo m K for i < 0. Then replace A by U 1 A(U ), with

U = In +

A1
0


Ai t

i/m

i>0

The matrices U then converge to the identity under the (t 1/m , m K )-adic
topology on o K t 1/m .
Finally, we note that the compatibility N q A(N )A1 = t (d/dt)(A)A1
from (17.1.2.1) implies that having A in the desired form forces N to be
in the desired form as well. Namely, the compatibility first implies that N

320

Galois representations and differential modules

modulo q involves only nonnegative powers of t 1/m . We then derive the same
conclusion modulo q 2 , q 3 , and so on.
The crux of the proof is the following lemma.
Lemma 19.3.4. Let M be a finite differential module over E admitting a
unit-root Frobenius structure. Suppose that M is c-constant for some c <
p 1/( p1) . Then M is trivial as a differential module.
Proof. We may assume that the Frobenius lift is standard. By Lemma 19.3.3
we may assume that there is a c-constant basis on which , t D act via matrices
A, N over E o K t 1 . Then A and N together represent a finite differential
module over (0,1) K /t equipped with a unit-root Frobenius structure.
Moreover, the commutation relation N A + t (d/dt)(A) = q A(N ) from
1
(17.1.2.1) and the fact that |A0 | = | A1
0 | = |A|1 = |A |1 = 1 together
force N0 = 0 and |N |1 c; consequently, there exists (0, 1) for which
|N | 1.
We now proceed as in the proof of Lemma 18.2.3. As in that proof, we
construct a matrix V, with entries in E K t 1  and with |V In |1 c and
|V In | 1, for which
A = V 1 A(V ),

N  = V 1 N V + V 1 t

d
(V )
dt

have entries in E K t p . Let : K t 1  K t 1  be the K -linear substitution t 1  t p . Since is standard, commutes with ; consequently,
A = 1 (A ), N  = p 1 1 (N  ) again satisfy the commutation relation
N  A + t (d/dt)(A ) = q A (N  ). Since | A In |1 = |A In |1 c,
we deduce that |N  |1 c; we also have |N  | p = p|N  | p. To choose
u [0, 1] such that p u c1u = 1 or, in other words, u log p + (1 u) log c = 0,
we take
u=

log c
.
log c log p

Using Proposition 8.2.3(b) we obtain |N  | 1 for

log = up log =

p log c
log .
log c log p

Since p log c < log c log p < 0, we have log (1 + )( log ) for
some fixed  > 0.
Consequently, we can replace A, N by another pair for which the desired
result can be derived equivalently, with arbitrarily small. In particular we
can force < p 1/( p1) . Now let M  be the differential module on the disc

19.4 Ramification and differential slopes

321

of radius 1 in the coordinate t 1 with the action of t 1 d/dt 1 = td/dt

given by N . The fact |N | 1 implies by Theorem 6.5.3 that the spectral
radius of d/dt 1 on M  is at most . Hence the generic radius of convergence
of M  at radius 1 is at least p 1/( p1) 1 > 1, so, using Theorem 9.6.1 it
can be proved that the local horizontal sections at infinity converge on a disc
of radius greater than 1 in the parameter t 1 . We may then restrict these to a
basis of horizontal sections of M.
Proof of Theorem 19.3.1. It follows from Proposition 19.1.5 that D is fully
faithful, so the key point is to show that D is essentially surjective. That
is, for any finite differential module M over E equipped with a unit-root
Frobenius structure, there exists a finite separable extension L of K ((t)) such
that M E E L is a trivial differential module (and so corresponds to an unramified representation of G K ((t)) ). Using Proposition 19.1.5, we may choose L
so that M E E L is c-constant for some c < p 1/( p1) . Then Lemma 19.3.4
implies the desired result.

19.4 Ramification and differential slopes

In the equal-characteristic case, we can relate the upper numbering ramification filtration (Definition 3.4.3) of a Galois representation with finite local
monodromy to the generic radius of convergence of an associated differential
module, as follows. We will not give a proof here; see the notes for attributions
and references plus some speculations about a mixed-characteristic analogue.
Theorem 19.4.1. Assume that K is perfect. Let V be a finite-dimensional
vector space over K , and let : G K ((t)) GL(V ) be a continuous homomorphism for the p-adic topology on GL(V ), with finite local monodromy.
Then, for (0, 1) sufficiently close to 1,
R(D (V ) F ) = b ,

b = max{i 1 : G i K ((t))  ker( )}.

Corollary 19.4.2. With the notation of Theorem 19.4.1, let V1 , . . . , Vm be the

constituents of V , and let j : G K ((t)) GL(V j ) be the corresponding homomorphisms for j = 1, . . . , m. For (0, 1) sufficiently close to 1, the multiset
of subsidiary radii of D (V ) F consists of b1 , . . . , bn , where the multiset
{b1 , . . . , bn } consists of max{i 1 : G K ((t)),i  ker( j )} with multiplicity
dim(V j ) for j = 1, . . . , m.
Remark 19.4.3. One interpretation of Theorem 19.4.1 is that the decomposition of V by ramification numbers matches up with the ChristolMebkhout

322

Galois representations and differential modules

decomposition of D (V ) R provided by Theorem 12.6.4. While the latter was inspired by analogues for meromorphic connections in the complex
analytic setting, the analogy with wild ramification was anticipated somewhat
before it was realized in Theorem 19.4.1.
Remark 19.4.4. Using the integrality properties of subsidiary radii
(Theorem 11.3.2(b)), we may deduce that, for (0, 1) sufficiently close to
1, the product of the subsidiary radii is an integral power of ; this amounts to
verifying the HasseArf theorem for V (the integrality of the Artin conductor).
An interesting corollary of Theorem 19.4.1 is the following.
Proposition 19.4.5. Let M be a finite free differential module, on the open
annulus with inner radius and outer radius , satisfying the Robba condition.
Suppose that for some closed interval [ , ] with < < < there
exists a finite tale extension R of K  /t, t/ such that the differential module
M K  /t,t/ R is unipotent (i.e., a successive extension of trivial modules).
Then there exists a positive integer m coprime to p such that the pullback of
M along the map t  t m is unipotent.
Proof. Note that the conclusion may be checked by replacing K by a finite
Galois extension K  ; this follows from the fact that
H 0 (M K K  )G K  /K = H 0 (M).
We may thus enlarge K and then rescale to ensure that 1 (, ). We may
also assume that R is Galois over K  /t, t/ and (possibly after enlarging K
again) geometrically connected. That is, R is connected and remains so after
any further finite enlargement of K .
Put G = R K  /t,t/ F1 . Our hypotheses so far ensure that G is a field
so, by Theorem 1.4.9, G carries a unique multiplicative norm extending | |1 .
Since that norm is computed using Newton polygons, we can replace K by a
finite extension to ensure that the multiplicative value group of G is the same
as that of K .
By Lemma 8.6.1 the norm on G can be defined as the supremum norm for
some basis of R K  /t,t/ K t, t 1 . After moving and closer to 1, we
can define the same norm for a basis e1 , . . . , en of R itself.
Let A be the matrix of the trace pairing of R in terms of this basis, i.e., Ai j is
the trace of multiplication by ei e j as an endomorphism of R over K  /t, t/.
On the one hand, because G has the same multiplicative value group as K , we
must have |det(A)|1 = 1. On the other hand, since R is tale over K  /t, t/,
det(A) must be a unit in R. We conclude that det(A) = ct m for some c o
K

Notes

323

and some m Z. In particular, R induces a finite tale extension R of

K [t, t 1 ].
Let 0 = M0 M1 Ml = M be a filtration of M such that, for
i = 1, . . . , l, Mi /Mi1 K  /t,t/ R is a trivial differential module. View the
K -vector space
%
H 0 ((Mi /Mi1 ) K  /t,t/ R)
i

as a representation of Aut(R/ K [t, t 1 ]). By restricting to the inertia groups at

t = 0 and t = , applying Theorem 19.4.1, and possibly making a different
choice of [ , ] (but still with 1 in its interior), we can construct a subextension R  of R which induces a tamely ramified extension of K [t, t 1 ] such that
M K  /t,t/ R  is unipotent. However, a finite tale extension of K [t, t 1 ]
that is tamely ramified at t = 0 and t = must be contained in an extension of the form K [t 1/m , t 1/m ] for some positive integer m and some finite
separable extension K of K . (This is usually deduced as a consequence of
Grothendiecks theory of the tame quotient of the tale fundamental group of
a scheme; see [99, Expos XIII].)
This proves the claim for the restriction of M to the open annulus with inner
radius and outer radius . The claim for M itself follows by Corollary 13.6.4.

Notes
A more detailed survey of most material in this chapter (excluding the unit-root
p-adic local monodromy theorem) is given in the article [128].
Proposition 19.1.5 was originally formulated by Fontaine [89, A1.2.6], in a
slightly less general form and with no reference to differential modules. Our
presentation more closely follows [207, Theorem 4.1.3]; a related result in the
language of F-isocrystals is the theorem of Crew [61, Theorem 2.1].
For K perfect, the rank 1 case of Theorem 19.3.1 is due to Crew [61,
Theorem 4.12], while the general case is due to Tsuzuki [207, Theorem 5.1.1];
however, both arguments can be extended easily to the general case. An alternate exposition was given by Christol [44] in the case of a standard Frobenius
lift, although without discussion of the fact that standardness is not stable under
the replacement of K ((t)) by a finite separable extension. (In our argument
this is treated by the invariance of the c-constant property under a change in
Frobenius lift, as in Definition 19.3.2. The representation-theoretic approach
we used is that taken in [207].) Yet another exposition may be inferred from
[135, Theorem 4.5.2], where a stronger result is proved. (The stronger result

324

Galois representations and differential modules

is used in the study of semistable reduction for isocrystals; see the notes for
Chapter 23.) All these proofs are similar in form to the proof given here; by
contrast, one may infer a rather different proof by specializing Theorem 20.1.4
below to the unit-root case (Remark 20.1.5).
The fact that one can give a direct quantitative relationship between wild
ramification and the spectral properties of differential modules fits nicely into
a well-developed analogy between the structures of irregular formal meromorphic connections and those of wildly ramified -adic tale sheaves. An early
suggestion along these lines was given by Katz [121] and pursued further by
Terasoma [204]; some further work in this direction is that of Beilinson, Bloch,
and Esnault [13].
Theorem 19.4.1 was originally stated in its present form by Matsuda [169,
Corollary 8.8]; a reformulation in the formalism of Tannakian categories
was given by Andr [5, Complement 7.1.2] as part of his formulation and
proof of the p-adic local monodromy theorem. However, thanks to the p-adic
global index theorem of Christol and Mebkhout [51, Theorem 8.41], [52,
Corollaire 5.012], this could have been deduced from the Grothendieck
OggShafarevich formula for unit-root overconvergent F-isocrystals in rigid
cohomology; such a formula was proved by Tsuzuki [208, Theorem 7.2.2]
(by Brauer induction, as is possible in the -adic case) and Crew [63,
Theorem 5.4] (using the KatzGabber theory of canonical extensions, as
also is possible in the -adic case). For a proof by direct computation and
Brauer induction (not using the ChristolMebkhout index theory), see [128,
Theorem 5.23].
In the case of an imperfect residue field, it was originally suggested by
Matsuda [170] that one should formulate an analogue of Theorem 19.4.1
relating the AbbesSaito conductor (discussed in the notes for Chapter 3) to
a suitable differential analogue. That differential analogue was described by
Kedlaya [133]; a comparison with the AbbesSaito conductor was established
by Chiarellotto and Pulita [40] for one-dimensional representations, and by
Xiao [219] in the general case. This has the side effect of establishing the integrality of the AbbesSaito conductor in the case of equal characteristic, which
is not evident from the original construction.
In mixed characteristic the appropriate analogue of the functor V  D(V )
is provided by the theory of (, )-modules; see Chapter 24. The sort of analogue of Theorem 19.4.1 that should exist in mixed characteristic is distinctly
less clear. Even in the case of a perfect residue field, where one is asking for
a differential interpretation of the usual conductor, only a partial answer is
known, by a result of Marmora [167]: one has a differential interpretation of
the conductor once one passes to a suitably large cyclotomic extension. In

Exercises

325

general, one does at least have integrality of the AbbesSaito conductor when
the residue characteristic is odd; see Xiao [220].

Exercises
(1) Suppose that the fixed field K 0 of on K has residue field Fq and the same
value group as K . Prove that E =1 = K 0 . (Hint: reduce to a corresponding
equality of residue fields.)
(2) Prove that any finite free unit-root difference module over E admits a
unique compatible differential structure.
(3) Prove that the kernel of d/dt on E equals K . In particular, for any finite
separable extension L of K ((t)), prove that the kernel of d on E L is the
integral closure of K in E L .
(4) Write out explicitly the isomorphism D (V )
= M implied by Example 19.2.6. (Hint: this module is generated by (1 + p f )1/ p for any lift
f E of f .)
(5) Prove that, for an arbitrary Frobenius lift, the property whereby a finite
differential module over E or E equipped with a unit-root Frobenius structure is c-constant remains invariant under a change in Frobenius lift. (Hint:
in Lemma 17.3.4, if the matrix of action of D/i has norm at most  then so
does the matrix of action of D i /i!. Apply this observation to (17.3.1.1).)

20
The p-adic local monodromy theorem

We are now ready to state the capstone theorem of this book, the p-adic local
monodromy theorem. This asserts that a finite differential module over an
annulus carrying a Frobenius structure has finite local monodromy, in the
sense that it becomes unipotent after making a suitable finite tale cover of the
annulus. In this chapter, we give the precise statement of the theorem, illustrate it with an example and a couple of basic applications, and discuss some
technical points that arise if the field K has imperfect residue field. We will
postpone discussion of the proof(s) of the theorem to the next chapter.
We will discuss two broad areas of application of the p-adic local monodromy theorem in later chapters. One of these is in the subject of rigid
cohomology, where the theorem plays a role analogous to the -adic local monodromy theorem of Grothendieck in the subject of tale cohomology (hence
the name); see Chapter 23. The other is in p-adic Hodge theory, where the
theorem clarifies the structure of certain p-adic Galois representations; see
Chapter 24.
Hypothesis 20.0.1. Throughout this chapter and the next, fix a homomorphism
: K ((t)) K ((t)) preserving (but not necessarily fixing) K and carrying
t to t q for some power q of p. We will assume that all the Frobenius lifts
considered are in fact lifts of this particular .

20.1 Statement of the theorem

Definition 20.1.1. Let L be a finite separable extension of K ((t)) to which
extends. Put
R L = R E E L .
326

20.1 Statement of the theorem

327

We say that a finite free differential module M over R is quasiconstant if

there exists L such that M R R L is trivial. We say M is quasiunipotent if
it is a successive extension of quasiconstant modules; this holds if and only
if M R R L is unipotent for some L (exercise).
Remark 20.1.2. It is more typical to make Definition 20.1.1 without reference
to , so that L may be any finite separable extension of K ((t)). This corresponds precisely to the case where is an absolute Frobenius morphism, as
then extends to any finite separable extension of K ((t)).
The condition of quasiunipotence implies solvability.
Proposition 20.1.3. Let M be a finite free quasiunipotent differential module
over R. Then M is solvable at 1.
Proof. We may reduce to the case where M is irreducible and hence quasiconstant. We may thus pick an L for which M R R L is trivial. It is then
clear that M R R L is solvable at 1, and one can argue using direct calculation that this implies the same for M. Here, we will give a slightly different
argument.
There is no harm in enlarging K , so we may assume that K is integrally
closed in R L . In this case, H 0 (M R R L ) is a vector space over K equipped
with a linear action of G L/ K ((t)) ; that is, we have a representation of a finite
group on a finite-dimensional vector space over a field of characteristic 0.
(Without the enlargement of K , we might have only a semilinear group action.)
It is a basic fact of the representation theory of finite groups that this representation can also be defined over a subfield of K that is finite over Q; in particular,
we can pick a finite extension K 0 of Q p contained in K and a K 0 -lattice T in
H 0 (M R R L ) that is stable under the Galois action.
Since the desired result makes no reference to , we can choose to be
an absolute Frobenius lift fixing K 0 . We then obtain a Frobenius structure on
M R R L by declaring T to be fixed under the Frobenius action. Since T is stable under G L/ K ((t)) , this action descends to a Frobenius structure on M itself.
We may thus deduce that M is solvable at 1, by applying Theorem 17.2.1. (The
reader may recognize this construction from the proof of Theorem 19.3.1.)
Conversely, the following important theorem asserts that many naturally
occurring solvable differential modules, including PicardFuchs modules, are
quasiunipotent. In the case of an absolute Frobenius lift, the theorem is due
independently to Andr [5], Kedlaya [125], and Mebkhout [171]; the generalization to nonabsolute Frobenius lifts is original to this book. See Chapter 21
for a proof and see the notes for further discussion.

328

The p-adic local monodromy theorem

Theorem 20.1.4 ( p-adic local monodromy theorem). Let M be a finite free

differential module over R admitting a Frobenius structure for some Frobenius
lift. Then M is quasiunipotent.
Remark 20.1.5. In the special case where is absolute and M = M E
R for M a finite differential module over E equipped with a unit-root
Frobenius structure, Theorem 20.1.4 almost recovers Tsuzukis theorem
(Theorem 19.3.1). The missing ingredients are as follows.
(a) We must show that M is quasiconstant, not just quasiunipotent. This
follows from the fact that a unipotent differential module over R
equipped with a unit-root Frobenius structure must be trivial (exercise).
(b) We must show that M E E L is a trivial differential module. This
follows from the triviality of M R R L , using Corollary 16.3.4.

20.2 An example
It may be worth seeing how Theorem 20.1.4 applies in a particular example.
This example, corresponding to a Bessel equation, was originally considered
by Dwork [77]; the analysis given here is due to Tsuzuki [209, Example 6.2.6]
and was cited in the introduction to [125].
Example 20.2.1. Let p be an odd prime, put K = Q p ( ) with p1 = p,
and take to be the absolute p-power Frobenius morphism. Let M be the
differential module of rank 2 over R with the action of D on a basis e1 , e2
given by


0
t 1
N=
.
2 t 2 0
Then M admits a Frobenius structure; this was shown by explicit calculation in
[77] but can also be derived by consideration of a suitable PicardFuchs module. Define the tamely ramified extension L of K ((t)), and the corresponding
extension E L of E , by adjoining to K ((t)) an element u such that 4u 2 = t;
then put
u = 1 +


(1)n
n=1

Define the matrix

(2n)!2
u n K {u}.
(32 )n n!3

u+
!
"
U = ud
(u + ) + 12 u 1 u +
du

u
!
"
ud
(u ) + 12 + u 1 u
du

20.3 Descent of sections

329

and use it to change basis; then the action of d/du on the new basis e+ , e of
M R R L is via the matrix

 1 1
0
2 u + u 2
.
0
12 u 1 u 2 .
That is, M R R L splits into two differential submodules of rank 1. To render these quasiconstant, we must adjoin to L a square root of u (to eliminate
the term 12 u 1 ) and a root of the polynomial z p z = u 1 (which, by
Example 19.2.6, eliminates the terms u 2 ).
By further analysis, carried out in [209, Example 6.2.6], one determines
that in this example the special Newton slopes are 12 log p, 12 log p. By contrast, the generic Newton slopes (obtained by writing M = M E R using
the chosen basis and then extending scalars to E) are 0, log p. This illustrates the conclusion of the semicontinuity theorem for Newton polygons
(Theorem 16.4.6).

20.3 Descent of sections

The property of quasiunipotence is rather handy because it often allows
complicated-looking statements about nontrivial differential modules to be
reduced to statements about trivial differential modules that can be checked
by direct calculation. Over a disc, this can be achieved using Dworks trick
(Corollary 17.2.2); over an annulus, Theorem 20.1.4 often serves as a usable
replacement. Here is a typical example, which builds on techniques of de Jong;
see the notes for further discussion.
We start with a calculation that is in some sense dual to Lemma 16.3.2.
Recall that E and E denote the -perfections of E and E , respectively
(Definition 15.4.1).
Lemma 20.3.1. Let A be an n n matrix over oE , and suppose that v En ,

w (E )n satisfy Av (v) = w. Then v (E )n .

Proof. Exercise.
We next bring to bear de Jongs reverse filtration.
Lemma 20.3.2. Let M be a finite dualizable difference module over E . Let
: M E be a nonzero -equivariant map (i.e., commutes with
using the specified -actions on M and E ). Then 1 (E )/ ker() has

rank 1 and Newton slope 0, and M/ 1 (E ) has all its Newton slopes less
than 0.

330

The p-adic local monodromy theorem

Proof. We may assume that is injective. Let M1 be the first step of the
filtration on M, given by Theorem 15.4.4. Then induces a nonzero element
of H 0 (M1 E E ), which can only exist if M1 is pure of norm 1. Moreover,

by Lemma 20.3.1 (applied after using Proposition 14.4.16 to choose a suitable

basis),
H 0 (M1 ) = H 0 (M1 E E ),

so is induced by a nonzero morphism M1 E . We deduce that 1 (E )

is of rank 1; since is assumed to be injective, M1 must also be of rank 1.
This proves the desired results.
To eliminate the perfection in the previous lemma, we use the projection
E E constructed in Lemma 15.4.5.
Proposition 20.3.3. Let M be a finite dualizable difference module over E .
Let : M E be a nonzero -equivariant map. Then 1 (E )/ ker() has
rank 1 and Newton slope 0, and M/ 1 (E ) has all Newton slopes less than 0.
Proof. Put M  = M E E , and let  : M  E be the composition
1

M  E E E E ,
the last map being multiplication in E . The map  is nonzero because
it restricts to on M. (Note that the multiplication map is injective by
Corollary 15.4.6, but we will use the proof technique of that result here rather
than its statement.)
By Lemma 20.3.2 the desired results will follow from the assertion that the
natural inclusions
1 (0) E E (  )1 (0),
1 (E ) E E (  )1 (E )
are surjective. To check this assertion, choose a basis m 1 , . . . , m n of M. Given

n
vi m i with v E .
v (  )1 (0) (resp. v (  )1 (E )), write v = i=1
We induct on the largest integer j for which v j  = 0, the case where no such j
exists being a trivial base case.
Define the map : E E as in Lemma 15.4.5, and put M = id M :
M  M, so that M =  . Then the quantity
M (v/v j ) =

j

i=1

(vi /v j )m i

20.3 Descent of sections

331

belongs to 1 (0) (resp. to 1 (E )). Moreover the coefficient of m j in

v/v j M (v/v j ) vanishes, so by the induction hypothesis the latter belongs
to 1 (0) E E (resp. to 1 (E ) E E ). Hence v does also.
We now add the p-adic local monodromy theorem in order to split some
exact sequences.
Lemma 20.3.4. Suppose that K is perfect. Let 0 M1 M M2 0
be a short exact sequence of finite differential modules over E equipped with
Frobenius structures. Suppose that every generic Newton slope of M1 is strictly
greater than every generic Newton slope of M2 . Then the exact sequence
splits.
Proof. As in Lemma 5.3.3, we may replace M1 , M2 with M2 M1 , E ; that
is, we may reduce to the case where M2 = E and every generic Newton slope
of M1 is positive. By Theorem 16.4.6 every special Newton slope of M1 is also
positive. By Theorem 20.1.4, we can find a finite Galois extension L of K ((t))
such that M1 E R L is unipotent. Since we have assumed that K is perfect,
R L can itself be written as a Robba ring over some finite extension of K (see
Remark 20.5.3 below). By a direct calculation (exercise), the exact sequence
splits over R L . By Corollary 16.3.3, in the category of difference modules the
map H 1 (M1 E E L ) H 1 (M1 E R L ) is injective, so the original exact
sequence splits in the category of difference modules over E L . That splitting is
unique (as can be seen for the case of E L ), so it descends to a splitting of the
original sequence in the category of difference modules over E .
We now check that this splitting is also a splitting of differential modules.
We start with an injective homomorphism E M of difference modules.
By comparing special slopes, we see that the resulting copy of R in M E
R must be the first step in the slope filtration of M E R, in the sense of
Theorem 17.4.3. In particular, the copy of R in M E R is a differential
submodule, as is its pure descent to E . The latter is just the image of E in M,
so the claim follows.
We now put the preceding lemmas together to get a theorem on the descent
of horizontal sections.
Theorem 20.3.5. Let M be a finite differential module over R = K t0 or
R = E admitting a Frobenius structure. Then, in the category of differential
modules, we have
H 0 (M) = H 0 (M R E).

332

The p-adic local monodromy theorem

Proof. For the case R = K t0 we have H 0 (M) = H 0 (M R E ) by

Proposition 17.2.5, so we may assume that R = E hereafter. We may enlarge
K to force K to be perfect. We first check that any nonzero -invariant
horizontal section v of M E E descends to M. The section v corresponds
to a nonzero -equivariant map : M E. By Proposition 20.3.3,
1 (E )  = 0 and the generic slopes of M / 1 (E ) are all negative. By
Lemma 20.3.4 the exact sequence
0 1 (E )/ ker() M / ker() M / 1 (E ) 0
must split. However, again by Proposition 20.3.3, any complement of
1 (E )/ ker() in M / ker() that is in the category of difference modules
must map to zero under . We conclude that in fact (M ) = E , forcing
v H 0 (M).
To check the original claim (still for R = E ), we may enlarge K to have
an algebraically closed residue field. In this case, Corollary 14.6.4 implies that
H 0 (M E E) is spanned by one-dimensional fixed subspaces for some power
of the Frobenius action. The previous argument shows that any generator of
one of these subspaces belongs to M, proving the claim.

20.4 Local duality

Here is another useful property of quasiunipotent differential modules, which
by Theorem 20.1.4 is present whenever one has a Frobenius structure.
Lemma 20.4.1. Put R = E or R. Let M be a finite free differential module
over R. Then, for i = 0, 1, the natural maps
H i (M) H i (M E E L )G L/ K ((t))
are bijections.
Proof. For i = 0, we clearly have
H 0 (M E E L )G L/ K ((t)) = H 0 ((M E E L )G L/ K ((t)) ) = H 0 (M).
For i = 1, we have injectivity because M is a direct summand of M E E L .
We have surjectivity, because if x M E E L represents a Galois-invariant
class in H 1 (M E E L ) then the average of g(x) over g G L/ K ((t)) is an
element of M representing the same class.
Proposition 20.4.2. Suppose that K is perfect. (This assumption can be
removed; see Section 20.5.) Let M be a finite free quasiunipotent differential

20.5 When the residue field is imperfect

333

module over R. Then the spaces H 0 (M), H 1 (M) are finite-dimensional, and
there is a perfect pairing
H 0 (M) H 1 (M ) H 1 (M M ) H 1 (R)
=K

dt
.
t

In particular, by Theorem 20.1.4 this holds whenever M admits a Frobenius

structure.
Proof. This is straightforward to check when M is unipotent (exercise). In
general, we may choose a finite Galois extension L of K ((t)) such that M R
R L is unipotent. Since we have assumed that K is perfect, R L can itself be
written as a Robba ring over some finite extension of K (see Remark 20.5.3
below), so the desired assertions hold for M R R L . We may then deduce the
desired results using Lemma 20.4.1. (See also [128, Proposition 4.26].)
This leads to the following result of Matsuda [169, Theorem 7.8].
Corollary 20.4.3 (Matsuda). Every indecomposable finite free quasiunipotent
differential module over R has the form M N with M quasiconstant and N
unipotent.
Proof. Exercise.

20.5 When the residue field is imperfect

We mentioned earlier (Remark 19.0.1) that the reader should assume that the
residue field K of K is perfect. We now discuss the possible confusion arising
when this hypothesis is omitted.
Remark 20.5.1. The Cohen structure theorem asserts that if F is a complete
discretely valued field of characteristic p > 0 then there exists an isomorphism
F
= F ((t)). However, this isomorphism is far from unique; it depends not
only on the choice of the series parameter t but also on the choice of embedding of the residue field F into F. This choice is unique when F is perfect
(because in that case one has Teichmller lifts; see the notes for Chapter 14),
but not otherwise.

Example 20.5.2. Suppose that F = F p (x)((t)). Then, for any y = i=1

yi t i ,
there is an embedding F p (x) F given by x  x + y.
Remark 20.5.3. Suppose further that E is a finite separable extension of F. By
the Cohen structure theorem, we can find copies of E and F inside E and F,
respectively, and use these to present E and F as power series fields (possibly

334

The p-adic local monodromy theorem

in different series parameters, if |E |  = |F |). If E is separable over F , we

can choose the copy of E to be the integral closure in E of the chosen copy
of F , by Proposition 3.2.3 (or simply by applying Hensels lemma directly).
If F = K ((t)) and E = L, we can then view R L as a copy of the Robba ring
with coefficients in the unramified extension of K = F with residue field
E . However, this may not be possible if E is not separable over F , as in the
following example.
Example 20.5.4. Suppose again that F = F p (x)((t)), and put
E = F[z]/(z p z xt p ).
F p (x 1/ p ), but there is no pth root of x within E itself. To write
Then E =
E
= E ((t)), we must make a different choice of the copy of E within E,
e.g., F p (zt).
Fortunately, in our setting there is a convenient way to skirt this issue using
the Frobenius map .
Lemma 20.5.5. Suppose that is bijective on K . Then, for any finite Galois
extension E of F = E equipped with an extension of , E is separable
over F .
Proof. Let U and T be the maximal unramified and tamely ramified subextensions, respectively, of E over F. Then the claim holds for the extensions U/F
and T /U (the former by the definition of an unramified extension, the latter
by Proposition 3.3.6). We may thus assume that E is totally wildly ramified
over F.
By Remark 3.3.10 the extension E/F can be written as a tower of Artin
Schreier extensions F = E 0 E 1 El = E. It may happen that
E 1 is not preserved by the action of on E; however, it must be carried to
another Z/ pZ-subextension of E over F, of which there are only finitely many.
Consequently E 1 is preserved by some power of ; similarly, we can choose a
power of preserving E 2 , . . . , El .
Since the desired result is insensitive to the replacement of by a power, we
may thus reduce to considering a single ArtinSchreier extension

P=
ci t i .
E = F[z]/(z p z P),
i

Since E admits an extension of , there must exist a F

p such that P
p
a(P) = y y for some y F. By replacing with a suitable power, we
may reduce to the case a = 1.
Remember that we do not change the extension by replacing P with P +
p
y y for y F. We may thus choose P so that, for any two indices i, j < 0

Notes

335

such that ci , c j  = 0, the ratio i/j is not a power of p. Now let j be the smallest
integer for which c j  = 0. We are done if either j 0, in which case E is
unramified, or j < 0 is not divisible by p, in which case #(|E |/|F |) > 1
and so, by Lemma 3.1.1, E = F .
Suppose that, on the contrary, j is divisible by p. By the choice of P and the
fact that P (P) = y p y for some y F, we must have (c j t j ) = (c j t j )q .
1/q
That is, c j = 1 (c j ) K since K was assumed to be bijective. We can
thus replace P by P + (c j t j )1/ p c j t j to increase j. This process repeats only
finitely many times, after which we may deduce the claim.

Notes
The p-adic local monodromy theorem (Theorem 20.1.4) can be viewed
as an archimedean analogue of the following theorem of Borel [see 194,
Theorem 6.1]. Given a vector bundle with connection over the punctured unit
disc in the complex plane, equipped with a polarized variation of (rational)
Hodge structures, Borels theorem asserts that the monodromy transformation
is forced to be quasiunipotent (i.e., its eigenvalues must be roots of unity). It
appears that the Frobenius structure in the p-adic setting plays the role of the
variation of Hodge structures in the complex analytic realm.
The p-adic local monodromy theorem was originally formulated and proved
only in the absolute case. Then, it is often referred to in the literature as
Crews conjecture because it emerged from the work of Crew [62] on the
finite dimensionality of rigid cohomology with coefficients in an overconvergent F-isocrystal. Crews original conjecture was more limited still, as it
concerned only modules such that the differential and Frobenius structures
were both defined over E ; it was restated in a more geometric form by de
Jong [66]. A closer analysis of Crews conjecture was then given by Tsuzuki
[209], who explained (using Theorem 19.3.1, which he had proved in [207])
how Theorem 20.1.4 in the absolute case would follow from a slope filtration
theorem [209, Theorem 5.2.1].
The original proofs of Theorem 20.1.4 in the absolute case can be
briefly described as follows. The proof of Kedlaya uses slope filtrations
(Theorem 17.4.3) to reduce everything to the unit-root case (Theorem 19.3.1).
Andr and Mebkhout proceed by using properties of solvable differential
modules to reduce everything to the p-adic Fuchs theorem of Christol and
Mebkhout (Theorem 13.6.1). Andrs proof was phrased in the language of
Tannakian categories, whereas Mebkhouts proof was more explicit; the proof
in the relative case that we will give in the next chapter is closely modeled on
Mebkhouts proof.

336

The p-adic local monodromy theorem

For applications to rigid cohomology, as far as we know only the absolute

case of Theorem 20.1.4 is of any relevance. However, the nonabsolute case
occurs in the context of relative p-adic Hodge theory. Namely, Berger and
Colmez [18] used the full strength of Theorem 20.1.4 to prove an analogue of
Fontaines conjecture on the potential semistability of de Rham representations
(Corollary 24.4.5 below) for a family of de Rham representations parametrized
by an affinoid base space.
To our knowledge, no proof of the nonabsolute case of Theorem 20.1.4 has
appeared prior to the one we are about to describe in Chapter 21. The proof
by Kedlaya does not apply because it relies on Theorem 19.3.1, whose given
proof does not extend to the relative case. The use of the nonabsolute case of
Theorem 20.1.4 in [18] provided a strong motivation for working at this level
of generality in the present book.
Lemma 20.3.1 is a mild generalization of [206, Proposition 2.2.2]. A weaker
result in the same spirit appears in the work of Cherbonnier and Colmez [38].
For an absolute Frobenius lift, the case of Theorem 20.3.5 with R = E
was originally conjectured by Tsuzuki [210, Conjecture 2.3.3] and proved by
Kedlaya [126, Theorem 1.1]. However, most of the ideas are already present in
the subcase R = K t0 , which was established by de Jong [65, Theorem 9.1];
the main difference is that de Jong could use Dworks trick (Corollary 17.2.2)
where we had to use the p-adic local monodromy theorem (Theorem 20.1.4).
Since a weak form of Dworks trick applies even without a differential structure (Theorem 16.1.1), it should be possible to extend the case R = K t0 of
Theorem 20.3.5 to difference modules; this was carried out in the absolute case
in [130], but we expect the general case to follow similarly.
The proof of Theorem 20.3.5 given here is in substance the same as the
proof in the absolute case given in [126]. In particular, Proposition 20.3.3
is essentially [126, Proposition 4.2], which in turn is close to de Jongs [65,
Corollary 8.2].
The original application of Theorem 20.3.5 in the case R = K t0 is de
Jongs proof of the equal-characteristic analogue of Tates extension theorem
on p-divisible groups (BarsottiTate groups). Tate proved [203] that for R a
complete discrete valuation ring of characteristic 0 and residual characteristic
p > 0, given any two p-divisible groups over R, any morphism between their
generic fibres extends to a full morphism. Tates proof introduced the seeds that
grew into the subject of p-adic Hodge theory (see Chapter 24); de Jong recognized that an appropriate analogue of Tates method for R of characteristic
p would proceed by means of crystalline Dieudonn theory. In this manner, de
Jong [65, Theorem 1.1] reduced this analogue of Tates theorem to an instance
of Theorem 20.3.5.

Exercises

337

The case R = E of Theorem 20.3.5 has applications in the theory of

overconvergent F-isocrystals (for more on which see Chapter 23). Namely,
it implies (after some work) that, on a smooth variety over a field of characteristic p > 0, the restriction functor from overconvergent F-isocrystals to
convergent F-isocrystals is fully faithful [137, Theorem 4.2.1].
The local duality for quasiunipotent differential modules (Proposition 20.4.2)
is due to Matsuda [169]. See also the treatment in [128, 4].

Exercises
(1) Prove that a finite free differential module M over R is quasiunipotent if
and only if M R R L is unipotent for some L. (Hint: produce a nonzero
quasiconstant submodule of M, e.g., by Galois descent.)
(2) Let 0 M1 M M2 0 be a nonsplit extension of differential
modules over R, where M1 and M2 are both trivial and of rank 1. Prove
that M cannot admit a Frobenius structure that induces unit-root Frobenius
structures on both M1 and M2 . (Hint: use the fact that H 1 (R) = K dt/t.)
(3) Prove Lemma 20.3.1. (Hint: reduce to the case where | A| 1 for some
(0, 1) for which |w| < . Then use |w| to bound the terms of v of
norm greater than some c > 0.)
(4) Complete the proof of Lemma 20.3.4 by proving the following assertion.
Let M be a finite unipotent differential module over R, equipped with a
Frobenius structure whose Newton slopes are all positive. Prove that any
exact sequence 0 M N R 0 in the category of differential
modules with Frobenius structures must split. (Hint: apply Lemma 9.2.3 to
N and then see how the resulting basis behaves under a standard Frobenius
lift.)
(5) In the notation of Theorem 20.3.5, suppose that (t) = t q . Prove that the
equality H 0 (M) = H 0 (M R E) also holds in the category of difference
modules. (Hint: if v H 0 (M) then t 1 D(v) H 0 (M) also.)
(6) Prove Proposition 20.4.2 in the case where M is unipotent. (Hint: use
Lemma 9.2.3.)
(7) Prove Corollary 20.4.3 using Proposition 20.4.2. (Hint: let P be an indecomposable finite free quasiunipotent differential module over R. First,
prove that P is a successive extension of copies of a single irreducible differential module M. Then construct an isomorphism P
= M N , with
N unipotent, by induction on the rank of P.)

21
The p-adic local monodromy theorem: proof

In this chapter we give a proof of the p-adic local monodromy theorem, at

the full level of generality at which we stated it (Theorem 20.1.4). After some
initial reductions, we start with the case of a module of differential slope 0,
i.e., one satisfying the Robba condition. We describe how this case can be
treated using either the p-adic Fuchs theorem for ChristolMebkhout annuli
(Theorem 13.6.1) or the slope filtration theorem of Kedlaya (Theorem 16.4.1).
We then treat the rank 1 case using the classification of rank 1 solvable modules
from Chapter 12. We then show that any module of rank greater than 1 and
prime to p can be made reducible, by comparing the module with its top
exterior power and using properties of refined differential modules. We finally
handle the case of a module M of rank divisible by p by considering M M
instead.
The reader may notice some similarities to the proof of the TurrittinLevelt
Hukuhara decomposition theorem (Theorem 7.5.1). In fact, this theorem is also
known as the p-adic Turrittin theorem for this reason.
Besides the running hypothesis for this part of the book (Hypothesis 14.0.1)
and the hypothesis from the previous chapter (Hypothesis 20.0.1), it will be
convenient to set several more hypotheses.

21.1 Running hypotheses

We are going to make a number of calculations under the same hypotheses.
Rather than repeat the hypotheses each time, we enunciate them once and for
all here. We first explain how to deal with the case where K is imperfect.
Remark 21.1.1. First, we point out that it suffices to prove quasiunipotence,
after replacing K by a complete extension K  to which K extends, either by
338

21.2 Modules of differential slope 0

339

Proposition 6.9.1 or a more elementary argument (see [125, Proposition 6.11]).

In particular we may take K  to be the -perfection of K , on which is
bijective.
Here is a variant of the previous remark.
Remark 21.1.2. One can also proceed without -perfecting K initially but
replacing K by its inverse image under at any time when needed. We end up
with a finite extension L of m ( K )((t)), for some positive integer m, such
that M R R L is quasiunipotent. In particular, H 0 (M R R L ) is a nonzero
space on which acts bijectively. Applying m gives a nonzero element of
H 0 (M R R L  ) for a finite separable extension L  of K ((t)). We may deduce
that M has a nonzero -stable differential submodule that is quasiunipotent;
repeating this argument recovers Theorem 20.1.4 in full.
We are now ready to introduce the new running hypotheses mentioned in the
introduction to this chapter.
Hypothesis 21.1.3. For the rest of the chapter, assume that K is bijective
on K ; by the preceding remarks this is harmless for the purpose of proving Theorem 20.1.4. Let F be a finite Galois (but not necessarily unramified)
extension of K . Put R F = R K F; we will not attempt to extend to R F . Let
M be a finite differential module over R equipped with a Frobenius structure.
Hypothesis 21.1.4. Within each lemma in this chapter, set the notation as follows. Let N be a nonzero differential submodule (that is not a subquotient)
of M R R F for some specified F. We will use L to indicate an initially
unspecified finite separable extension of K ((t)) to which extends; since K
is bijective on K , Lemma 20.5.5 implies that R L may be viewed as a Robba
ring over a finite unramified extension of K . We will use F  to indicate an initially unspecified finite extension of F that is Galois over the integral closure
K  of K in R L . (This may require the identification of a subfield of F larger
than K with an isomorphic subfield of R L .) Write R L ,F  = R L K  F  .
Notation 21.1.5. We write N0 to refer to the component of N of differential
slope 0, as provided by Theorem 12.6.4.

21.2 Modules of differential slope 0

We start by describing objects of differential slope 0. This requires the use
of either of two pieces of heavy machinery: the theory of p-adic exponents
(Chapter 13) or the theory of slope filtrations for difference modules over the
Robba ring (Theorem 17.4.3).

340

The p-adic local monodromy theorem: proof

Lemma 21.2.1. Suppose that N has differential slope 0. Then we may choose
L such that N R F R L ,F is unipotent.
Before giving either proof, we insert a reduction common to both.
Remark 21.2.2. In Lemma 21.2.1 the existence of N forces M to have a nontrivial summand M0 of differential slope 0, such that N appears in M R R F .
It thus suffices to prove that if M is of differential slope 0 then we may choose
L such that M R R L is unipotent.
We first give the proof using p-adic exponents.
First proof of Lemma 21.2.1. As in Remark 21.2.2, we may assume that M
itself is of differential slope 0. By Corollary 17.3.2 we may change to a
standard Frobenius lift. We may then apply Corollary 13.6.2 to deduce that
M R R[t 1/m ] is unipotent for some positive integer m coprime to p.
We next give the proof using slope filtrations.
Second proof of Lemma 21.2.1. Again as in Remark 21.2.2, we may assume
that M itself is of differential slope 0. By Theorem 17.4.3 (and changing the
Frobenius lift, as in Corollary 17.3.2), we may reduce to the case where M is
pure of norm 1 as a difference module.
In other words, Lemma 21.2.1 is reduced to the following claim. Let M be
a finite differential module over E admitting a unit-root Frobenius structure
for a standard Frobenius lift. Suppose that I R(M F ) = 1 for (0, 1)
sufficiently close to 1. Then there exists a positive integer m coprime to p such
that M R R[t 1/m ] is unipotent.
To prove this claim, apply Lemma 19.3.3 to construct a basis of M R
E [t 1/m ] for some m, on which , t D act via matrices A, N such that A, A1 ,
and N have entries in E o K t 1/m . As in the proof of Lemma 19.3.4, this
forces N0 = 0.
For notational simplicity, assume hereafter that m = 1. Let f : K t 1 
K t denote the substitution t 1  t. We may then view f (N ) as defining
a differential module M  over K t/ for the operator d/dt for some > 1
such that I R(M  K t/ F ) = 1. By Theorem 9.6.1 this module is trivial on
the open disc of radius ; this implies that the original module M is trivial, and
in particular unipotent.
Remark 21.2.3. It would be interesting to know whether one can prove
Lemma 21.2.1 without using either p-adic exponents or slope filtrations but
instead simply using the fact that the hypothesis forces M to extend across

21.3 Modules of rank 1

341

the entire punctured open unit disc (by the pasting together of Frobenius
antecedents).

21.3 Modules of rank 1

We next consider the rank-1 case, using the classification of solvable rank 1
differential modules (Theorem 12.7.2). This proof is the only point in the
course of the proof of Theorem 20.1.4 at which we will introduce any specific wildly ramified extensions of K ((t)). (We made some tamely ramified
extensions in Lemma 21.2.1 and will make some unramified extensions in
Lemma 21.4.1 below.)
Lemma 21.3.1. Suppose that rank(N ) = 1. Then we may choose L such that
N R F R L ,F is trivial.
Proof. We may assume that F contains all pth roots of 1. By Theorem 17.2.1,
M is solvable at 1 as then is M R R F and also N . By Theorem 12.7.2 there
h
exists a nonnegative integer h such that N p has differential slope 0. If h = 0,
we may deduce the claim from Lemma 21.2.1. It suffices to check the case
h = 1, as we may apply this case repeatedly to deduce the general case.
In the case h = 1, by Theorem 17.1.6 (or the special case of the latter
discussed in Remark 17.1.7), there exist c1 , . . . , cb {0} o
F , with ci = 0
whenever i is divisible by p, such that the rank 1 differential module N1 =
M1,c1 Mb,cb over R F has the property that N1 N has differential
slope 0. Since N1 R F R L 0 ,F is unipotent for L 0 equal to the ArtinSchreier
extension L 0 of F ((t)) defined by the parameter c1 t 1 + + cb t b , so is
N R F R L 0 ,F .
Unfortunately, this construction does not guarantee that L 0 admits an action
of . However, suppose that we run over all possible choices of N for the given
M and F and define each L 0 we see, using the unique possible ArtinSchreier
parameter in the additive group
% alg
K t i .
i<0, p i

The resulting parameters then fill out a finite set S.

Pick an integer i with i < 0, p  i. Let Si be the set of coefficients of
t i appearing in the elements of S. Let P i (x) K [x] be the product of the
distinct minimal polynomials of the elements of Si over K . Since M admits
a Frobenius structure, the roots of (P i (x)), when multiplied by t qi , must
define ArtinSchreier parameters that are equivalent to the roots of P i (x) when
multiplied by t i . In other words,

342

The p-adic local monodromy theorem: proof

(P i (x)) = P i (x 1/q )q .

We may thus extend to the splitting field of P i in such a way that, for each
root a of P i , we have (a) = bq for some root b of P i .
This allows us to define a compositum of ArtinSchreier extensions L to
which extends; this includes all the choices of L 0 made above. For such
an L, N R F R L ,F has differential slope 0, so we can make it trivial by
Lemma 21.2.1. This gives the desired result.

21.4 Modules of rank prime to p

We next pass from rank 1 to rank prime to p, using what we know about refined
differential modules over F .
Lemma 21.4.1. Suppose that rank(N ) is coprime to p. Then we may choose
L , F  such that N R F R L ,F  is either unipotent or reducible.
Proof. Put n = rank(N ). The case n = 1 is covered by Lemma 21.3.1, so
we may assume that n is greater than 1 and coprime to p. Suppose by way of
contradiction that N R F R L ,F  is irreducible and nontrivial for all L , F  . By
Lemmas 21.2.1 and 21.3.1, we may reduce to the case where (M M)0 is
unipotent and n N is trivial.
Let us realize M, M0 , N as differential modules on some annulus < |t| <
1. For (, 1), let F = F/t, t/an , F = F  /t, t/an denote the
fields of analytic elements on the circle of radius over the respective base
fields F, F  . We claim that N F must be refined for every (, 1).
Otherwise, by Theorem 10.6.7, for a suitable choice of F  and some positive
integer m coprime to p we could split N nontrivially over F (t 1/m ). Each
projector for this splitting would define a horizontal section of (N N )
F (t 1/m ). Since (M M)0 is unipotent, so is its subquotient (N N )0 ;
consequently we would have
H 0 (N N F (t 1/m )) = H 0 (N N )
in terms of a basis of (N N )0 , as in Lemma 9.2.3. We would thus get a
nontrivial splitting of N itself, contrary to hypothesis.
We now have that N is refined for every (, 1). Choose (, 1) such
h
h+1 /( p1)
that p p /( p1) < I R(N F ) < p p
for some nonnegative integer
h. Apply Theorem 10.4.2 to form the h-fold Frobenius antecedent P of N F ,
which is still refined. We may then apply Proposition 6.8.4(a) to deduce that
P n has the same spectral radius as P and Proposition 6.8.4(b) to deduce that
(n P ) P n has a strictly greater spectral radius than P. However, these two

Notes

343

contradict each other because n P is the h-fold Frobenius antecedent of the

trivial module (n N ) F . This contradiction yields the desired result.

21.5 The general case

We now make the step from rank prime to p to arbitrary rank. The trick used
here is one familiar from elementary group theory, e.g., it is used in the proof
of the Sylow theorems. See the exercises for another example of its use.
Lemma 21.5.1. For arbitrary N , we may choose L , F  such that N R F R L ,F 
is either unipotent or reducible.
Proof. By Lemma 21.4.1 it suffices to consider N of rank n divisible by p.
Then the trace-0 component of N N has rank n 2 1, which is not divisible by p. By repeated application of Lemma 21.4.1, we can force the trace-0
component of (N N ) R F R L ,F  to acquire a unipotent component. In particular, the space V = H 0 ((N N ) R F R L ,F  ) has F  -dimension greater
than 1.
We may view V as a finite-dimensional, not necessarily commutative, F  algebra. A standard fact about such algebras (exercise) is that, for some finite
extension F  of F  , V F  F  fails to be a division algebra. Thus for suitable
F  we can find a pair of nonzero horizontal that endomorphisms of N R L ,F 
which compose to zero. This forces N to be reducible.
Proof of Theorem 20.1.4. It suffices to consider M as irreducible. By
Lemma 21.5.1, we may choose L , F  such that M R R L ,F  is either unipotent
or reducible. In the former case we are done, as this implies that M R R L is
also unipotent. Otherwise, M R R L ,F  contains a proper nonzero differential
submodule N . By applying Lemma 21.5.1 repeatedly, we may keep replacing
N by a submodule (after changing L and F  ) until N becomes unipotent.
Apply  to (M R R L ) K  F  on the left and Gal(F  /K  ) on the right. The
images of N fill out a submodule of the form N0 K  F  , where N0 is a nonzero
differential submodule of M stable under . The module N0 is unipotent
because N0 K  F  is unipotent. Consequently M R R L has a nonzero unipotent submodule N0 stable under , so we may apply the induction hypothesis
to (M R R L )/N0 to conclude.

Notes
The approach to Theorem 20.1.4 presented here is modeled on that given by
Mebkhout [171] in the absolute case, except that Mebkhout describes only

344

The p-adic local monodromy theorem: proof

the first of our two proofs of Lemma 21.2.1. The approach of Andr [5] is
substantively similar but formally different, as it is phrased in the language of
Tannakian categories.
In the course of proving Lemma 21.3.1, we noticed that if is not an
absolute Frobenius lift, then the condition of admitting an extension of is
a highly nontrivial restriction on finite separable extensions of K ((t)). This
phenomenon was noted in the context of p-adic Hodge theory by Berger and
Colmez [18, Proposition 6.2.2].

Exercises
(1) Let G be a finite p-group and let : G GL(V ) be a complex linear
representation of G. Prove that if is nontrivial then either or
has a nontrivial one-dimensional subrepresentation. (Hint: consider the
possible dimensions of irreducible representations of G.)
(2) Let V be a finite-dimensional, not necessarily commutative, F-algebra.
Prove that there exists a finite extension F  of F such that V F  F is not a
division algebra. (Hint: we may assume that V itself is a division algebra.
Pick any nonzero x V , view x as a linear transformation from V to V
via left multiplication, and subtract an eigenvalue of this transformation.)

Part VI
Areas of Application

22
PicardFuchs modules

In this final part of the book, we touch briefly on some areas of application of
the theory of p-adic differential equations. These chapters are intended more to
inspire than to inform, with statements that are more illustrative than definitive.
In the present chapter we revisit the territory of Chapter 0, briefly discussing
how PicardFuchs modules give rise to differential equations with Frobenius
structures and what this has to do with zeta functions.

22.1 Origin of PicardFuchs modules

The original source for p-adic differential equations, which inspired the
general theory, was the following construction.
Definition 22.1.1. Let K be a field of characteristic 0. Let t be a coordinate
on the projective line P1K . Let f : X P1K be a proper, flat, generically
smooth morphism of algebraic varieties. Let S P1K be a zero-dimensional
subscheme containing (for convenience) and all points over which f is
not smooth. The PicardFuchs modules on P1K \ S associated with f are certain finite locally free differential modules Mi for i = 0, . . . , 2 dim( f ) over
(P1K \ S, O) with respect to the derivation d/dt; they also have regular singularities with rational exponents at each point of S. For
/ S, the fibre of
Mi at can be canonically identified with the ith de Rham cohomology of the
fibre f 1 ().
Although the classical construction of the PicardFuchs module is analytic
(it involves viewing f as an analytically locally trivial fibration and integrating
differentials against moving homology classes), there is an algebraic construction, due to Katz and Oda [123], involving a Leray spectral sequence for the
algebraic de Rham cohomology of the total space.
347

348

PicardFuchs modules

Example 22.1.2. Perhaps the most fundamental example of a PicardFuchs

module comes from the Legendre family of elliptic curves
y 2 = x(x 1)(x t),
with S = {0, 1, }. In this case, the PicardFuchs module is precisely the
differential module derived from the hypergeometric differential equation
y  +

c (a + b + 1)z 
ab
y
y = 0,
z(1 z)
z(1 z)

with parameters (a, b, c) = (1/2, 1/2, 1), as considered in Chapter 0. See

[213, 7] for an algebraic derivation of this fact, in the manner of Katz
and Oda.

22.2 Frobenius structures on PicardFuchs modules

The example considered in Chapter 0 was just one of a number of explicit
examples studied by Dwork and others, in which there seemed to be some
strong relationship between the PicardFuchs equations derived in characteristic 0 and the zeta functions observed over finite fields. Dwork was able to give
a somewhat systematic explanation, in some cases, in terms of Frobenius structures; nowadays, the technology of p-adic cohomology (about which more in
Chapter 23) can be used to give a fairly general explanation.
We first give an explicit statement to the effect that PicardFuchs modules
always carry Frobenius structures. See the notes for a more detailed discussion.
Theorem 22.2.1. With notation as above, assume that the field K is complete
for a discrete valuation. Also suppose that f extends to a proper morphism
X P1oK such that the intersection of P1k with the nonsmooth locus is contained in the intersection of P1k with the Zariski closure of S (i.e., the morphism
is smooth over all points of P1k that are not the reductions of points in S). Let Mi
be the ith PicardFuchs module for f , and let : P1oK P1oK be a Frobenius
lift (e.g., t  t p ) that acts on o K as a lift of the absolute Frobenius morphism.
Then, for some (0, 1), there exists an isomorphism (Mi )
= Mi over the
1
1
Frchet completion R of (P K \ S, O) for the -Gauss norm and the Gauss
norms |t | = , [, 1) and S.
Remark 22.2.2. Geometrically, the Frobenius structure is defined on the complement in P1K of a union of discs around the points of S, each of radius less
than 1. (This includes a disc of radius less than 1 around , by which we mean
the complement of a disc of radius greater than 1 around 0.) In particular, by

22.3 Relationship to zeta functions

349

working in a unit disc not containing any points of S, we obtain a differential module with Frobenius structure over K t0 . In a unit disc containing one
or more points of S, we obtain a differential module with Frobenius structure
only over >0 K /t, t0 . If the disc contains exactly one point of S and the
exponents at that point are all 0, we can also obtain a differential module with
Frobenius structure over K t0 for the derivation td/dt, provided that fixes
that point.
Example 22.2.3. In Example 22.1.2, Theorem 22.2.1 applies directly except
when p = 2. In that case the reduction modulo p fails to be generically
smooth; one must change the defining equation to get a usable description
mod 2.

22.3 Relationship to zeta functions

We next give an explicit statement to the effect that the Frobenius structure on
a PicardFuchs module can be used to compute zeta functions. (The condition
on allows for a unique choice in each residue disc, by Lemma 15.2.6.)
Theorem 22.3.1. Retain the notation of Theorem 22.2.1, but now assume that
K = Fq with q = pa and that is a qth-power Frobenius lift on P1oK .
Suppose that o K satisfies (t ) 0 (mod t ) and that f extends
smoothly over the residue disc containing . Then
( f 1 (), T ) =

2 dim(
f )

det(1 T , (Mi ) )(1)

i+1

i=0

This suggests an interesting strategy for computing zeta functions, which

was described by Alan Lauder.
Remark 22.3.2. Suppose that we have in hand the differential module Mi ,
plus the action of  on some individual (Mi ) . (This data would ordinarily be
specified by a basis of Mi , the matrix of action of D, and the matrix of action
of  modulo t .) View the equation
NA+

d(t)
dA
=
A(N )
dt
dt

as a differential equation with initial condition provided by (Mi ) ; we may

then solve for A and evaluate at another .
More explicitly, let us suppose for simplicity that = 0 is the starting value
and = 1 is the target value. In the open unit disc around 0, we can compute
U such that

350

PicardFuchs modules
U 1 N U + U 1

dU
=0
dt

and then write down

A = U A0 (U 1 ).
This gives only a power series representation around t = 0 with radius of
convergence 1; it does not give any way in which we can specialize to = 1.
However, Theorem 22.2.1 implies that the entries of A can be written as
uniform limits of rational functions with limited denominators. For the purposes of computing the zeta function, we can limit how much p-adic accuracy
is needed in the computations by giving some bounds on the degrees of the
polynomials that appear and the sizes of the coefficients (e.g., using the Weil
conjectures, Theorem 0.2.5). To obtain this much accuracy, we must compute
A modulo some particular power of p. This means that we must determine
some rational function whose degree we can (in principle) control, so it suffices to determine suitably many terms in the power series expansion around
0. We can then reconstruct a sufficiently good rational function approximation
to A and evaluate at = 1.
Remark 22.3.3. One can recover from Theorem 22.3.1 the example of Dwork
discussed in Chapter 0. In that example one is separating the PicardFuchs
module, which has rank 2, into a unit-root component and a component
of slope log p. For this to be possible, one must be in the situation of
Theorem 15.3.4; this fails precisely at the residue discs at which the Igusa
polynomial vanishes, which is why one must invert the Igusa polynomial in
the course of the computation.

Notes
The notion of algebraic de Rham cohomology was introduced by Grothendieck
in [98], where he gave an algebraic construction of the topological de Rham
cohomology of a complex algebraic variety. The construction involves commingling the straightforward cohomology of the de Rham complex (constructed from a module of Khler differentials) with the cohomology of
coherent algebraic sheaves. The subject was further developed by Hartshorne
[105, 106]. A thorough treatment of Grothendiecks theorem can be found
in [97].
In the final footnote to [98], Grothendieck gave what we believe was the
first suggestion that one should consider in general what we call PicardFuchs

Notes

351

modules. His suggestion was based on Manins work on the Mordell conjecture over function fields; partly for this reason, PicardFuchs modules are also
commonly known as GaussManin connections. (This terminology is apparently due to Grothendieck. The name presumably refers to Gausss study of
hypergeometric differential equations, which incorporated the original Picard
Fuchs equation governing the periods of elliptic curves.) A good collection of
onward references can be found in [109].
The fact that a PicardFuchs module has regular singularities with rational
exponents can be proved in several ways, both geometric (Griffiths, Landman)
and arithmetic (Katz). One can also formulate a more abstract version, concerning polarized variations of Hodge structures; this was proved by Borel and
extended by Schmid (see the notes to Chapter 20 for further discussion of this
last point). Again, see [109] for references.
The construction of Frobenius structures on PicardFuchs modules is a
consequence of general results in the theory of p-adic cohomology, which
we discuss further in the next chapter. For the moment we point out that
Theorem 22.2.1 in its stated form can be found in [211, Theorem 3.3.1].
Lauders strategy for computing zeta functions (also called the deformation method) was introduced in [154]; it has been worked out in detail for
hyperelliptic curves by Hubrechts [116]. Hubrechts implemented the resulting
algorithm in version 2.14 of the computer algebra system Magma. A version
for hypersurfaces was described by Gerkmann [96].

23
Rigid cohomology

It has been suggested several times in this book that the study of p-adic
differential equations is deeply connected with the theory of p-adic cohomology for varieties over finite fields. In particular, the Frobenius structures arising
on PicardFuchs modules, discussed in the previous chapter, appear within this
theory.
In this chapter, we introduce a little of the theory of rigid p-adic cohomology, as developed by Berthelot and others. In particular, we illustrate the role
played by the p-adic local monodromy theorem in a fundamental finiteness
problem in the theory.

23.1 Isocrystals on the affine line

We start with a concrete description of p-adic cohomology in a very special
case, namely the cohomology of locally constant coefficient systems on the
affine line over a finite field. This is due to Crew [62].
Definition 23.1.1. Let k be a perfect (for simplicity) field of characteristic p >
0. Let K be a complete discrete (again for simplicity) nonarchimedean field of
characteristic 0 with K = k. An overconvergent F-isocrystal on the affine line
over k (with coefficients in K ) is a finite differential module with Frobenius
structure on the ring A = >1 K t/, for some absolute Frobenius lift ;
as in Proposition 17.3.1 the resulting category is independent of the choice of
Frobenius lift.
Definition 23.1.2. Let M be an overconvergent F-isocrystal on the affine
line over k. Let R be a copy of the Robba ring with series parameter t 1 ,
so that we can identify A as a subring of R. (We can write explicitly R =
>1 K  1 /t 1 , t 1 }.) Define
352

23.2 Crystalline and rigid cohomology

353

H 0 (A1k , M) = ker(D, M),

H 1 (A1k , M) = coker(D, M),
0
Hloc
(A1k , M) = ker(D, M A R),
1
Hloc
(A1k , M) = coker(D, M A R),

Hc1 (A1K , M) = ker(D, M A (R/A)),

Hc2 (A1K , M) = coker(D, M A (R/A)).
By taking kernels and cokernels in the short exact sequence
0 M M A R M A (R/A) 0
and applying the snake lemma, we get an exact sequence
0
(A1k , M) Hc1 (A1k , M)
0 H 0 (A1k , M) Hloc
1
H 1 (A1k , M) Hloc
(A1k , M) Hc2 (A1k , M) 0.

Remark 23.1.3. Crew [62] showed that in this construction H i computes the
rigid cohomology of M, Hci computes the rigid cohomology with compact
i computes some sort of local cohomology at .
supports, and Hloc
Crews main result in this setting was the following.
i (A1 , M)
Theorem 23.1.4 (Crew). The spaces H i (A1k , M), Hci (A1k , M), Hloc
K
are all finite-dimensional over K . Moreover, the Poincar pairings

H i (A1k , M) Hc2i (A1k , M ) Hc2 (A1k , A)

= K,
H i (A1 , M) H 1i (A1 , M ) H 1 (A1 , A)
=K
loc

loc

loc

are perfect.
The key ingredient is the fact that, by the p-adic local monodromy theorem
(Theorem 20.1.4), M A R is quasiunipotent which implies the finiteness
i (A1 , M). This further implies the finite dimensionalities, except for
of Hloc
k
1
Hc (A1k , M) and H 1 (A1k , M); however, these are related by a map with finitedimensional kernel and cokernel. Moreover, they carry incompatible topologies: the former is a Frchet space while the latter is dual to a Frchet space.
This incompatibility can be resolved only if both spaces are finite-dimensional.

23.2 Crystalline and rigid cohomology

Next, we briefly discuss how the previous example fits into a broader theory of
p-adic cohomology, deferring most references to the notes.

354

Rigid cohomology

Motivated by the work of Dwork and also by related work of Monsky and
Washnitzer, Grothendieck proposed a method of constructing an analogue of
algebraic de Rham cohomology in positive characteristic. This was developed
by Berthelot and Ogus into the theory of crystalline cohomology. An important example of this is the fact that if X is a smooth proper scheme over Z p
then the algebraic de Rham cohomology of X carries a Frobenius action that
computes the zeta function of the special fibre. (This generalizes to explain
the results on Frobenius structures and zeta functions discussed in the previous
chapter.)
One defect of crystalline cohomology is that it gives a coherent cohomology theory only for schemes over a finite field that are smooth and proper.
By contrast, the work of Monsky and Washnitzer had given a good theory
for smooth affine schemes. To fuse these theories, Berthelot introduced the
theory of rigid cohomology as well as a theory of locally constant coefficient
objects, overconvergent F-isocrystals. The example in the previous paragraph
demonstrates the computation of the cohomology of coefficients on the affine
line; a theory of cohomology with compact supports, and a local cohomology
theory, are also demonstrated by Crews construction.
For trivial coefficients, it was shown by Berthelot that rigid cohomology has
all the desired properties of a Weil cohomology: finite dimensionality, Poincar
duality, Knneth formula, cycle class maps, the Lefschetz trace formula for
Frobenius morphisms, etc. These were extended to nonconstant coefficients
by Kedlaya, using a relative version of Theorem 23.1.4.
Berthelot also suggested a theory of constructible coefficients, based on
ideas from the theory of algebraic D-modules. (These are modules over a ring
of differential operators; they are the natural coefficient objects in algebraic de
Rham cohomology.) Recent work of Caro, Kedlaya, and Tsuzuki has shown
that they form a good theory of coefficients, enjoying the same formal properties as their -adic tale counterparts. For example, one can execute a proof of
the Weil conjectures entirely using p-adic cohomology [132].

23.3 Machine computations

In recent years, interest has emerged in explicit computation of the zeta functions of algebraic varieties defined over finite fields. Some of this interest
has come from cryptography, in particular the use of the Jacobians of elliptic
(and later hyperelliptic) curves over finite fields as black box abelian groups
for certain public-key cryptography schemes (Diffie-Hellman, ElGamal). For
elliptic curves, a good method for doing this was proposed by Schoof [196]. It
amounts to computing the trace of the Frobenius morphism on the -torsion
points, the set of which is otherwise known as the tale cohomology with

Notes

355

F -coefficients, for values of  small enough to determine uniquely the one

unknown coefficient of the zeta function within the range prescribed by the
HasseWeil bound.
It turns out to be somewhat more difficult to execute Schoofs scheme for
curves of higher genus, as discovered by Pila [180]. One is forced to work with
higher-division polynomials in order to compute the torsion of the Jacobian
of the curve; an interpretation in terms of tale cohomology is of little value
because the definition of tale cohomology is not intrinsically computable. (It
is easy to write down cohomology classes, but it is difficult to test two such
classes for equality.)
It has been noticed by several authors that rigid cohomology is intrinsically
more computable and so lends itself better to this sort of task. Specifically,
Kedlaya [124] proposed an algorithm using rigid cohomology to compute the
zeta function of a hyperelliptic curve over a finite field of small odd characteristic. The limitation to odd characteristic was lifted by Denef and Vercauteren
[72]; the limitation to small characteristic was somewhat remedied by Harvey
[108], who improved the dependence on the characteristic p from O( p) to
O( p 1/2+ ).
More recently, interest has emerged in considering also higher-dimensional
varieties; this has come partly from potential applications in the study of mirror symmetry for CalabiYau varieties. In this case tale cohomology is of no
help at all, since there is no geometric interpretation of Heti for i > 1 analogous to the interpretation for i = 1 in terms of the Jacobian. Rigid cohomology
should still be computable, but relatively little progress has been made in making these computations practical (one exception being the treatment of smooth
surfaces in projective 3-space in [3]). It may be necessary to combine these
techniques with Lauders deformation method (see Remark 22.3.2) to obtain
the best results.

Notes
Crews work, and subsequent work that builds on it (e.g., [131]), makes essential use of nonarchimedean functional analysis, as was evident in the discussion
of Theorem 23.1.4. We recommend Schneiders book [195] as a user-friendly
introduction to this topic.
For general surveys of the subject of p-adic cohomology, we recommend [117] (for crystalline cohomology only) and [141] (broader but more
advanced). Also useful at this level of generality is Berthelots original article
outlining the theory of rigid cohomology [22].
For the basics of crystalline cohomology, see [27] and references within
(largely to Berthelots thesis [21]). For the work of Monsky and Washnitzer

356

Rigid cohomology

which motivated it, in addition to the original papers [172, 173, 174] there is
also a useful survey by van der Put [213].
Until recently no comprehensive introductory text on rigid cohomology was
available. That state of affairs was remedied by the appearance of the book
of Le Stum [157]. Berthelots own attempt at a foundational treatise remains
incomplete, but what does exist [23] may also be helpful. Some of the function
of a foundational text has been served by Berthelots articles concerning finite
dimensionality [24] and Poincar duality [25].
For coefficients in an overconvergent F-isocrystal, the basic properties
(finite dimensionality, Poincar duality, the Knneth formula) are proved in
[131] using techniques similar to those of Crew in [62]. However, these
techniques do not appear to suffice for the construction of a full coefficient
theory. For this, one needs a higher-dimensional analogue of Theorem 20.1.4,
called the semistable reduction theorem for overconvergent F-isocrystals. This
asserts that an overconvergent F-isocrystal on a noncomplete variety can
always be extended to a logarithmic isocrystal on some compactification, after
pulling back along a generically finite cover; such a result allows reductions
to arguments about (logarithmic) crystalline cohomology. (Except for the part
of the theorem that concerns the cover, this is analogous to Delignes theory
of canonical extensions in [68].) See [134, 137, 140, 143] for the proof; the
original results from the present book feature prominently in [143].
Berthelot wrote a useful survey of his theory of arithmetic D-modules [26].
It is this theory that has been developed by Caro (in part jointly with Tsuzuki)
into a full theory of p-adic coefficients. This work is quite expansive and
requires a summary more detailed than we can provide here; see [35, 36] for a
representative sample.
On the subject of machine calculations, as a companion to our original paper
on hyperelliptic curves [124] we recommend Edixhovens course notes [83].
Some discussion is also included in [93, Chapter 7]. We gave a condensed
summary of the general approach in [127]. For more on the role of elliptic and
hyperelliptic curves in cryptography (including the relevance of the problem of
the machine computation of zeta functions), the standard first reference is [55]
even though it is some years behind the frontiers of this fast-moving subject.
(For instance, it predates the growing area of pairing-based cryptography.)

24
p-adic Hodge theory

For our last application we turn to the subject of p-adic Hodge theory. Recall
that in Chapter 19 we described a nonabelian ArtinSchreier construction,
giving an equivalence of categories between continuous representations of the
absolute Galois group of a positive characteristic local field on a p-adic vector space and certain differential modules with Frobenius structures. In this
chapter, we describe an analogous construction for the Galois group of a
mixed-characteristic local field. We also mention a couple of applications of
this construction.
Hypothesis 24.0.1. Throughout this chapter, let K be a finite extension of Q p ,
let V be a finite-dimensional Q p -vector space, and let : G K GL(V ) be a
continuous homomorphism for the p-adic topology on V .

24.1 A few rings

We begin with the field of norms construction of Fontaine and Wintenberger.
Definition 24.1.1. Put K n = K ( pn ) and K = n K n . Let F = Frac W ( K )
and F  be the maximal subfields of K and K , respectively, that are unramified
over Q p . Put HK = G K and  K = G K /K = G K /HK .
+ be the inverse limit of the system
Definition 24.1.2. Put o = oC p . Let E
o/ po o/ po,
in which each map is the p-power Frobenius morphism (which is a ring homo + are sequences (x0 , x1 , . . . )
morphism). More explicitly, the elements of E
p
of elements of o/ po for which xn+1 = xn for all n. In particular, for any
+ , the quantity p n v p (xn ) is the same for all n for which
nonzero x E
357

358

p-adic Hodge theory

xn  = 0; we call this quantity v(x) and conventionally put v(0) = +. Choose

+ with 0 = 1 and 1  = 1. (This choice is somewhat
 = (0 , 1 , . . . ) E
analogous to the choice of a square root of 1 in C.)
The following observations were made by Fontaine and Wintenberger [92].
Proposition 24.1.3. The following are true.
(a) The ring E + is a domain in which p = 0, with fraction field E =
+ [( 1)1 ].
E
under
+ [0, +] extends to a valuation on E
(b) The function v : E
+
.
is complete and o = E
which E
E

(c) The field E is the algebraic closure of K (( 1)). (The embedding of

exists because v( 1) = p/( p 1) > 0.)
K (( 1)) into E
be the ring of Witt vectors of E,
i.e., the unique
Definition 24.1.4. Let A
The
complete discrete valuation ring with maximal ideal p and residue field E.
is algebraically closed, hence perfect.
uniqueness follows from the fact that E
lifts to an automorphism
In particular, the p-power Frobenius morphism on E

of A. (See the notes for Chapter 14 for further discussion of Witt vectors.)
can be uniquely written as the sum
Definition 24.1.5. Each element of A

n=0 p [x n ], where x n E and [x n ] denotes the Teichmller lift of x n (the

for all positive integers m); note
unique lift of xn that has a p m th root in A
p
p
with a weak topology,
that ([x]) = [x ] = [x] . We may thus equip A

n
in which a sequence xm = n=0 p [xm,n ] converges to zero if, for each n,
v(xm,n ) + as m +. Let AQ p be the completion of Z p [([] 1) ]
for the weak topology; as a topological ring, it is isomorphic to the ring oE
in A
defined over the base field Q p with its own weak topology. It is also -stable
because ([]) = [] p .
Definition 24.1.6. Let A be the completion of the maximal unramified
Put
extension of AQ p , viewed as a subring of A.
A K = A HK ,
where the right-hand side makes sense because we have chosen all the rings so
far in a functorial fashion, so that they indeed carry a G K -action. Note that A K
can be written as a ring of the form oE , but with coefficients in F  rather than
in Q p .
Definition 24.1.7. For any ring denoted with a boldface A so far, define the
corresponding ring with A replaced by B by tensoring over Z p with Q p . For

Z Q p is the fraction field of A.

instance, B = A
p

24.2 (, )-modules

359

24.2 (, )-modules
We are now ready to describe the mechanism, introduced by Fontaine, for converting Galois representations into modules over various rings equipped with
much simpler group actions.
Definition 24.2.1. Recall that V is a finite-dimensional vector space equipped
with a continuous G K -action. Put
D(V ) = (V Q p B) HK ;
by the HilbertNoether theorem, D(V ) is a finite-dimensional B K -vector
space, and the natural map D(V ) B K B V Q p B is an isomorphism.
Since we have taken only HK -invariants, D(V ) retains a semilinear action of
G K /HK =  K ; it also inherits an action of from B. That is, D(V ) is a
(, )-module over B K , i.e., a finite free B K -module equipped with semilinear and  K -actions that commute with each other. It is also tale, which is to
say the -action is tale (unit-root), because, as in Definition 19.2.4, one can
find a G K -invariant lattice in V .
Theorem 24.2.2 (Fontaine). The functor D, from the category of continuous
representations of G K on finite-dimensional Q p -vector spaces to the category
of tale (, )-modules over B K , is an equivalence of categories.
Proof. The inverse functor is
V = (D(V ) B K B)=1 .
The argument is similar to that in the proof of Proposition 19.1.5; see [89].
In much the same way that Proposition 19.1.5 was refined by
Theorem 19.3.1, Theorem 24.2.2 was refined by Cherbonnier and Colmez as
follows [38]. The big difference is that no additional restriction on the Galois
representations is imposed.
Definition 24.2.3. Let BQ p be the image of the bounded Robba ring E , under
the identification of E (having coefficients in Q p ) with BQ p , sending t to []
1. Let BK be the integral closure of BQ p in B K . Again, BK carries the actions
of and  K .

n
Definition 24.2.4. Let A be the set of x =
n=0 p [x n ] A such that
lim inf{v(xn )/n} > .
n

Define
D (V ) = (V Q p B ) HK ;
it is an tale (, )-module over BK .

360

p-adic Hodge theory

The following is the main result of [38].

Theorem 24.2.5 (CherbonnierColmez). The functor D , from the category
of continuous representations of G K on finite-dimensional Q p -vector spaces to
the category of tale (, )-modules over BK , is an equivalence of categories.
Remark 24.2.6. By Theorem 24.2.2, to check Theorem 24.2.5 it suffices to
check that the base extension functor from tale (, )-modules over BK
to tale (, )-modules over B K is an equivalence. The full faithfulness of
this functor is elementary and it follows from Lemma 16.3.2. The essential
surjectivity is much deeper; it amounts to the fact that the natural map
D (V ) B B V Q p B
K

is an isomorphism. Verifying this requires the development of an appropriate

analogy to Sens theory of decompletion; this analogy has been developed into
a full abstract Sen theory by Berger and Colmez [18].
A further variant was proposed by Berger [14].
Definition 24.2.7. Using the identification BQ p
= E , put
Brig,K = BK B R,
Qp

where R is the Robba ring over Q p ; the subscript rig indicates rigid in
the sense of rigid analytical geometry. Note that Brig,K admits continuous
extensions (for the LF-topology) of the actions of and  K . Define

(V ) = D (V ) B Brig,K ;
Drig
K

it is an tale (, )-module over Brig,K .

, from the category of continuous

Theorem 24.2.8 (Berger). The functor Drig
representations of G K on finite-dimensional Q p -vector spaces to the category
of tale (, )-modules over Brig,K , is an equivalence of categories.

Remark 24.2.9. The principal new content in Theorem 24.2.8 is that the base
extension functor from tale -modules over E to tale -modules over R is
fully faithful; this follows from Corollary 16.3.4. The essential surjectivity of
the functor is almost trivial, since the taleness of the -action is defined over
the Robba ring by base extension from E (Definition 16.3.1). One need only
check that the  K -action also descends to any tale lattice, but this is easy: the
proof is similar to that of Lemma 17.4.2.

24.3 Galois cohomology

361

24.3 Galois cohomology

Since the functor D and its variants lose no information about Galois representations, it is unsurprising that they can be used to recover basic invariants of a
representation, such as its Galois cohomology.
Definition 24.3.1. Assume for simplicity that  K is procyclic; this eliminates
only the case where p = 2 and {1} , for which see Remark 24.3.2
below. Let be a topological generator of . Define the Herr complex over
B K associated with V as the complex (with the first nonzero term placed in
degree 0)
0 D(V ) D(V ) D(V ) D(V ) 0;
here the first map is m  (( 1)m, ( 1)m) and the second is (m 1 , m 2 )
( 1)m 1 ( 1)m 2 . (The fact that this is a complex follows from the
commutativity of and .) Similarly, define the Herr complex over BK or

Brig,K by replacing D(V ) by D (V ) or Drig

(V ), respectively.
Remark 24.3.2. In a more conceptual description, which also covers the case
where  K need not be profinite, one takes the total complex associated with
1

0 C ( K , D(V )) C ( K , D(V )) 0,
where C ( K , D(V )) is the usual cochain complex for computing Galois
cohomology (as in [199, I.2.2]). One might think of this as the monoid
cohomology of  K Z0 acting on D(V ).
Theorem 24.3.3. The cohomology of the Herr complex computes the Galois
cohomology of V .
Proof. For B K , the desired result was established by Herr [111]. The argument
proceeds in two steps. One takes first the cohomology of the ArtinSchreier
sequence
1

0 Q p B B 0,
after tensoring with V . This reduces the claim to the fact that the inflation
homomorphisms
H i ( K , D(V )) H i (G K , V Q p B)
are bijections; this is proved by adapting a technique introduced by Sen.
For BK and Brig,K , the desired result was established by Liu [160]; his argument proceeds by comparison with the original Herr complex rather than by
imitation of the above argument, though one could probably do that also.

362

p-adic Hodge theory

Remark 24.3.4. As in [111, 160], one can make Theorem 24.3.3 more precise.
For instance, the construction of Galois cohomology is functorial; there is also
an interpretation in the Herr complex of the cup product in cohomology.
Remark 24.3.5. One can also use the Herr complex to recover Tates fundamental theorems on Galois cohomology (finite dimensionality, the Euler
Poincar characteristic formula, local duality). This was done by Herr in [112].

24.4 Differential equations from (, )-modules

One of the original goals of p-adic Hodge theory was to associate finer invariants with p-adic Galois representations, so as, for instance, to distinguish those
representations which arise in geometry (i.e., from the tale cohomology of
varieties over K ). This was originally done using a collection of period rings
introduced by Fontaine; more recently, Bergers work has demonstrated that
one can reproduce these constructions using (, )-modules. Here is a brief
description of an example that shows the relevance of p-adic differential equations to this study. We will make reference to Fontaines rings BdR , Bst without
giving definitions of them, for which see [15].
Definition 24.4.1. Let :  K Z
p denote the cyclotomic character; that is,
for all nonnegative integers m and all  K ,
( )

( pm ) = pm .
For  K sufficiently close to 1, we may compute
=

log
log ( )

as an endomorphism of D(V ), using the power series for log(1+ x). The result
does not depend on .
Remark 24.4.2. If one views  K as a one-dimensional p-adic Lie group over
Z p then is the action of the corresponding Lie algebra.
Definition 24.4.3. Note that acts on Brig,K with respect to
f  [] log[]

df
.
d[]

As a result, it does not induce a differential module structure with respect

to d/dt on D(V ), but only on D(V ) Brig,K [(log[])1 ]. We say that V is
de Rham if there exists a differential module with Frobenius structure M over
Brig,K and an isomorphism

24.5 Beyond Galois representations

363

D(V ) Brig,K [(log[])1 ] M Brig,K [(log[])1 ]

of differential modules with Frobenius structure.
One then has the following results of Berger [14].
Theorem 24.4.4 (Berger).
(a) The representation V is de Rham if and only if it is de Rham in
Fontaines sense, i.e., if
DdR (V ) = (V Q p BdR )G K
satisfies
DdR (V ) K BdR
= V Q p BdR .
(b) Suppose that V is de Rham. If V is semistable in Fontaines sense,
i.e., if
Dst (V ) = (V Q p Bst )G K
satisfies
Dst (V ) F Bst
= V Q p Bst
then there exists an M as in Definition 24.4.3 that is unipotent. Conversely, if such an M exists then V is potentially semistable, i.e., it
becomes semistable upon restriction to G L for some finite extension
L of K .
Applying Theorem 20.1.4 then yields the following corollary, which was
previously a conjecture of Fontaine [90, 6.2].
Corollary 24.4.5 (Berger). Every de Rham representation is potentially
semistable.
Remark 24.4.6. The descriptor de Rham is meant to convey the fact that
if V = Heti (X K K alg , Q p ) for X a smooth proper variety over K then
V is de Rham and one can use the aforementioned constructions to recover
i (X, K ) functorially from V (solving Grothendiecks problem of the
HdR
mysterious functor). See [15] for more of this story.

24.5 Beyond Galois representations

The category of arbitrary (, )-modules over Brig,K turns out to have its own
representation-theoretic interpretation: it is equivalent to the category of Bpairs introduced by Berger [16]. One can associate Galois cohomology with

364

p-adic Hodge theory

such objects using the Herr complex; it has been shown by Liu [160] that the
analogues of Tates theorems (see Remark 24.3.5) still hold. These functors

can be interpreted as the derived functors of Hom(Drig

(V0 ), ) when V0 is the
trivial representation [142, Appendix].
It may be wondered why we should be interested in (, )-modules over
Brig,K if ultimately we have in mind an application concerning only Galois
representations. One answer is that converting Galois representations into
(, )-modules exposes extra structure that is not visible without such a
conversion.

(V ) is a
Definition 24.5.1 (Colmez). We say that V is trianguline if Drig

successive extension of (, )-modules of rank 1 over Brig,K . The point is

that these need not be tale, so V need not be a successive extension of
representations of dimension 1.
Trianguline representations have the dual benefits of being relatively easy to
classify and also somewhat commonplace. On one hand, Colmez [58] classified the two-dimensional trianguline representations of G Q p ; the classification
included a parameter (the L-invariant) relevant to p-adic L-functions. On
the other hand, a result of Kisin [146] shows that the Galois representations
associated with many classical modular forms are in fact trianguline.

Notes
Our presentation here was largely a summary of Bergers [15], which we highly
recommend. For detailed study, the recent notes by Brinon and Conrad for the
2009 Clay Mathematics Institute summer school [32] may be of considerable
value.
A variant of the theory of (, )-modules was introduced by Kisin [147],
n
using the Kummer tower K ( p 1/ p ) instead of the cyclotomic tower K ( pn ).
This leads to certain advantages, particularly when studying crystalline representations. For instance, Kisin was able to use his construction to establish
some classification theorems for finite flat group schemes and for p-divisible
groups, as conjectured by Breuil. Kisins work was based on an earlier paper
of Berger [17]; both of these use slope filtrations (as in Theorem 16.4.1)
to recover a theorem of Colmez and Fontaine classifying semistable Galois
representations in terms of certain linear algebraic data.
After [14] had appeared, Fontaine succeeded in giving a direct proof of
Corollary 24.4.5 that did not involve p-adic differential equations; see [91].

References

[1] A. Abbes and T. Saito, Ramification of local fields with imperfect residue fields, Amer. J.
Math. 124 (2002), 879920.
[2] A. Abbes and T. Saito, Ramification of local fields with imperfect residue fields, II, Doc.
Math. extra vol. (2003), 572.
[3] T. G. Abbott, K. S. Kedlaya, and D. Roe, Bounding Picard numbers of surfaces using
p-adic cohomology, in Arithmetic, Geometry and Coding Theory (AGCT 2005), Societ
Mathmatique de France, 2009, pp. 125159.
[4] Y. Andr, Diffrentielles non commutatives et thorie de Galois diffrentielle ou aux
diffrences, Ann. Sci. cole Norm. Sup. (4) 34 (2001), 685739.
[5] Y. Andr, Filtrations de type HasseArf et monodromie p-adique, Invent. Math. 148 (2002),
285317.
[6] Y. Andr, Dworks conjecture on the logarithmic growth of solutions of p-adic differential
equations, Compos. Math. 144 (2008), 484494.
[7] Y. Andr and F. Baldassarri, De Rham Cohomology of Differential Modules on Algebraic
Varieties, Progress in Math. 189, Birkhuser, 2001.
[8] Y. Andr and L. Di Vizio, q-difference equations and p-adic local monodromy, Astrisque
296 (2004), 55111.
[9] M. F. Atiyah and I. G. Macdonald, Introduction to Commutative Algebra, Addison-Wesley,
Reading, 1969.
[10] F. Baldassarri, Continuity of the radius of convergence of differential equations on p-adic
analytic curves, arXiv:0809.2479v3 (2009).
[11] F. Baldassarri and L. Di Vizio, Continuity of the radius of convergence of p-adic differential
equations on Berkovich analytic spaces, arXiv:0709.2008v3 (2008).
[12] M. Baker and R. Rumely, Potential Theory and Dynamics on the Berkovich Projective Line,
Math. Surveys and Monographs 159, Amer. Math. Soc., 2010.
[13] A. Beilinson, S. Bloch, and H. Esnault, -factors for GaussManin determinants, Moscow
J. Math. 2 (2002), 477532.
[14] L. Berger, Reprsentations p-adiques et quations diffrentielles, Invent. Math. 148 (2002),
219284.
[15] L. Berger, An introduction to the theory of p-adic representations, in Geometric Aspects of
Dwork Theory, Vols. I, II, de Gruyter, Berlin, 2004, 255292.
[16] L. Berger, Construction de (, )-modules: reprsentations p-adiques et B-paires, Algebra
and Num. Theory 2 (2008), 91120.
[17] L. Berger, quations diffrentielles p-adiques et (, N )-modules filtrs, Astrisque 319
(2008), 1338.
[18] L. Berger and P. Colmez, Familles de reprsentations de de Rham et monodromie p-adique,
Astrisque 319 (2008), 303337.

365

366

References

[19] V. G. Berkovich, Spectral Theory and Analytic Geometry over Non-Archimedean Fields,
Math. Surveys and Monographs 33, Amer. Math. Soc., Providence, 1990.
[20] V. G. Berkovich, tale cohomology for non-Archimedean analytic spaces, Publ. Math. IHS
78 (1993), 5161.
[21] P. Berthelot, Cohomologie Cristalline des Schmas de Caractristique p > 0, Lecture Notes
in Math. 407, Springer-Verlag, Berlin, 1974.
[22] P. Berthelot, Gomtrie rigide et cohomologie des varits algbriques de caractristique
p, in Introductions aux cohomologies p-adiques (Luminy, 1984), Mm. Soc. Math. France
(N.S.) 23 (1986), 732.
[23] P. Berthelot, Cohomologie rigide et cohomologie rigide support propre. Premire partie, Prpublication IRMAR 96-03, available at http://perso.university-rennes1.fr/pierre.
berthelot/.
[24] P. Berthelot, Finitude et puret cohomologique en cohomologie rigide (with an appendix in
English by A. J. de Jong), Invent. Math. 128 (1997), 329377.
[25] P. Berthelot, Dualit de Poincar et formule de Knneth en cohomologie rigide, C.R. Acad.
Sci. Paris 325 (1997), 493498.
[26] P. Berthelot, Introduction la thorie arithmtique des D-modules, cohomologies p-adiques
et applications arithmtiques, II, Astrisque 279 (2002), 180.
[27] P. Berthelot and A. Ogus, Notes on Crystalline Cohomology, Princeton University Press,
Princeton, 1978.
[28] R. Bhatia, Matrix Analysis, Graduate Texts in Math. 169, Springer-Verlag, New York,
1997.
[29] J. M. Borger, Conductors and the moduli of residual perfection, Math. Annalen 329
(2004), 130.
[30] S. Bosch, Lectures on formal and rigid geometry, preprint (2008) available at
http://wwwmath1.uni-muenster.de/sfb/about/publ/bosch.html.
[31] S. Bosch, U. Gntzer, and R. Remmert, Non-Archimedean Analysis, Grundlehren der Math.
Wiss. 261, Springer-Verlag, Berlin, 1984.
[32] O. Brinon and B. Conrad, p-adic Hodge theory, notes from the 2009 Clay Mathematics
Institute summer school, available online at http://math.stanford.edu/ conrad/.
[33] S. Boucksom, C. Favre, and M. Jonsson, Valuations and plurisubharmonic singularities,
Publ. Res. Inst. Math. Sci. 44 (2008), 449494.
[34] K. Buzzard and F. Calegari, Slopes of overconvergent 2-adic modular forms, Compos. Math.
141 (2005), 591604.
[35] D. Caro, F-isocristaux surconvergents et surcohrence diffrentielle, Invent. Math. 170
(2007), 507539.
[36] D. Caro, Dvissages des F-complexes de D-modules arithmtiques en F-isocristaux
surconvergents, Invent. Math. 166 (2006), 397456.
[37] N. L. Carothers, A Short Course on Banach Space Theory, Cambridge University Press,
2004.
[38] F. Cherbonnier and P. Colmez, Reprsentations p-adiques surconvergentes, Invent. Math.
133 (1998), 581611.
[39] B. Chiarellotto and G. Christol, On overconvergent isocrystals and F-isocrystals of rank
one. Compos. Math. 100 (1996), 7799.
[40] B. Chiarellotto and A. Pulita, Arithmetic and Differential Swan conductors of rank one
representations with finite local monodromy, arXiv:0711.0701v1 (2007); to appear in Amer.
J. Math.
[41] B. Chiarellotto and N. Tsuzuki, Logarithmic growth and Frobenius filtrations for solutions
of p-adic differential equations, J. Inst. Math. Jussieu 8 (2009), 465505.
[42] G. Christol, Modules Diffrentiels et Equations Diffrentielles p-adiques, Queens Papers
in Pure and Applied Math. 66, Queens University, Kingston, 1983.

References

367

[43] G. Christol, Un thorme de transfert pour les disques singuliers rguliers, in Cohomologie
p-adique, Astrisque 119120 (1984), 151168.
[44] G. Christol, About a Tsuzuki theorem, in p-adic Functional Analysis (Ioannina, 2000),
Lecture Notes in Pure and Appl. Math. 222, Dekker, New York, 2001, 6374.
[45] G. Christol, Thirty years later, in Geometric Aspects of Dwork Theory, Vol. I, de Gruyter,
Berlin, 2004, 419436.
[46] G. Christol, Le Thorme de Turritin p-adique, in preparation; draft available at http://
people.math.jussieu.fr/christol.
[47] G. Christol and B. Dwork, Effective p-adic bounds at regular singular points, Duke Math.
J. 62 (1991), 689720.
[48] G. Christol and B. Dwork, Differential modules of bounded spectral norm, in p-adic Methods in Number Theory and Algebraic Geometry, Contemp. Math. 133, Amer. Math. Soc.,
Providence, 1992.
[49] G. Christol and B. Dwork, Modules diffrentielles sur les couronnes, Ann. Inst. Fourier 44
(1994), 663701.
[50] G. Christol and Z. Mebkhout, Sur le thorme de lindice des quations diffrentielles, II,
Annals of Math. 146 (1997), 345410.
[51] G. Christol and Z. Mebkhout, Sur le thorme de lindice des quations diffrentielles, III,
Annals of Math. 151 (2000), 385457.
[52] G. Christol and Z. Mebkhout, Sur le thorme de lindice des quations diffrentielles, IV,
Invent. Math. 143 (2001), 629672.
[53] R. C. Churchill and J. J. Kovacic, Cyclic vectors, in Differential Algebra and Related Topics
(Newark, NJ, 2000), World Scientific, River Edge NJ, 2002, 191218.
[54] D. Clark, A note on the p-adic convergence of solutions of linear differential equations,
Proc. Amer. Math. Soc. 17 (1966), 262269.
[55] H. Cohen and G. Frey (eds.), Handbook of Elliptic and Hyperelliptic Curve Cryptography,
Discrete Math. and its Applications 34, Chapman & Hall/CRC, 2005.
[56] R. M. Cohn, Difference Algebra, John Wiley & Sons, New York, London, Sydney, 1965.
[57] R. F. Coleman, Dilogarithms, regulators and p-adic L-functions, Invent. Math. 69 (1982),
171208.
[58] P. Colmez, Reprsentations triangulines de dimension 2, Astrisque 319 (2008), 213258.
[59] B. Conrad, Several approaches to non-archimedean geometry, in M. Baker et al.
(eds.), p-adic Geometry: Lectures from the 2007 Arizona Winter School, University Lecture Series 45, Amer. Math. Soc., 2008.
[60] F. T. Cope, Formal solutions of irregular linear differential equations, II, Amer. J. Math. 58
(1936), 130140.
[61] R. Crew, F-isocrystals and p-adic representations, in Algebraic Geometry (Brunswick,
Maine, 1985), Proc. Symp. Pure Math. 46, Part 2, Amer. Math. Soc., Providence, 1987,
111138.
[62] R. Crew, Finiteness theorems for the cohomology of an overconvergent isocrystal on a curve,
Ann. Sci. c. Norm. Sup. 31 (1998), 717763.
[63] R. Crew, Canonical extensions, irregularities, and the Swan conductor, Math. Ann. 316
(2000), 1937.
[64] N. E. Csima, NewtonHodge filtration for self-dual F-crystals, arXiv:0706/2530v1 (2007).
[65] A. J. de Jong, Homomorphisms of BarsottiTate groups and crystals in positive characteristic, Invent. Math. 134 (1998), 301333; erratum, ibid. 138 (1999), 225.
[66] A. J. de Jong, BarsottiTate groups and crystals, in Proc. Int. Congress of Mathematicians,
Vol. II (Berlin, 1998), Doc. Math. Extra Vol. II (1998), 259265.
[67] A. J. de Jong and F. Oort, Purity of the stratification by Newton polygons, J. Amer. Math.
Soc. 13 (2000), 209241.
[68] P. Deligne, Equations Diffrentielles Points Singuliers Rguliers, Lecture Notes in Math.
163, Springer-Verlag, Berlin, 1970.

368

References

[69] P. Deligne, La conjecture de Weil, II, Publ. Math. IHS 52 (1980), 137252.
[70] P. Deligne and N. Katz (eds.), Groupes de Monodromie en Gomtrie Algbrique II, Proc.
Sminaire de Gomtrie Algbrique du Bois-Marie 19671969 (SGA 7 II), Lecture Notes in
Math. 340, Springer-Verlag, Berlin, New York, 1973.
[71] M. Demazure, Lectures on p-divisible Groups, Lecture Notes in Math. 302, SpringerVerlag, New York, 1972.
[72] J. Denef and F. Vercauteren, An extension of Kedlayas algorithm to hyperelliptic curves in
characteristic 2, J. Cryptology 19 (2006), 125.
[73] L. Di Vizio, Local analytic classification of q-difference equations with |q| = 1, J. Noncommutative Geom. 3 (2009), 125149.
[74] B. Dwork, p-adic cycles, Publ. Math. IHS 37 (1969), 27115.
[75] B. Dwork, On p-adic differential equations, II. The p-adic asymptotic behavior of solutions
of ordinary linear differential equations with rational function coefficients, Ann. Math. (2)
98 (1973), 366376.
[76] B. Dwork, On p-adic differential equations, III. On p-adically bounded solutions of ordinary linear differential equations with rational function coefficients, Invent. Math. 20 (1973),
3545.
[77] B. Dwork, Bessel functions as p-adic functions of the argument, Duke Math. J. 41 (1974),
711738.
[78] B. Dwork, Lectures on p-adic Differential Equations, Grundlehren der Math. Wissenschaften 253, Springer-Verlag, New York, 1982.
[79] B. M. Dwork, On exponents of p-adic differential modules, J. reine angew. Math. 484
(1997), 85126.
[80] B. Dwork, G. Gerotto, and F. Sullivan, An Introduction to G-Functions, Ann. Math. Studies
133, Princeton University Press, Princeton, 1994.
[81] B. Dwork and P. Robba, On ordinary linear p-adic differential equations, Trans. Amer. Math.
Soc. 231 (1977), 146.
[82] B. Dwork and P. Robba, Effective p-adic bounds for solutions of homogeneous linear
differential equations, Trans. Amer. Math. Soc. 259 (1980), 559577.
[83] B. Edixhoven, Point counting after Kedlaya, course notes (2006) available at http://
www.math.leidenuniv.nl/ edix/oww/mathofcrypt/carls_edixhoven/kedlaya.pdf.
[84] D. Eisenbud, Commutative Algebra with a View Toward Algebraic Geometry, Graduate
Texts in Math. 150, Springer-Verlag, New York, 1995.
[85] P. Enflo, A counterexample to the approximation problem in Banach spaces, Acta Math. 130
(1973), 309317.
[86] G. Faltings, Coverings of p-adic period domains, research announcement (2007) available
at http://www.mpim-bonn.mpg.de/preprints/.
[87] C. Favre and M. Jonsson, The Valuative Tree, Lecture Notes in Math. 1853, Springer-Verlag,
Berlin, 2004.
[88] C. Favre and M. Jonsson, Valuations and multiplier ideals, J. Amer. Math. Soc. 18 (2005),
655684.
[89] J.-M. Fontaine, Reprsentations p-adiques des corps locaux, I, in The Grothendieck
Festschrift, Vol. II, Progress in Math. 87, Birkhuser, Boston, 1990, 249309.
[90] J.-M. Fontaine, Reprsentations p-adiques semi-stables, in Priodes p-adiques (Bures-surYvette, 1988), Astrisque 23 (1994), 113184.
[91] J.-M. Fontaine, Reprsentations de de Rham et reprsentations semi-stables, Orsay preprint
number 2004-12, available online at http://www.math.u-psud.fr/ biblio/pub/2004/.
[92] J.-M. Fontaine and J.-P. Wintenberger, Le corps de normes de certaines extensions
algbriques de corps locaux, C.R. Acad. Sci. Paris Sr. A, B 288 (1979), A367A370.
[93] J. Fresnel and M. van der Put, Rigid Analytic Geometry and its Applications, Progress in
Mathematics 218, Birkhuser, Boston, 2004.

References

369

[94] W. Fulton, Intersection Theory, second edition, Springer-Verlag, Berlin, 1998.

[95] W. Fulton, Eigenvalues, invariant factors, highest weights, and Schubert calculus, Bull.
Amer. Math. Soc. (N.S.) 37 (2000), 209249.
[96] R. Gerkmann, Relative rigid cohomology and deformation of hypersurfaces, Int. Math. Res.
Papers 2007 (2007), article ID rpm003 (67 pp.).
[97] P. Griffiths and J. Harris, Principles of Algebraic Geometry, Wiley Interscience, New York,
1978.
[98] A. Grothendieck, On the de Rham cohomology of algebraic varieties, Publ. Math. IHS 29
(1966), 95103.
[99] A. Grothendieck, in Revtements tales et Groupe Fondamental, Proc. Sminaire de
Gomtrie Algbrique du Bois-Marie 19601961 (SGA 1), Lecture Notes in Math. 224,
Springer-Verlag, Berlin, 1971.
[100] H. Hahn, ber die nichtarchimedische Grensysteme, in Gesammelte Abhandlungen I,
Springer, Vienna, 1995.
[101] U. Hartl, Period spaces for Hodge structures in equal characteristic, arXiv:math/0511686v2
(2006).
[102] U. Hartl, On a conjecture of Rapoport and Zink, arXiv:0605254v1 (2006).
[103] U. Hartl, On period spaces for p-divisible groups, C.R. Math. Acad. Sci. Paris 346 (2008),
11231128.
[104] U. Hartl and R. Pink, Vector bundles with a Frobenius structure on the punctured unit disc,
Compos. Math. 140 (2004), 689716.
[105] R. Hartshorne, Residues and Duality, Lecture Notes in Math. 20, Springer, New York,
1966.
[106] R. Hartshorne, On the de Rham cohomology of algebraic varieties, Publ. Math. IHS 45
(1975), 599.
[107] R. Hartshorne, Algebraic Geometry, Graduate Texts in Math. 52, Springer-Verlag, New
York, 1977.
[108] D. Harvey, Kedlayas algorithm in larger characteristic, Int. Math. Res. Notices 2007 (2007),
article ID rnm095 (29 pages).
[109] M. Hazewinkel, GaussManin connection, in Encyclopedia of Mathematics, third edition,
Kluwer, 2001; available online at http://eom.springer.de/G/g043470.htm.
[110] M. Hazewinkel, Witt vectors, part 1, arXiv:0804.3888v1 (2008).
[111] L. Herr, Sur la cohomologie galoisienne des corps p-adiques, Bull. Soc. Math. France 126
(1998), 563600.
[112] L. Herr, Une approche nouvelle de la dualit locale de Tate, Math. Ann. 320 (2001),
307337.
[113] W. V. D. Hodge and D. Pedoe, Methods of Algebraic Geometry, Vol. III (reprint of the 1954
original), Cambridge University Press, Cambridge, 1994.
[114] A. Horn, On the eigenvalues of a matrix with prescribed singular values, Proc. Amer. Math.
Soc. 5 (1954), 47.
[115] A. Horn, Eigenvalues of sums of Hermitian matrices, Pacific J. Math. 12 (1962), 225241.
[116] H. Hubrechts, Point counting in families of hyperelliptic curves, Found. Comp. Math. 8
(2008), 137169.
[117] L. Illusie, Crystalline cohomology, in Motives (Seattle, WA, 1991), Proc. Symp. Pure Math.
55, Part 1, Amer. Math. Soc., Providence, 1994, 4370.
[118] I. Kaplansky, Maximal fields with valuations, Duke Math. J. 9 (1942), 303321.
[119] K. Kato, Swan conductors for characters of degree one in the imperfect residue field case,
in Algebraic K -theory and algebraic number theory (Honolulu, HI, 1987), Contemp. Math.
83, Amer. Math. Soc., Providence, 1989, 101131.
[120] N. M. Katz, Slope filtration of F-crystals, in Journes de Gomtrie Algbrique de Rennes
(Rennes, 1978), Vol. I, Astrisque 63 (1979), 113163.

370

References

[121] N. M. Katz, On the calculation of some differential Galois groups, Invent. Math. 87 (1987),
1361.
[122] N. M. Katz, A simple algorithm for cyclic vectors, Amer. J. Math. 109 (1987), 6570.
[123] N. M. Katz and T. Oda, On the differentiation of de Rham cohomology classes with respect
to parameters, J. Math. Kyoto Univ. 8 (1968), 199213.
[124] K. S. Kedlaya, Counting points on hyperelliptic curves using MonskyWashnitzer cohomology, J. Ramanujan Math. Soc. 16 (2001), 323338; errata, ibid. 18 (2003), 417418.
[125] K. S. Kedlaya, A p-adic local monodromy theorem, Annals of Math. 160 (2004), 93184.
[126] K. S. Kedlaya, Full faithfulness for overconvergent F-isocrystals, in Geometric Aspects of
Dwork Theory (Vol. II), de Gruyter, Berlin, 2004, 819835.
[127] K. S. Kedlaya, Computing zeta functions via p-adic cohomology, in Algorithmic Number
Theory (ANTS VI), Lecture Notes in Comp. Sci. 3076, Springer-Verlag, 2004, 117.
[128] K. S. Kedlaya, Local monodromy of p-adic differential equations: an overview, Int. J.
Number Theory 1 (2005), 109154; errata at http://math.mit.edu/ kedlaya/papers/.
[129] K. S. Kedlaya, Slope filtrations revisited, Documenta Math. 10 (2005), 447525; errata, ibid.
12 (2007), 361362.
[130] K. S. Kedlaya, Frobenius modules and de Jongs theorem, Math. Res. Lett. 12 (2005),
303320.
[131] K. S. Kedlaya, Finiteness of rigid cohomology with coefficients, Duke Math. J. 134 (2006),
1597.
[132] K. S. Kedlaya, Fourier transforms and p-adic Weil II, Compos. Math. 142 (2006),
14261450.
[133] K. S. Kedlaya, Swan conductors for p-adic differential modules, I: A local construction,
Algebra and Number Theory 1 (2007), 269300.
[134] K. S. Kedlaya, Semistable reduction for overconvergent F-isocrystals, I: Unipotence and
logarithmic extensions, Compos. Math. 143 (2007), 11641212.
[135] K. S. Kedlaya, The p-adic local monodromy theorem for fake annuli, Rend. Sem. Math.
Padova 118 (2007), 101146.
[136] K. S. Kedlaya, Slope filtrations for relative Frobenius, Astrisque 319 (2008), 259301.
[137] K. S. Kedlaya, Semistable reduction for overconvergent F-isocrystals, II: A valuationtheoretic approach, Compos. Math. 144 (2008), 657672.
[138] K. S. Kedlaya, Swan conductors for p-adic differential modules, II: Global variation,
arXiv:0705.0031v3 (2008); to appear in J. Institut Math. Jussieu.
[139] K. S. Kedlaya, Effective p-adic cohomology for cyclic cubic threefolds, preprint (2008)
available at http://math.mit.edu/ kedlaya/papers/.
[140] K. S. Kedlaya, Semistable reduction for overconvergent F-isocrystals, III: Local semistable
reduction at monomial valuations, Compos. Math. 145 (2009), 143172.
[141] K. S. Kedlaya, p-adic cohomology, in Algebraic Geometry (Seattle 2005), Proc. Symp. Pure
Math. 80, Amer. Math. Soc., 2009, 667684.
[142] K. S. Kedlaya, Some new directions in p-adic Hodge theory, J. Thorie Nombres Bordeaux
21 (2009), 285300.
[143] K. S. Kedlaya, Semistable reduction for overconvergent F-isocrystals, IV: Local semistable
reduction at nonmonomial valuations, arXiv:0712.3400v3 (2009).
[144] K. S. Kedlaya, Good formal structures for flat meromorphic connections, I: Surfaces,
arXiv:0811.0904v4(2009); to appear in Duke Math. J.
[145] K. S. Kedlaya and Liang Xiao, Differential modules on p-adic polyannuli, J. Inst. Math.
Jussieu 9 (2010), 155201.
[146] M. Kisin, Overconvergent modular forms and the FontaineMazur conjecture, Invent. Math.
153 (2003), 373454.
[147] M. Kisin, Crystalline representations and F-crystals, in Algebraic Geometry and Number
Theory, Progress in Math. 253, Birkhuser, Boston, 2006, 459496.

References

371

[148] R. E. Kottwitz, Isocrystals with additional structure, Comp. Math. 56 (1985), 201220.
[149] R. E. Kottwitz, Isocrystals with additional structure. II, Comp. Math. 109 (1997), 255339.
[150] R. E. Kottwitz, On the HodgeNewton decomposition for split groups, Int. Math. Res.
Notices 26 (2003), 14331447.
[151] W. Krull, Allgemeine Bewertungstheorie, J. fr Math. 167 (1932), 160196.
[152] F.-V. Kuhlmann, Elimination of ramification I: The generalized stability theorem, preprint
available at http://math.usask.ca/ fvk/; to appear in Trans. Amer. Math. Soc.
[153] S. Lang, Algebraic groups over finite fields, Amer. J. Math. 78 (1956), 555563.
[154] A. G. B. Lauder, Deformation theory and the computation of zeta functions, Proc. London
Math. Soc. 88 (2004), 565602.
[155] G. Laumon, Cohomology of Drinfeld Modular Varieties I, Cambridge Studies in Advanced
Math. 41, Cambridge University Press, 1996.
[156] M. Lazard, Les zros dune fonction analytique dune variable sur un corps valu complet,
Publ. Math. IHS 14 (1962), 4775.
[157] B. Le Stum, Rigid Cohomology, Cambridge Tracts in Math. 172, Cambridge University
Press, 2007.
[158] A. H. M. Levelt, Jordan decomposition for a class of singular differential operators, Ark.
Mat. 13 (1975), 127.
[159] J. Lindenstrauss and L. Tzafriri, Classical Banach Spaces, Lecture Notes in Math. 338,
Springer, 1973.
[160] R. Liu, Cohomology and duality for (, )-modules over the Robba ring, Int. Math. Res.
Not. 2007, article ID rnm150 (32 pages).
[161] F. Loeser, Exposants p-adiques et thormes dindice pour les quations diffrentielles
p-adiques (daprs G. Christol et Z. Mebkhout), in Proc. Sminaire Bourbaki, Vol. 1996/97,
Astrisque 245 (1997), 5781.
[162] A. Loewy, ber einen Fundamentalsatz fr Matrizen oder lineare homogene Differentialsysteme, Sitzungsberichte der Heidelberger Akademie der Wissenschaften, 5. Abhandlung
(1918), 136.
[163] E. Lutz, Sur lquation y 2 = x 3 + Ax + B sur les corps p-adiques, J. reine angew Math.
177 (1937), 238247.
[164] B. Malgrange, Sur les points singuliers des quations diffrentielles, Enseign. Math. 20
(1974), 147176.
[165] B. Malgrange, Connexions mromorphes 2: le rseau canonique, Invent. Math. 124 (1996),
367387.
[166] Yu. I. Manin, The theory of commutative formal groups over fields of finite characteristic
(in Russian), Usp. Math. 18 (1963), 390; English translation, Russian Math. Surveys 18
(1963), 180.
[167] A. Marmora, Irrgularit et conducteur de Swan p-adiques, Doc. Math. 9 (2004), 413433.
[168] S. Matsuda, Local indices of p-adic differential operators corresponding to ArtinSchreier
Witt coverings, Duke Math. J. 77 (1995), 607625.
[169] S. Matsuda, Katz correspondence for quasi-unipotent overconvergent isocrystals, Comp.
Math. 134 (2002), 134.
[170] S. Matsuda, Conjecture on AbbesSaito filtration and ChristolMebkhout filtration, in
Geometric Aspects of Dwork Theory. Vols. I, II, de Gruyter, Berlin, 2004, 845856.
[171] Z. Mebkhout, Analogue p-adique du thorme de Turrittin et le thorme de la monodromie
p-adique, Invent. Math. 148 (2002), 319351.
[172] P. Monsky, Formal cohomology, II: the cohomology sequence of a pair, Ann. Math. 88
(1968), 218238.
[173] P. Monsky, Formal cohomology, III: Fixed point theorems, Ann. Math. 93 (1971),
315343.
[174] P. Monsky and G. Washnitzer, Formal cohomology, I, Ann. Math. 88 (1968), 181217.

372

References

[175] M. Nagata, Local Rings, Interscience Tracts in Pure and Applied Math. 13, John Wiley &
Sons, New York, 1962.
[176] F. Oort and T. Zink, Families of p-divisible groups with constant Newton polygon, Doc.
Math. 7 (2002), 183201.
[177] O. Ore, Theory of non-commutative polynomials, Ann. Math. 34 (1933), 480508.
[178] B. Osserman, The Weil conjectures, in W. T. Gowers (ed.), The Princeton Companion to
Mathematics, Princeton University Press, 2008.
[179] A. Ostrowski, Untersuchungen zur arithmetischen Theorie der Krper, Math. Z. 39 (1935),
269404.
[180] J. Pila, Frobenius maps of abelian varieties and finding roots of unity in finite fields, Math.
Comp. 55 (1990), 745763.
[181] E. Pons, Modules diffrentiels non solubles. Rayons de convergence et indices, Rend. Sem.
Mat. Univ. Padova 103 (2000), 2145.
[182] B. Poonen, Maximally complete fields, Enseign. Math. (2) 39 (1993), 87106.
[183] S. Priess-Crampe and P. Ribenboim, Differential equations over valued fields (and more),
J. reine angew. Math. 576 (2004), 123147.
[184] A. Pulita, Frobenius structure for rank one p-adic differential equations, in Ultrametric functional analysis, Contemp. Math. 384, Amer. Math. Soc., Providence, 2005,
247258.
[185] A. Pulita, Rank one solvable p-adic differential equations and finite Abelian characters via
LubinTate groups, Math. Ann. 337 (2007), 489555.
[186] J.-P. Ramis, Thormes dindices Gevrey pour les quations diffrentielles ordinaires, Mem.
Amer. Math. Soc. 48 (1984).
[187] P. Ribenboim, The Theory of Classical Valuations, Springer-Verlag, New York, 1999.
[188] J. F. Ritt, Differential Algebra, Colloq. Pub. XXXIII, Amer. Math. Soc., New York, 1950.
[189] P. Robba, Lemmes de Hensel pour les oprateurs diffrentiels. Application la rduction
formelle des quations diffrentielles, Enseign. Math. (2) 26 (1980), 279311.
[190] P. Robba, Indice dun oprateur diffrentiel p-adique. IV. Cas des systmes. Mesure de
lirrgularit dans un disque. Ann. Inst. Fourier (Grenoble), 35 (1985), 1355.
[191] A. M. Robert, A Course in p-adic Analysis, Graduate Texts in Math. 198, Springer-Verlag,
New York, 2000.
[192] R. T. Rockafellar, Convex Analysis, Princeton University Press, Princeton, 1970.
[193] W. H. Schikhof, Ultrametric Calculus: An Introduction to p-adic Analysis, Cambridge
Studies in Advanced Math. 4, Cambridge University Press, 1984.
[194] W. Schmid, Variation of Hodge structure: the singularities of the period mapping, Invent.
Math. 22 (1973), 211319.
[195] P. Schneider, Nonarchimedean Functional Analysis, Springer-Verlag, Berlin, 2002.
[196] R. Schoof, Elliptic curves over finite fields and the computation of square roots mod p,
Math. Comp. 44 (1985), 483494.
[197] E. Schrner, Ultrametric fixed point theorems and applications, in Valuation Theory and its
Applications, Vol. II (Saskatoon, SK, 1999), Fields Inst. Communications 33, Amer. Math.
Soc., 2003, 353359.
[198] J.-P. Serre, Local Fields, Graduate Texts in Math. 67, Springer-Verlag, New York, 1979.
[199] J.-P. Serre, Galois Cohomology, Corrected reprint of the 1997 English edition, Springer
Monographs in Math., Springer-Verlag, Berlin, 2002.
[200] J. H. Silverman, The Arithmetic of Elliptic Curves, second printing, Graduate Texts in Math.
106, Springer-Verlag, New York, 1991.
[201] M. F. Singer and M. van der Put, Galois Theory of Difference Equations, Lecture Notes in
Math. 1666, Springer-Verlag, Berlin, 1997.
[202] M. F. Singer and M. van der Put, Galois Theory of Linear Differential Equations,
Grundlehren der Math. Wiss. 328, Springer-Verlag, Berlin, 2003.

References

373

[203] J. Tate, p-divisible groups, in Proc. Conf. on Local Fields (Driebergen, 1966), SpringerVerlag, Berlin, 1967, 158183.
[204] T. Terasoma, Confluent hypergeometric functions and wild ramification, J. Algebra 185
(1996), 118.
[205] A. Thuillier, Thorie du potentiel sur les courbes en gomtrie analytique non archimdienne. Applications la theorie dArakelov, thesis, Universit de Rennes 1, 2005.
[206] N. Tsuzuki (as T. Nobuo), The overconvergence of morphisms of tale --spaces on a
local field, Compos. Math. 103 (1996), 227239.
[207] N. Tsuzuki, Finite local monodromy of overconvergent unit-root F-isocrystals on a curve,
Amer. J. Math. 120 (1998), 11651190.
[208] N. Tsuzuki, The local index and the Swan conductor, Comp. Math. 111 (1998), 245288.
[209] N. Tsuzuki, Slope filtration of quasi-unipotent overconvergent F-isocrystals, Ann. Inst.
Fourier (Grenoble) 48 (1998), 379412.
[210] N. Tsuzuki, Morphisms of F-isocrystals and the finite monodromy theorem for unit-root
F-isocrystals, Duke Math. J. 111 (2002), 385419.
[211] N. Tsuzuki, On base change theorem and coherence in rigid cohomology, Doc. Math. extra
volume in honour of Kazuya Katos fiftieth birthday (2003), 891918.
[212] H. L. Turrittin, Convergent solutions of ordinary linear homogeneous differential equations
in the neighborhood of an irregular singular point, Acta Math. 93 (1955), 2766.
[213] M. van der Put, The cohomology of Monsky and Washnitzer, in Introductions aux
cohomologies p-adiques (Luminy, 1984), Mm. Soc. Math. France 23 (1986), 3359.
[214] A. C. M. van Rooij, Non-Archimedean Functional Analysis, Monographs and Textbooks in
Pure and Applied Math. 51, Marcel Dekker, New York, 1978.
[215] M. Vaqui, Valuations, in Resolution of Singularities (Obergurgl, 1997), Progr. Math. 181,
Birkhuser, Basel, 2000, 539590.
[216] C. A. Weibel, An Introduction to Homological Algebra, Cambridge Studies in Advanced
Math. 38, Cambridge University Press, Cambridge, 1994.
[217] H. Weyl, Das asymptotische Verteilungsgesetz der Eigenwerte linearer partieller Differentialgleichungen (mit einer Anwendung auf die Theorie der Hohlraumstrahlung), Math. Ann.
71 (1912), 441479.
[218] H. Weyl, Inequalities between the two kinds of eigenvalues of a linear transformation, Proc.
Nat. Acad. Sci. USA 35 (1949), 408411.
[219] L. Xiao, On ramification filtrations and p-adic differential equations, I: equal characteristic
case, arXiv:0801.4962v2 (2008).
[220] L. Xiao, On ramification filtrations and p-adic differential equations, II: mixed characteristic
case, arXiv:0811.3792v1 (2008).
[221] L. Xiao, Non-archimedean differential modules and ramification theory, thesis, Massachusetts Institute of Technology, 2009.
[222] P. T. Young, Radii of convergence and index for p-adic differential operators, Trans. Amer.
Math. Soc. 333 (1992), 769785.
[223] H. Zassenhaus, ber eine Verallgemeinerung des Henselschen Lemmas, Arch. Math.
(Basel) 5 (1954), 317325.
[224] I. B. Zhukov, A approach to higher ramification theory, in Invitation to higher local
fields (Mnster, 1999), Geom. Topol. Monographs 3, Geom. Topol. Publ., Coventry, 2000,
143150.
[225] I. B. Zhukov, On ramification theory in the case of an imperfect residue field (in Russian),
Mat. Sb. 194 (2003), 330; translation, Sb. Math. 194 (2003), 17471774.
[226] T. Zink, On the slope filtration, Duke Math. J. 109 (2001), 7995.

Notation

| | (-Gauss (semi)norm), 26
|A| (norm of a matrix), 56, 63
C p (completed algebraic closure of
Q p ), 28
D(V )
equal characteristic, 314
mixed characteristic, 359
D (V )
equal characteristic, 317, 318
mixed characteristic, 359

(V ), 360
Drig
Diag(1 , . . . , n ) (diagonal
matrix), 55
disc( f, r ) (discrepancy), 202
E (completion of oK ((t)) o K K ), 156
E (overconvergent series), 263
E L (unramified extension of E), 314
E L (unramified extension of E ), 316
E (perfection of E), 269

E (perfection of E ), 269
E (completed maximal unramified extension
of E), 314
[E : F] (degree of a field extension), 27
Ext(, ) (Yoneda extension group), 82
F alg (algebraic closure), 45
F (completion of K (t) for | | ), 155
F (completion of K (t p ) for | | p ), 171
F (completion of K ((t 1) p 1) for
| | p ), 181
F sep (separable closure), 45
|F | (multiplicative value group), 16

374

G E/F (Galois group), 45

G F (absolute Galois group), 45
 K (in p-adic Hodge theory), 357
H 0, H 1
of difference modules, 245
of differential modules, 78
H K (in p-adic Hodge theory), 357
I R(V ) (intrinsic radius), 157
K (complete discretely valued field), 243
K (complete nonarchimedean
field), 135
K /t, t0 (bounded series on an
annulus), 137
K /t, t/ (series on a disc or
annulus), 136
K /t, t/} (series on a half-open annulus),
143
K /t, t/an (analytic elements on a
half-open annulus), 144
K /t, t/an (analytic elements on an open
annulus), 144
K t/ (series on a disc), 137
K {t/} (series on an open disc), 143
K t/0 (series with bounded
coefficients), 137
K t/an (analytic elements on a
disc), 144
K t (series with logarithmic
growth), 308
mF (maximal ideal), 17
oF (valuation ring), 17
Q p (field of p-adic numbers), 26

Notation
R (Robba ring), 264
R L (extension of R), 327
R(M) (radius of convergence), 154
R{T } (ring of twisted polynomials)
over a difference ring, 246
over a differential ring, 85
(extended Robba ring), 282
R
R(V ) (generic radius of convergence), 156
R (group of units), 13
s ( f ) (slope of a function), 185
T (transpose), 55
(transpose, Hermitian), 56
T (inverse transpose), 55
|T | (operator norm), 94
T (translation), 185

375

|T |sp, (spectral radius), 94

vr (Gauss semivaluation), 26
Wm (differential module), 171
Z p (ring of p-adic integers), 26
Zq (unramified extension of Z p ), 7
F (residue field), 17
(type of field element), 17
(Frobenius pullback), 171
(Frobenius pushforward), 171
(cyclotomic character), 362
(Frobenius lift), 362
(off-center Frobenius pullback), 181
(off-center Frobenius pushforward), 181
(equals p 1/( p1) when p > 0), 156

Index

Boldface page numbers indicate definitions of terms or statements of results

abelian group
complete, 14
separated, 14
absolute Frobenius lift, 265
absolute value, see also norm
additive value group, 16
adjoint form (of a twisted polynomial), 86, 246
analytic elements, 144, 156
approximation lemmas, 20, 30, 147, 206
ArtinHasse exponential series, 162, 293
ArtinSchreier extension, 51, 317, 334, 341
ArtinSchreier theory
nonabelian, 259, 315, 357
average norm (of a difference module), 281
B-pair, 363
Bzout domain, 276
Baer sum, 82
Banach contraction mapping theorem, 41, 269
Banach space, 23
BarsottiTate group, see also p-divisible
group
base change
of a difference module, 244
of a differential module, 81, 98, 107, 178
Berkovich spaces, 8
Bessel differential equation, 120, 328
binomial series, 50
Birkhoff factorization, 149
bounded homomorphism, 14, 94
bounded Robba ring, 217, 263, 359
CalabiYau variety, 355
Cauchy sequence, 14
Cauchy theorem, 120
p-adic, 154, 302

376

center (of a Frobenius lift), 265

change-of-basis matrix, 81, 244
clean differential module, 214
Cohen structure theorem, 333
coherent locally free module, 144
compactness (of a closed interval), 136,
203, 208
companion matrix, 64
compatible seminorm on a module, 17
complete abelian group, 14
completion (of an abelian group), 15
connection, 81, 91
constant subring (of a differential ring), 78
constituents (of a differential module), 80
contraction mapping theorem, 41, 269
convexity, 190
Cramers rule, 66
Crews conjecture, 335
crystalline cohomology, 353
cyclic vector
of a difference module, 246
of a differential module, 84
theorem, 84, 89, 246, 260, 311
cyclotomic character, 362
de Rham cohomology, 347
de Rham representation, 336, 362
decomposition theorem
for difference modules, 252
refined strong, 179, 214
refined visible, 113
strong, 177
visible, 107
derivation of rational type, 164
DieudonnManin classification theorem, 258,
283, 316, 332

Index
difference
field, 244
difference-closed, 247
isometric, 249
Galois theory, 258
ideal, 244
module, 244
ring, 244
differential
Galois theory, 91
ideal, 78
module, 78
normed, 96
on an open annulus, 152
trivial, 78
unipotent, 78, 153
ring or field, 77
slopes, 211
dimensioned constants, 27
discrepancy
of a differential module, 203
of a function, 202
discretely valued field, 16
dual
of a difference module, 245
of a differential module, 82
duality for convex functions, 43
dualizable difference module, 245
Dwork
exponential series, 181, 292
trick, 273, 275, 280, 281, 295,
329, 336
eigenvalues
prepared, 123
weakly prepared, 123
elementary divisors, 63
elementary matrix, 55
elementary row or column
operation, 56
elliptic curve, 3
elliptic integral, 5
equivalence (of p-adic exponents), 224
tale cohomology, 5, 354
tale difference module, see also unit-root
difference module
tale fundamental group, 323
exponents
of a differential module, 122
of a p-adic differential module, 226
extended Robba ring, 282
exterior power, 57, 82
exterior product , 57
extrinsic generic radius of convergence, 157
Feketes lemma, 94
field of norms, 357

377

filtration
lower numbering, 52
upper numbering, 52, 321
finite local monodromy ( p-adic), 318
formal adjoint (of a twisted polynomial), 86,
246
Frchet completeness, 139
Frobenius
antecedent, 173, 181, 213, 225, 303, 341,
342
descendant, 171, 181, 193, 196, 206
lift, 265
absolute, 265
standard, 297
zero-centered, 266
off-center, 181, 196
structure (on a differential module), 160,
291, 327
Fuchs theorem, 125
p-adic, 221, 231, 302
Fuchsian singularities, 130
full spectrum, 100
fundamental solution matrix, 123, 221, 234,
295, 302, 308
Galois cohomology, 363
Galois descent, 129, 173
Gauss
(semi)norm, 26, 138, 282
for twisted polynomials, 104
(semi)valuation, 26
for twisted polynomials, 104
GaussJordan elimination, 63, 141
GaussManin connection, 351
generalized power series, 31, 282
generic Newton polygon, 266, 280
generic point, 160
generic radius of convergence, 156
extrinsic, 157
intrinsic, 157
Gevrey functions, 130
gluing lemma, 142, 208
Hadamard three-circle theorem, 139
HahnBanach theorem, 32
HasseArf theorem, 322
Hasses theorem, 3
henselian field, 46, 263
Hensels lemma, 41
Hermitian transpose, 56
Herr complex, 361
HilbertNoether theorem, 173, 314, 359
Hodge polygon
of a difference module, 254
of a matrix, 63
HodgeNewton decomposition, 65
for difference modules, 254

378
homomorphism
bounded, 14
isometric, 14
horizontal
element (of a differential module), 78
morphism (of differential modules), 81
Horn inequalities, 71
hypergeometric
differential equation, 5, 120, 159,
222, 348
series, 5, 159
Igusa polynomial, 6
index
of a differential module, 126
of nilpotency, 123
inertia group, 49, 323
intrinsic
generic radius of convergence, 157
radius, 157
subsidiary generic radii of
convergence, 162
invariant factors, 63
inversive difference ring, 245
irregularity, 119, 162
isoclinic difference module, 250
isometric
difference field, 249
homomorphism, 14
JordanHlder constituents, 80
KiehlTate theorems, 149
Kummer theory, 50
Lang torsor, 259
Lefschetz
principle, 119
trace formula, 5
limit (of a sequence), 14
Liouville number, 219
LittlewoodRichardson
numbers, 71
local horizontal section, 154
locally free module, 84
lower convex hull, 36, 105
lower numbering filtration, 52
MalcevNeumann series, 31, 282
master factorization theorem, 39, 104, 140,
142, 246
matrix of action, 80, 244
maximally complete field, 28
metrically equivalent seminorms, 14
minor (of a matrix), 61
mirror symmetry, 355
monodromy transformation, 122

Index
multiplicative
seminorm, 15
value group, 16
mysterious functor, 363
Newton above Hodge theorem, 64
for difference modules, 254
Newton polygon
generic, 266, 280
of a difference module, 254, 279
of a matrix, 64
of a polynomial, 36
of a power series, 138
of a twisted polynomial, 104, 246
semicontinuity theorem, 266, 280, 311, 329
special, 266, 280
Newtons method, 264
nilpotency index, 123
nonarchimedean
group, ring, field, 16
seminorm, 16
norm, 14
Gauss, 26
for twisted polynomials, 104
p-adic, 26
spectral, 95
supremum, 17
t-adic, 25
trivial, 25
normal (matrix), 60
normed differential module, 96
off-center Frobenius antecedents and
descendants, 181, 196
operator norm, 56, 94, 121
opposite
difference ring/module, 245
ring, 86
optimal basis, 198
order of log-growth, 308
Ostrowskis theorem, 17, 27
outer provinces, rebellious, 300
overconvergent F-isocrystal (on the affine
line), 352
p-adic
Cauchy theorem, 154, 302
Fuchs theorem, 221, 231, 302
Hodge theory, 357
integers, ring of, 26
Liouville number, 219
local monodromy theorem, 2, 114, 234,
279, 327
unit-root, 318
norm, 26
numbers, field of, 26
Turrittin theorem, 338

Index
p-divisible group, 285, 336, 364
p-power roots of unity, 226
perfection, 249, 339
(, )-module, 359
PicardFuchs
equation, 7
module, 88, 122, 294, 347
piecewise affine function, 136
Poincar pairing, 353
potentially semistable representation,
336, 363
prepared eigenvalues, 123
principal minor (of a matrix), 61
product formula, 26
product seminorm, 23
projective module, 84
pseudoconvergent sequence, 30
pseudolimit, 30
pure
difference module, 250, 277
differential module, 100
quasiconstant differential module, 327
quasiunipotent
differential module, 327
monodromy, 122
quotient seminorm, 16
radii of optimal convergence, 198
radius of convergence (of a differential
module), 154
ramification numbers, 211
RapoportZink spaces, 285
rational type, 164
real
semivaluation, 16
valuation, 16
rebellious outer provinces quelled by
Frobenius, 300
refined decomposition theorem
strong, 179, 214
visible, 113
refined differential module, 100, 342
regular differential module, 119,
234, 347
residue field (of a nonarchimedean field), 17
rigid analytic geometry, 8
rigid cohomology, 353
Robba condition, 223
Robba ring, 264, 360
bounded, 217, 263, 359
extended, 282
Schauder basis, 32
semiabsolute value, see also seminorm
semicontinuity theorem for Newton polygons,
266, 280, 311, 329

379

seminorm, 13
compatible, 17
Gauss, 26
metrically equivalent, 14
multiplicative, 15
product, 23
quotient, 16
submultiplicative, 15
supremum, 17
topologically equivalent, 14
semistable
difference module, 281
semivaluation, 16
Gauss, 26
separated abelian group, 14
shearing transformation, 125, 226, 227,
234, 310
Shilov boundary, 149
singular value, 95
archimedean, 56
decomposition, 56
nonarchimedean, 63
slope (of a polynomial), 36
slope filtration theorem, 279, 298,
340, 364
Smith normal form, 63
solvable differential module, 211, 294, 302
special Newton polygon, 266, 280
spectral norm, 95
spectral radius, 94
spectrum
full, 100
visible, 105
spherically complete metric space, 28
standard Frobenius lift, 297
strong decomposition theorem, 177
strong triangle inequality, 16
strongly difference-closed difference
field, 247
subharmonicity, 190
submultiplicative seminorm, 15
subsidiary (generic) radii (of convergence),
162
supersingular elliptic curve, 6
supremum (semi)norm, 17
supremum-equivalent seminorm, 18
t-adic valuation, 25
tame degree (of an extension), 49
tamely ramified extension, 49, 101, 113, 179,
323
Taylor series, 90, 159, 173, 226, 296
Teichmller lift, 7, 259, 358
tensor product
of difference modules, 244
of differential modules, 81
topologically equivalent seminorms, 14

380
totally ramified extension, 48
totally wildly ramified extension, 49
trace (element or component), 82
trace-zero component, 82
transfer theorem, 158
regular singular, 234, 295
triangle inequality, 14
strong, 16
trianguline representation, 364
trivial
difference module, 244
differential module, 78
TurrittinLeveltHukuhara decomposition
theorem, 114, 128, 338
twisted polynomial
over a difference ring, 246
over a differential ring, 85
type (of a field element), 218
ultrametric seminorm, 16
unipotent differential module,
78, 153
unit-root
difference module, 250, 298,
315, 318
p-adic local monodromy theorem, 318
unramified extension, 47
upper numbering filtration, 52, 321

Index
valuation, 16
Gauss, 26
for twisted polynomials, 104
t-adic, 25
valuation ring (of a nonarchimedean field), 17
value group, 16
variation of Hodge structures, 335
visible decomposition theorem, 107, 176
refined, 113
visible spectrum, 105
weak equivalence (of p-adic exponents), 224
weak topology, 358
weakly difference-closed difference field, 247
weakly prepared eigenvalues, 123, 221
wedge power, see also exterior power
Weierstrass preparation, 140
Weil conjectures, 4
well-ordered set, 31, 282
width (of a polynomial), 35
wild inertia subgroup, 49
wildly ramified extension, 49
Witt vectors, 258, 293, 358
Yoneda extension group, 82, 245
zero-centered Frobenius lift, 266
zeta function, 4, 349, 354