p-adic Analysis in Arithmetic Geometry

Winter Semester 2015/2016

University of Bayreuth
Michael Stoll

Contents

1. Introduction 2
2. p-adic numbers 3
3. Newton Polygons 14
4. Multiplicative seminorms and Berkovich spaces 19
5. The Berkovich affine and projective line 24
6. Analytic spaces and functions 34
7. Berkovich spaces of curves 39
8. Integration 45
References 52

Screen Version of March 29, 2020, 11:29.

 § 1. Introduction 2

1. Introduction

We all know the fields R and C of real and complex numbers as the completion of
the field Q of rational numbers and its algebraic closure. In particular, we have
a canonical embedding Q ,→ R ⊂ C, which we can sometimes use to get number
theoretic results by applying analysis over R or C.
Now R is not the only completion of Q. Besides the usual absolute value, there are
more absolute values on Q; to be precise, up to a natural equivalence (and except
for the trivial one), there is one absolute value |·|p for each prime number p. (We
will explain what an absolute value is in due course.) Completing Q with respect
to |·|p leads to the field Qp of p-adic numbers; we can then take its algebraic
closure Q̄p and the completion Cp of that (Q̄p is, in contrast to R̄ = C, not
complete), which has similar properties as C: it is the smallest extension field
of Q that is algebraically closed and complete with respect to the p-adic absolute
value.
There is a general philosophy in Number Theory that ‘all completions are created
equal’ and should have the same rights. In many situations, one gets the best
results by considering them all together. Since Qp and Cp are complete metric
spaces (and the former is, like R or C, locally compact), we can try to do analysis
over them. Many concepts and results of ‘classical’ analysis carry over without
great problems. But we will see that there is one feature of the p-adic topology
that is a stumbling block for an easy transfer of certain parts of analysis (like
for example line integrals) to the p-adic setting: this topology is totally discon-
nected. One goal of this lecture course is to explain a way to resolve this problem,
which is to embed Cp (say) into a larger ‘analytic space’ Canp that is (in this case,
even uniquely) path-connected. This approach is due to Vladimir Berkovich; the
analytic spaces constructed in this way are also known as Berkovich Spaces. V.G. Berkovich
c MFO 2012
 § 2. p-adic numbers 3

2. p-adic numbers

In this section we will recall (or introduce) the field Qp of p-adic numbers and its
properties. We begin with a rather general definition.

2.1. Definition. Let K be a field. An absolute value on K is a map DEF

absolute
|·| : K −→ R≥0 , x 7−→ |x| value
with the following properties.

(1) |x| = 0 ⇐⇒ x = 0.
(2) (Multiplicativity) |xy| = |x| · |y|.
(3) (Triangle Inequality) |x + y| ≤ |x| + |y|.

If |·| satisfies the stronger inequality

(30 ) (Ultrametric Triangle Inequality) |x + y| ≤ max{|x|, |y|},

then the absolute value is said to be ultrametric or non-archimedean, otherwise it

is archimedean.
Two absolute values |·|1 and |·|2 on K are said to be equivalent, if there is a
constant c > 0 such that |x|2 = |x|c1 for all x ∈ K. ♦

Properties (1), (2) and (3) imply that d(x, y) = |x − y| defines a metric on K.
((2) is needed for the symmetry: it implies that | − 1| = 1, so that |y − x| =
| − 1| · |x − y| = |x − y|.) The absolute value is then continuous as a real-valued
function on the metric space (K, d).
One can show that two absolute values on K are equivalent if and only if they
induce the same topology on K (Exercise).

2.2. Lemma. Let (K, |·|) be a field with a non-archimedean absolute value. Then LEMMA
R = {x ∈ K : |x| ≤ 1} is a subring of K. R is a local ring (i.e., it has exactly one valuation ring
maximal ideal) and has K as its field of fractions.

Proof. Exercise! q

This ring is the valuation ring of (K, | · |). DEF

valuation ring
Recall the following definition.

2.3. Definition. A metric space (X, d) is said to be complete, if every Cauchy DEF
sequence in X converges in X: if (xn ) is a sequence in X such that complete
metric space
lim sup d(xn , xn+m ) = 0 ,
n→∞ m≥1

then there exists x ∈ X such that limn→∞ d(xn , x) = 0.

If K is a field with absolute value |·|, then (K, |·|) is said to be complete, if the
metric space (K, d) is, where d is the metric induced by the absolute value. ♦
 § 2. p-adic numbers 4

2.4. Examples. EXAMPLES

Every field K has the trivial absolute value |·|0 with |x|0 = 1 for all x 6= 0. absolute
values
The usual absolute value is an (archimedean) absolute value on Q, R and C. The
latter two are complete, Q is not.
|f (x)/g(x)| = edeg(f )−deg(g) defines a non-archimedean absolute value on K(x), the
field of rational functions in one variable over K. (Exercise!)
If |·| is an absolute value and 0 < c ≤ 1, then |·|c is also an absolute value, which
is equivalent to |·|. If |·| is non-archimedean, then this remains true for c > 1.
(Exercise!) ♣

We can now introduce the p-adic absolute values on Q.

2.5. Definition. Let p be a prime number. If a 6= 0 is an integer, we define its DEF

p-adic valuation to be p-adic
vp (a) = max{n ∈ Z≥0 : pn | a} . absolute
value
For a = r/s ∈ Q× (with r, s ∈ Z, s 6= 0), we set vp (a) = vp (r) − vp (s). The p-adic
absolute value on Q is given by
(
0 if x = 0,
|x|p = −vp (x)
♦
p otherwise.

So the p-adic absolute value of x ∈ Q is small when (the numerator of) x is divisible
by a high power of p, and it is large when the denominator of x is divisible by a
high power of p. If x ∈ Z, then we clearly have |x|p ≤ 1: contrary to the familiar
situation with the usual absolute value, the integers form a bounded subset of Q
with respect to |·|p ! This explains the word ‘non-archimedean’ — the Archimedean
Axiom states that if x, y ∈ R>0 , then there is n ∈ Z such that nx > y; this is
equivalent to saying that Z is unbounded.

2.6. Lemma. The p-adic absolute value is a non-archimedean absolute value LEMMA
on Q. |·|p is
abs. value
Proof. We check the properties in Definition 2.1. Property (1) is clear from the
definition. Property (2) follows from vp (ab) = vp (a)+vp (b), which is a consequence
of unique factorization. Property (30 ) follows from vp (a + b) ≥ min{vp (a), vp (b)},
which is a consequence of the elementary fact that pn | a and pn | b together imply
that pn | a + b. q

Note that Q is not complete with respect to |·|p (Exercise!).

2.7. Example. Let (K, |·|) be a field with absolute value. Then for every x ∈ K EXAMPLE
with |x| < 1, the series geometric
∞
X series
xn
n=0
converges to the limit 1/(1 − x): we have
N −1
X 1 −xN
xn − = = |1 − x|−1 · |x|N ,
n=0
1−x 1−x
 § 2. p-adic numbers 5

which tends to zero, since |x| < 1 (note that |1 − x| ≥ 1 − |x| > 0, so the fraction
makes sense).
For example,
1
1 + p + p2 + p 3 + p4 + . . . =
1−p
in (Q, |·|p ). ♣

From the viewpoint of analysis, it is desirable to work with a complete field.

Indeed, one possible construction of the field of real numbers is as the smallest
complete field containing (Q, |·|∞ ), where (from now on) |·|∞ denotes the usual
absolute value. In fact, this construction works quite generally.

2.8. Theorem. Let (K, |·|) be a field with an absolute value. Then there is a THM
field (K 0 , |·|0 ) extending K such that |·|0 restricts to |·| on K, (K 0 , |·|0 ) is complete completion
and K is dense in K 0 .

Proof. Let C(K) be the ring of Cauchy sequences over K (with term-wise addition
and multiplication). The set N (K) of null sequences (i.e., sequences converging
to zero) forms an ideal in the ring C(K). We show that this ideal is actually
maximal. It does not contain the unit (all-ones) sequence, so it is not all of C(K).
If (xn ) ∈ C(K)\N (K), then it follows from the definition of ‘Cauchy sequence’ that
there are n0 ∈ Z>0 and c > 0 such that |xn | ≥ c for all n ≥ n0 . Then the sequence
(yn ) given by yn = 0 for n < n0 and yn = 1/xn for n ≥ n0 is a Cauchy sequence
and (1) = (xn ) · (yn ) + (zn ) with (zn ) ∈ N (K), so N (K) + C(K) · (xn ) = C(K).
Since N (K) is a maximal ideal in C(K), the quotient ring K 0 := C(K)/N (K) is a
field. There is a natural inclusion K ,→ C(K) by mapping a ∈ K to the constant
sequence (a), which by composition with the canonical epimorphism C(K) → K 0
gives an embedding i : K ,→ K 0 . We define |·|0 by
|[(xn )]|0 = lim |xn |
n→∞

(where [(xn )] denotes the residue class mod N (K) of the sequence (xn ) ∈ C(K)).
The properties of absolute values and of Cauchy sequences imply that this is well-
defined (i.e., the limit exists and does not depend on the choice of the represen-
tative sequence). That |·|0 is an absolute value follows easily from the assumption
that |·| is, and it is clear that |·|0 restricts to |·| on K.
It is also easy to see that K is dense in K 0 . Let x = [(xn )] ∈ K 0 , then |i(xn ) − x|0 =
limm→∞ |xn − xn+m | tends to zero as n tends to infinity, so we can approximate x
arbitrarily closely by elements of K.
It remains to show that K 0 is complete. So let (x(ν) )ν be a Cauchy sequence in K 0
(ν)
and represent each x(ν) by a Cauchy sequence (xn )n in K. Since K is dense
(ν)
in K 0 , we can do this in such a way that |i(xn ) − x(ν) |0 ≤ 2−n for all ν and n.
(The reason for this requirement is that we need some uniformity of convergence
(n)
to make the proof work.) Let yn = xn . We claim that (yn ) is a Cauchy sequence
(in K) and that y = [(yn )] = limν→∞ x(ν) (in K 0 ). To see the first, pick ε > 0.
There is ν0 such that |x(ν) − x(ν+µ) |0 < ε for all ν ≥ ν0 and µ ≥ 0. Then
(n+m)
|yn − yn+m | = |x(n)
n − xn+m |
(n) 0 (n+m)
≤ |i(x(n)
n )−x | + |x(n) − x(n+m) |0 + |x(n+m) − i(xn+m )|0
≤ 2−n + ε + 2−n−m < 2ε
 § 2. p-adic numbers 6

when 2−n < ε/2 and n ≥ ν0 . To see the second, note that
|x(ν) − y|0 = lim |x(ν) (n)
n − xn |
n→∞
(ν) 0
≤ lim sup |i(x(ν) (ν)
− x(n) |0 + |x(n) − i(x(n) 0


n ) − x | + |x n )|
n→∞

≤ lim sup(2−n + |x(ν) − x(n) |0 + 2−n ) = lim sup |x(ν) − x(n) |0 ,

n→∞ n→∞

which tends to zero as ν → ∞. q

One can show that (K 0 , |·|0 ) is determined up to unique isomorphism (of extensions
of K with absolute value), but we will not need this in the following.
Since |·|0 extends |·|, we will usually use the same notation for both. We will also
consider K as a subfield of K 0 .

2.9. Definition. Let p be a prime number. The completion Qp of Q with respect DEF
to the p-adic absolute value is called the field of p-adic numbers. Its valuation field of
ring Zp is the ring of p-adic integers. ♦ p-adic
numbers
Since |x|p takes a discrete set of values for x 6= 0, it follows that
Zp = {x ∈ Qp : |x|p ≤ 1} = {x ∈ Qp : |x|p < p}
is closed and open in Qp . This implies that for any two distinct elements x, y ∈ Qp ,
there are disjoint open neighborhoods X of x and Y of y such that Qp = X ∪ Y :
Qp is totally disconnected. To see this, let δ = |x − y|p > 0. Then the open
ball X around x of radius δ (which is x + pn+1 Zp , if δ = p−n ) is also closed,
so its complement Y = Qp \ X is open as well, and y ∈ Y . One consequence
of this is that any continuous map γ : [0, 1] → Qp is constant (otherwise let x
and y be two distinct elements in the image and let X and Y be as above; then
[0, 1] is the disjoint union of the two open non-empty subsets γ −1 (X) and γ −1 (Y ),
contradicting the fact that intervals are connected).
Next we want to show that Zp is compact. For this we need a fact about approx-
imation of p-adic numbers by rationals.

2.10. Lemma. Let x ∈ Qp \ {0} with |x|p = p−n . There is a ∈ {0, 1, 2, . . . , p − 1} LEMMA
such that |x − apn |p < |x|p . approximation
of p-adic
numbers
Proof. Replacing x by p−n x, we can assume that |x|p = 1, in particular, x ∈ Zp .
Since Q is dense in Qp , there is r/s ∈ Q such that |x − r/s|p < 1. By the
ultrametric triangle inequality, |r/s|p ≤ 1, so that we can assume that p - s. Let
a ∈ {0, 1, . . . , p − 1} be such that as ≡ r mod p. Then
|x − a|p = (x − r/s) + (r/s − a) p
≤ max{|x − r/s|p , |r/s − a|p } < 1 ,

since |r/s − a|p = |r − as|p /|s|p < 1. q

It follows that that Z is dense in Zp : iterating the statement of the lemma, we

find that for x ∈ Zp and n > 0 there is a ∈ Z (with 0 ≤ a < pn ) such that
|x − a|p < p−n .
 § 2. p-adic numbers 7

2.11. Theorem. The ring Zp is compact with respect to the metric induced by THM
the p-adic absolute value. In particular, Qp is locally compact. Zp is compact

Proof. We show that any sequence (xn ) in Zp has a convergent subsequence. Since
Zp is complete (Zp = {x ∈ Qp : |x|p ≤ 1} is a closed subset of the complete
space Qp ), it is enough to show that there is a subsequence that is a Cauchy
sequence. We do this iteratively. Let n1 be the smallest index n such that there are
infinitely many m > n with |xn − xm |p < 1. Such an n must exist by Lemma 2.10,
which implies that there is some a such that |xn − a|p < 1 for infinitely many n.
Now let n2 be the smallest index n > n1 such that |xn − xn1 |p < 1 and such
that there are infinitely many m > n with |xn − xm | < p−1 . This again exists by
Lemma 2.10, where we restrict to the infinitely many n such that |xn − xn1 |p < 1.
We continue in this way: nk+1 is the smallest n > nk such that |xn −xnk |p < p−(k−1)
and such that there are infinitely many m > n with |xn − xm |p < p−k . Then
|xnk+1 − xnk |p < p−(k−1) for all k ≥ 1, which by the ultrametric triangle inequality
implies that (xnk )k is a Cauchy sequence:
n o
xnk+l − xnk p ≤ max xnk+m − xnk+m−1 p : 1 ≤ m ≤ l < p−(k−1) .
That Qp is locally compact follows, since any x ∈ Qp has the compact neighbor-
hood x + Zp . q

2.12. Definition. Let (K, |·|) be a complete field with a non-archimedean abso- DEF
lute value. By Lemma 2.2 the valuation ring R = {x ∈ K : |x| ≤ 1} is a local ring residue field
with unique maximal ideal M = {x ∈ K : |x| < 1}. The quotient ring k = R/M
is therefore a field, the residue field of K. The canonical map R → k is called the
reduction map and usually denoted x 7→ x̄. ♦

Before we continue we note an easy but important fact on non-archimedean abso-

lute values.

2.13. Lemma. Let (K, |·|) be a field with a non-archimedean absolute value and LEMMA
let x, y ∈ K with |x| > |y|. Then |x + y| = |x|. all triangles
If we have x , x , . . . , x ∈ K such that x + x + . . . + x = 0 and n ≥ 2, then are isosceles
1 2 n 1 2 n
there are at least two indices 1 ≤ j < k ≤ n such that
|xj | = |xk | = max{|x1 |, |x2 |, . . . , |xn |} .

Proof. We have |x + y| ≤ max{|x|, |y|} = |x|. Assume that |x + y| < |x|. Then
|x| = |(x + y) + (−y)| ≤ max{|x + y|, | − y|} < |x| ,
a contradiction.
For the second statement observe that if we had just one j such that |xj | is
maximal, then by the second statement we would have
X
0 = |x1 + x2 + . . . + xn | = xj + xk = |xj | ,
k6=j

which is a contradiction, since xj 6= 0 in this situation. q

If |·| is a non-archimedean absolute value on a field K, then we can extend it to

the polynomial ring K[x].
 § 2. p-adic numbers 8

2.14. Lemma. Let (K, |·|) be a field with a non-archimedean absolute value. For LEMMA
f = a0 + a1 x + a2 x2 + . . . + an xn ∈ K[x] we set extension of
abs. value
|f | := max{|a0 |, |a1 |, |a2 |, . . . , |an |} .
to K[x]
0
Then properties (1), (2) and (3 ) of Definition 2.1 are satisfied for elements
of K[x].
If K is complete with respect to |·|, then for each n ∈ Z≥0 , the space K[x]<n of
polynomials of degree < n is a complete metric space for the metric induced by |·|
on K[x].

Proof. Exercise. q

The following is an important tool when working in complete non-archimedean

fields.

2.15. Theorem. Let (K, |·|) be a complete non-archimedean field with valuation THM
ring R and residue field k. Let F ∈ R[x] be a polynomial such that |F | = 1 Hensel’s
and suppose that we have a factorization F̄ = f1 f2 in k[x] such that f1 is monic Lemma
and f1 and f2 are coprime. (The notation F̄ means the polynomial in k[x] whose
coefficients are obtained by reduction from those of F .) Then there are unique
polynomials F1 , F2 ∈ R[x] such that F = F1 F2 , F1 is monic and F̄1 = f1 , F̄2 = f2 .

Proof. Let n = deg F , n1 = deg f1 , n2 = n − n1 ≥ deg f2 . We write R[x]<n for the

R-module of polynomials of degree < n.
(0) (0) (0) (0)
We choose polynomials F1 , F2 ∈ R[x] such that deg F1 = n1 , deg F2 = n2 ,
(0) (0) (0) (0)
F̄1 = f1 , F̄2 = f2 , F1 is monic and the leading coefficient of F2 is the same
(0) (0)
as that of F ; then F − F1 F2 ∈ R[x]<n and
(0) (0)
δ := |F − F1 F2 | < 1 .
We claim that the k-linear map
k[x]<n1 × k[x]<n2 −→ k[x]<n , (h1 , h2 ) 7−→ h1 f2 + h2 f1
is an isomorphism: it is injective, since h1 f2 + h2 f1 = 0 implies that f1 divides
h1 f2 , which in turn implies that f1 divides h1 (since f1 and f2 are coprime). But
deg h1 < n1 = deg f1 , so h1 = 0. Since f1 6= 0, it follows that h2 = 0, too. Finally
we observe that the dimensions of source and target are the same.
If M is the matrix representing this linear map with respect to the k-bases
((1, 0), (x, 0), . . . (xn1 −1 , 0), (0, 1), (0, x), . . . , (0, xn2 −1 )) and (1, x, x2 , . . . , xn1 +n2 −1 ),
then det(M ) 6= 0. Let now
(0) (0)
Φ : R[x]<n1 × R[x]<n2 −→ R[x]<n , (H1 , H2 ) 7−→ H1 F2 + H2 F1 .
Its matrix M̃ with respect to the ‘power bases’ reduces to M , so | det(M̃ )| = 1,
which means that M̃ is in GL(n, R), and Φ is invertible. It also follows that if
(H1 , H2 ) = Φ−1 (H), then |H1 |, |H2 | ≤ |H|.
(0) (0)
We now want to find F1 and F2 by adjusting F1 and F2 . So we would like to
determine H1 ∈ R[x]<n1 and H2 ∈ R[x]<n2 with |H1 |, |H2 | < 1 such that
(0) (0) (0) (0) (0) (0)
F = (F1 + H1 )(F2 + H2 ) = F1 F2 + H1 F2 + H2 F1 + H1 H2 .
 § 2. p-adic numbers 9

We ignore the nonlinear term H1 H2 and choose (H1 , H2 ) such that the linear terms
(0) (0)
correct the mistake, i.e., (H1 , H2 ) = Φ−1 (F − F1 F2 ). From the above we know
that |H1 |, |H2 | ≤ δ and therefore
(0) (0)
F − (F1 + H1 )(F2 + H2 ) = |H1 H2 | ≤ δ 2 .
(1) (0) (2)
Repeating this with the new approximations Fj = Fj + Hj we obtain Fj such
(1) (2) (2) (2)
that |Fj − Fj | ≤ δ 2 and |F − F1 F2 | ≤ δ 4 . Iterating this procedure, we
(m)
construct sequences (Fj )m≥0 in R[x]<nj such that
(m) (m+1) m (m) (m+1) m (m) (m) m
|F1 − F1 | ≤ δ2 , |F2 − F2 | ≤ δ2 and |F − F1 F2 | ≤ δ2
for all m. Since R[x]<nj is complete by Lemma 2.14, the sequences converge to
polynomials F1 and F2 with F = F1 F2 and F̄1 = f1 , F̄2 = f2 . This shows existence.
To show uniqueness, assume that F̃1 and F̃2 are another solution. Then
0 = F − F = F1 F2 − F̃1 F̃2 = (F1 − F̃1 )F2 + (F2 − F˜2 )F̃1 .
The map Φ as above, but using F̃1 and F2 , is still invertible (this only uses that the
reductions are f1 and f2 ), which immediately gives F1 −F̃1 = 0 and F2 −F̃2 = 0. q

We draw some conclusions from this.

2.16. Corollary. Let (K, |·|) be a field that is complete with respect to a non- COR
archimedean absolute value, let R be its valuation ring and k its residue field and Hensel’s
let f ∈ R[x] be monic. Assume that f¯ ∈ k[x] has a simple root a ∈ k. Then f has Lemma
a unique root α ∈ R such that ᾱ = a. for roots

Proof. This is the case f1 = x − a of Theorem 2.15. Note that the assumption
that a is a simple root of f¯ implies that the cofactor f2 = f¯/(x − a) is coprime
to f1 . q

2.17. Corollary. Let (K, |·|) be a field that is complete with respect to a non- COR
archimedean absolute value and let f = a0 + a1 x + . . . + an xn ∈ K[x] with an 6= 0. reducibility
Assume that there is 0 < m < n with |am | = |f | and that |a0 | < |f | or |an | < |f |. of certain
Then f is reducible. polynomials

Proof. After scaling f we can assume that |f | = 1. If |a0 | < 1, then let m be
minimal with |am | = 1. Then f¯ = xm f2 with f2 (0) 6= 0, so that f1 = xm and f2
are coprime. By Theorem 2.15 there is a factorization f = F1 F2 with deg F1 = m.
Since 0 < m < n = deg f , this shows that f is reducible.
If |an | < 1, then let m be maximal with |am | = 1. Then f¯ = f1 · ām with f1 monic
and (trivially) coprime to f2 = ām . Again by Theorem 2.15 there is a factorization
f = F1 F2 with deg F1 = m. Since again 0 < m < n = deg f , this shows that f is
reducible also in this case. q

The following example really belongs right after Definition 2.12.

2.18. Example. The residue field of Qp is Fp . This essentially follows from EXAMPLE
Lemma 2.10; the details are left as an exercise. ♣ Qp has Fp as
residue field
We will now look at field extensions of complete fields with absolute value. First
we introduce the norm and trace of an element in a finite field extension.
 § 2. p-adic numbers 10

2.19. Definition. Let K ⊂ L be a finite field extension and let α ∈ L. Then DEF
multiplication by α induces a K-linear map mα : L → L. We define the norm N(α) norm and
and trace Tr(α) of α to be the determinant and trace of mα , respectively. trace
If we want to make clear which fields are involved, we write NL/K (α) and TrL/K (α).
♦

Norms and traces have the following properties.

• If α ∈ K, then clearly NL/K (α) = α[L:K] (since mα can be taken to be αI[L:K] ).
• If K ⊂ L is separable, then we can also write
Y X
N(α) = σ(α) and Tr(α) = σ(α) ,
σ : L,→K̄ σ : L,→K̄

where σ runs through all embeddings of L into a fixed algebraic closure K̄ of K.

(This is because the σ(α) are the eigenvalues of mα .)
• Norms and traces are transitive: if K ⊂ L ⊂ L0 , then for α ∈ L0 we have
NL/K (NL0 /L (α)) = NL0 /K (α) and TrL/K (TrL0 /L (α)) = TrL0 /K (α).
• The norm is multiplicative and the trace is additive (even K-linear); this follows
from the corresponding properties of determinants and traces of linear endomor-
phisms.
• If L = K(α), then the characteristic polynomial of mα agrees with the minimal
polynomial of α, so its constant term is ± N(α).

2.20. Lemma. Let (K, |·|) be a field that is complete with respect to a non- LEMMA
archimedean absolute value and let R be its valuation ring. Let K ⊂ L be a finite integrality
field extension and let α ∈ L have norm N(α) ∈ R. Then α is integral over R,
i.e., α is a root of a monic polynomial with coefficients in R.

0
Proof. Let L0 = K(α) ⊂ L. Then N(α) = NL0 /K (α)[L:L ] . Let f ∈ K[x] be the
minimal polynomial of α. Its constant term a0 is ± NL0 /K (α) and so the assumption
implies |a0 | ≤ 1. If we had |f | > 1, then f would be reducible by Corollary 2.17,
a contradiction. So |f | = 1, meaning that f ∈ R[x], and α is a root of the monic
polynomial f . q

2.21. Corollary. In the situation of Lemma 2.20 we have | N(1 + α)| ≤ 1. COR
bound for
N(1 + α)
Proof. We know that α is a root of a monic polynomial f ∈ R[x]. Then 1 + α is
a root of the monic polynomial f (x − 1) ∈ R[x], and since a power of its constant
term, which is in R, is ± N(1 + α), it follows that | N(1 + α)| ≤ 1. q

2.22. Theorem. Let (K, |·|) be a complete field with a nontrivial non-archimedean THM
absolute value and let L be an algebraic extension of K. Then there is a unique extensions
absolute value |·|0 on L that extends |·|. If the extension K ⊂ L is finite, then of complete
(L, |·|0 ) is complete. fields
 § 2. p-adic numbers 11

Proof. Since any algebraic extension of K can be obtained as an increasing union

of finite extensions, it suffices to consider the finite case. So let [L : K] = n. We
first show existence. To this end we define |α|0 = | N(α)|1/n . Then it is clear that
|·|0 satisfies properties (1) and (2) of Definition 2.1. To show property (30 ), let
α, β ∈ L and assume that |α|0 ≥ |β|0 . Then | N(β/α)| ≤ 1, hence by Corollary 2.21
n
|α + β|0 = | N(α + β)| = | N(α)| · | N(1 + β/α)| ≤ | N(α)| = (max |α|0 , |β|0 )n .
It is also clear that |·|0 agrees with |·| on K.
To show uniqueness, let |·|00 be another absolute value on L extending |·|. Let
α ∈ L with |α|0 ≤ 1, so that | N(α)| ≤ 1. Then the minimal polynomial f of α
is in R[x]. If we had |α|00 > 1, then by the ultrametric triangle inequality applied
to f (α) = 0 we would obtain a contradiction, since then αn would be the unique
term of maximal absolute value, compare Lemma 2.13. So we must have |α|00 ≤ 1.
If |α|0 ≥ 1, then |α−1 |0 ≤ 1, so |α−1 |00 ≤ 1 and |α|00 ≥ 1. This implies that
|α|0 < 1 ⇐⇒ |α|00 < 1, from which it follows that the two absolute values on L
are equivalent (compare Problem (1) on Exercise sheet 1). Since they have the
same restriction to K, they must then be equal.
It remains to show that (L, |·|0 ) is complete. Let (b1 , b2 , . . . , bm ) be K-linearly
independent elements of L. We claim that there are constants c, C > 0 (depending
on b1 , b2 , . . . , bm ) such that
(2.1) c max{|a1 |, . . . , |am |} ≤ |a1 b1 + . . . + am bm |0 ≤ C max{|a1 |, . . . , |am |}
for all a1 , a2 , . . . , am ∈ K. We prove this by induction on m. The cases m ≤ 1 are
trivial. So let m ≥ 2. The upper bound is easy:
|a1 b1 + . . . + am bm |0 ≤ max{|a1 b1 |0 , . . . , |am bm |0 }
≤ max{|b1 |0 , . . . , |bm |0 } · max{|a1 |, . . . , |am |} .
To prove the lower bound, we argue by contradiction. If there is no lower bound,
then there are a1 , . . . , am with max{|a1 |, . . . , |am |} = 1 and |a1 b1 + . . . + am bm |0
(k) (k)
arbitrarily small. Pick a sequence (a1 , . . . , am ) of such tuples so that
(k) 0 −k
(2.2) |a1 b1 + . . . + a(k)
m bm | < 2 .
(k)
By passing to a sub-sequence and scaling we can assume that aj = 1 for all k
and for some j, say j = 1. Taking differences, we see that
(k+1) (k) (k+1) −k 0
(a2 − a2 )b2 + . . . + (am − a(k)
m )bm < 2 for all k.
By induction, there is c > 0 such that |a2 b2 + . . . + am bm |0 ≥ c max{|a2 |, . . . , |am |}.
(k+1) (k) (k)
This implies that |aj − aj | ≤ c−1 2−k for all 2 ≤ j ≤ m, so (aj )k is a Cauchy
sequence and converges to a limit aj . Taking the limit in (2.2) we find that
b 1 + a2 b 2 + . . . + am b m = 0 ,
which contradicts the linear independence of b1 , b2 , . . . , bm .
Now let (b1 , b2 , . . . , bn ) be a K-basis of L, and let c > 0 be the associated constant.
If (αm ) is a Cauchy sequence in L, write αm = a1m b1 + . . . + anm bn . Then
|ajm0 − ajm | ≤ c−1 |αm0 − αm |0 ,
so each sequence (ajm )m is a Cauchy sequence in K and so converges to some
aj ∈ K. But then αm → a1 b1 + · · · + an bn converges as well. q
 § 2. p-adic numbers 12

We remark that (2.1) implies that the topology on L induced by |·|0 is the same as
the topology induced by any K-linear isomorphism L → K n , where K n has the
product topology.
Since |·|0 is uniquely determined and extends |·|, one simply writes |·| for the
absolute value on L.
We note that it can be shown that a field that is complete with respect to an
archimedean absolute value must be isomorphic to R or C (with the usual absolute
value) and that a locally compact (and then necessarily complete) non-archimedean
field of characteristic 0 must be isomorphic to a finite extension of Qp for some
prime number p. (In characteristic p, it will be isomorphic to a finite extension of
the field Fp ((t)) of formal Laurent series over Fp , which is the field of fractions of
the ring of formal power series Fp [[t]].)

2.23. Corollary. Let Q̄p be the algebraic closure of Qp . Then Q̄p has a unique COR
absolute value extending |·|p on Qp . If σ ∈ Aut(Q̄p /Qp ) is any automorphism, Q̄p has unique
then |σ(α)|p = |α|p for all α ∈ Q̄p . abs. value

Proof. Q̄p is an algebraic extension of Qp , so Theorem 2.22 applies. The second

statement follows from the uniqueness of the absolute value, since α 7→ |σ(α)|p is
another absolute value on Q̄p extending the p-adic absolute value on Qp . q

Next we want to show that Q̄p is not complete (contrary to the algebraic closure C
of the completion R of Q with respect to the usual absolute value).

2.24. Lemma. Q̄p is not complete with respect to |·|p . LEMMA

Q̄p not
n complete
Proof. For n ≥ 1, let ζn ∈ Q̄p be a primitive (p2 − 1)-th root of unity. It follows
from properties of finite fields that ζn is a root of a monic polynomial fn ∈ Zp [x]
of degree 2n that reduces to an irreducible polynomial in Fp [x]; in particular, fn
is irreducible itself. Since ζn is a power of ζn+1 , we obtain a tower of fields
Qp = Qp (ζ1 ) ⊂ Qp (ζ2 ) ⊂ Qp (ζ3 ) ⊂ . . . ⊂ Q̄p

P∞ thatn [Qp (ζn+1 ) : Qp (ζn )] = 2 for all n ≥ 1. We now consider the series
such
n=1 ζn p . If Q̄p were complete, then series would converge (since |ζn |p = 1).
So we will show that the series does not converge in Q̄p . The proof will be by
contradiction. So we assume that the series has a limit α ∈ Q̄p . Then m :=
[Qp (α) : Qp ] is finite. Let σ ∈ Aut(Q̄p /Qp ). By Corollary 2.23, σ is continuous
with respect to the p-adic topology. This implies that
X∞
σ(α) = σ(ζn )pn .
n=2

Since [Qp (α) : Qp ] = m, there are exactly m values that σ(α) can take (namely,
the roots of the minimal P polynomial of α over Qp ). Now pick some n such that
2 > m. Note that sn = nk=1 ζk pk generates the field Qp (ζn ) (this is easily seen
n

by induction). So, as σ runs through the automorphisms of Q̄p over Qp , σ(sn )

takes exactly 2n distinct values. Since the various possible values of σ(ζn ) all
differ mod p (this is because fn has only simple roots in F̄p ), we conclude that the
various values of σ(sn ) all differ mod pn+1 . On the other hand,
σ(α) ≡ σ(sn ) mod pn+1 ,
 § 2. p-adic numbers 13

and there are only m < 2n possibilities for the left hand side. This gives the
desired contradiction. q

2.25. Definition. We define Cp to be the completion of Q̄p . ♦ DEF

Cp
2.26. Lemma. Let (K, |·|) be an algebraically closed field with a non-archimedean LEMMA
absolute value, and let K 0 be its completion. Then K 0 is also algebraically closed. completion
stays
Proof. Let K ⊂ L be a finite field extension and let α ∈ L; after scaling α by an alg. closed
0

element of K, we can assume that |α| ≤ 1 (recall that L has a unique absolute
value extending that of K 0 ). We must show that α ∈ K 0 . Let f ∈ K 0 [x] be the
minimal polynomial of α. Since K is dense in K 0 , there is a sequence (fn ) of monic
polynomials in K[x] such that |fn − f | < 2−n . Since K is algebraically closed,
each fn splits into linear factors. By the ultrametric triangle inequality (and using
|α| ≤ 1), we have
Y
|α − α0 | = |fn (α)| = |fn (α) − f (α)| ≤ |fn − f | < 2−n .
α0 : fn (α0 )=0

There must then be a root αn of fn such that |α − αn | < 2−n/ deg(f ) . This implies
that (αn ) converges in L to α. But K 0 is the closure of K, therefore the limit α
must already be in K 0 . q

So Cp is the unique (up to isomorphism) minimal extension of Qp that is complete

and algebraically closed. So in this sense, Cp is the p-adic analogue of the field of
complex numbers.
The residue field of Q̄p and of Cp is F̄p . Since this is infinite, it follows that nei-
ther of these two fields is locally compact: the closed ball around zero of radius 1
contains infinitely many elements whose pairwise distance is 1 (a system of rep-
resentatives of the residue classes), and so no sub-sequence of a sequence of such
elements can converge.
 § 3. Newton Polygons 14

3. Newton Polygons

Let (K, |·|) be a complete field with a nontrivial non-archimedean absolute value.
Consider a polynomial 0 6= f ∈ K[x]. Let α ∈ K̄ be a root of f . We have seen
that there is a unique extension of |·| to K̄, so it makes sense to consider |α|. We
will now describe a method that determines the absolute values of the roots of f
and how often they occur.
In this context it is advantageous to switch to a ‘logarithmic’ version of the absolute
value. We fix a positive real number c and set
v(x) = −c log |x| for x ∈ K × and v(0) = +∞ .

3.1. Definition. The map v : K → R ∪ {+∞} is the (additive) valuation asso- DEF
ciated to |·|. ♦ valuation

Of course, changing c will scale v by a positive factor, so v is not uniquely deter-

mined.
Corresponding to the properties of absolute values, the valuation satisfies

(1) v(x) < ∞ if x ∈ K × ;

(2) v(xy) = v(x) + v(y) for all x, y ∈ K;
(3) v(x + y) ≥ min{v(x), v(y)} for all x, y ∈ K, with equality when v(x) 6= v(y).

The adjective ‘additive’ refers to the second property.

When dealing with p-adic fields like Qp , Q̄p or Cp , we choose c = 1/ log p; then
v = vp on Qp , so v(Q× × ×
p ) = Z and (as is easily seen) v(Q̄p ) = v(Cp ) = Q.

Now let α ∈ K̄ be a root of f = a0 + a1 x + . . . + an xn ∈ K[x] with an 6= 0. Then

a0 + a1 α + a2 α 2 + . . . + an α n = 0 ,
so by Lemma 2.13 there must be (at least) two terms in the sum whose absolute
value is maximal, or equivalently, whose valuation is minimal. These valuations
are
v(a0 ), v(a1 ) + v(α), v(a2 ) + 2v(α), . . . , v(an ) + nv(α) .
So we need to have 0 ≤ k < m ≤ n such that
v(ak ) + kv(α) = v(am ) + mv(α) ≤ v(aj ) + jv(α) for all 0 ≤ j ≤ n.
This is equivalent to saying that the line of slope −v(α) through the points
(k, v(ak )) and (m, v(am )) in the plane has no point (j, v(aj )) below it. This
prompts the following definition.

3.2. Definition. Let 0 6= f = a0 + a1 x + . . . + an xn ∈ K[x] with K as above. DEF

The Newton polygon of f is the lower convex hull of the set of the points (j, v(aj )) Newton
for 0 ≤ j ≤ n such that aj 6= 0, i.e., the union of all line segments joining two polygon
of these points and such that the line through these points does not run strictly
above any of the other points. A maximal such line segment is a segment of the
Newton polygon; it has a slope (which is just its usual slope) and a length, which
is the length of its projection to the x-axis. ♦
 § 3. Newton Polygons 15

3.3. Example. Consider the polynomial EXAMPLE

5 4 3 2 Newton
f = x + 3x + 4x + 6x + 8 ∈ Q2 [x] .
polygon
The points (j, v(aj )) with aj 6= 0 are
(0, 3), (2, 1), (3, 2), (4, 0), (5, 0) .
We find three segments, forming a broken line with vertices (0, 3), (2, 1), (4, 0),
(5, 0). The slopes are −1, −1/2 and 0, and the lengths are 2, 2 and 1.

0 1 2 3 4 5
♣

What we did above amounts to the statement that the valuations of the roots of f
are among the slopes taken negatively of the Newton polygon. We now want to
prove a converse and give a more precise statement. We first introduce a variation
of the absolute value on the polynomial ring.

3.4. Definition. Let (K, |·|) be a field with a nontrivial non-archimedean abso- DEF
lute value and let r > 0. For f = a0 + a1 x + . . . + an xn ∈ K[x] we define |·|r on K[x]
|f |r := max{|aj |rj : 0 ≤ j ≤ n} .
When f 6= 0, then we set
`r (f ) = max{j : |aj |rj = |f |r } − min{j : |aj |rj = |f |r } ∈ Z≥0 . ♦

We observe that `r (f ) is strictly positive if and only if c log r is the slope of a

segment of the Newton polygon of f , and that in this case `r (f ) is the length of
the corresponding segment.
One can easily adapt the proof for the case r = 1 to show that |·|r is an abso-
lute value on K[x] (in the sense that it satisfies properties (1), (2) and (30 ) of
Definition 2.1). It also restricts to |·| on K.

3.5. Lemma. Let f, g ∈ K[x] be nonzero polynomials and let r > 0. Then we LEMMA
have additivity
`r (f g) = `r (f ) + `r (g) . of `r

Proof. If f = a0 + a1 x + . . . + an xn , write
n− (f ) = min{j : |aj |rj = |f |r } and n+ (f ) = max{j : |aj |rj = |f |r }
and similarly for g and f g. Precisely as in the proof of property (2) for |·|r (i.e.,
basically Gauss’ Lemma) one sees that
n− (f g) = n− (f ) + n− (g) and n+ (f g) = n+ (f ) + n+ (g) ,
 § 3. Newton Polygons 16

which implies the claim, since `r (f ) = n+ (f ) − n− (f ) and similarly for g and f g.

q

3.6. Theorem. Let 0 6= f ∈ K[x] and let r > 0. Then the number of roots α ∈ K̄ THM
of f (counted with multiplicity) such that |α| = r is exactly `r (f ). roots and
Newton
In terms of the Newton polygon, this says that f has roots of valuation s if and polygon
only if its Newton polygon has a segment of slope −s, and the number of such
roots (counted with multiplicity) is exactly the length of the segment.

Proof. We can assume that K = K̄. The proof is by induction on the degree of f .
If f is constant, there is nothing to show. If f = x − α, then `r (f ) = 0 if |α| =
6 r
and `r (f ) = 1 if |α| = r.
Now assume f is not constant and let β ∈ K̄ be a root of f . Then f = (x − β)f1
for some 0 6= f1 ∈ K[x]. By the inductive hypothesis, the number of roots α of f1
such that |α| = r is `r (f1 ). If |β| = r, then the number of such roots of f is
`r (f1 ) + 1 = `r (f1 ) + `r (x − β) = `r (f ) (using Lemma 3.5). If |β| = 6 r, then the
number of such roots of f is `r (f1 ) = `r (f1 ) + `r (x − β) = `r (f ) again. q

3.7. Lemma. Let f ∈ K[x] be irreducible. Then the Newton polygon of f consists LEMMA
of a single segment. Newton
polygon of
Proof. We assume for simplicity that K has characteristic zero. Then K̄ is sepa- irreducible
rable over K and the roots of f form one orbit under the Galois group Aut(K̄/K). polynomial
Since the absolute value is invariant under the action of this group, we see that all
roots of f have the same absolute value. The claim then follows from Theorem 3.6.
n
(In characteristic p, let α be a root of f . Then for some n ≥ 0, αp is separable
n
over K and f = g(xp ) for an irreducible polynomial g ∈ K[x]. The previous
argument applies to g, but this then implies the claim also for f .) q

This leads to the following consequence.

3.8. Lemma. Let 0 6= f ∈ K[x] with f (0) 6= 0 and let σ1 , . . . , σm be the segments LEMMA
of the Newton polygon of f . Then there is a factorization slope
factorization
f = f1 f2 · · · fm
such that the Newton polygon of fj is a single segment with the same slope and
length as σj .

Proof. We can assume that f is monic. Let f = h1 · · · hn be the factorization of f

into monic irreducible polynomials over K. By Lemma 3.7, the Newton polygon
of each hj consists of a single segment. Let s1 , s2 , . . . , sm be the distinct slopes
that occur for these segments; we can number them so that sj is the slope of σj .
We then define fj to be the product of the hi whose slope is sj . The claim follows
from the additivity of the lengths of segments of the same slope, Lemma 3.5. q

We extend |·|r to an absolute value on the field K(x) of rational functions in one
variable over K in the usual way. For a given rational function f , we can then
study how |f |r varies with r.
 § 3. Newton Polygons 17

3.9. Theorem. Let f ∈ K(x)× . Then the function THM

variation of
ϕf : R −→ R , s 7−→ −c log |f |e−s/c |f |r with r
is piecewise affine with integral slopes. If ∂− ϕf (s) − ∂+ ϕf (s) = ν, then ν is the
number of zeros α of f with v(α) = s minus the number of poles α of f with
v(α) = s (each counted with multiplicity).

Here
ϕf (s + ε) − ϕf (s) ϕf (s) − ϕf (s − ε)
∂+ ϕf (s) = lim and ∂− ϕf (s) = lim
ε&0 ε ε&0 ε
are the right and left derivatives of the piecewise affine function ϕf .
Qn ej
Proof. We can write f = γ j=1 (x − αj ) for some γ ∈ K, αj ∈ K̄ and ej ∈ Z, so
n
X n
X
ϕf (s) = −c log |γ| − c ej log |x − αj |e−s/c = v(γ) + ej ϕx−αj (s) ,
j=1 j=1

and it suffices to prove the claim for f = x − α. In this case, we have

ϕx−α (s) = −c log max{e−s/c , |α|} = min{s, v(α)}
(recall that v(0) = +∞). This is a piecewise affine function with slopes 1 and 0
(unless α = 0). If α = 0, the slope is constant, so ν = 0 for all s, and there are no
zeros or poles α with v(α) = s. If α 6= 0, then the slope changes from 1 to 0 at
s = v(α), so ν = 1 there and ν = 0 for all other s. q

3.10. Example. For the polynomial EXAMPLE

ϕf for f from
f = x5 + 3x4 + 4x3 + 6x2 + 8 ∈ Q2 [x] Example 3.3
from Example 3.3, the graph of ϕf looks as follows.

3 0
2
2

1 4

0 1/2 1 3/2 2
-1
5

-2
 § 3. Newton Polygons 18

The green lines are the graphs of s 7→ v(aj ) + js (for j = 0, 2, 3, 4, 5); ϕf is the
minimum of these functions. The segments of the graph of ϕf correspond to the
vertices of the Newton polygon of f and the vertices of the graph of ϕf correspond
to the segments of the Newton polygon of f . ♣

We give another characterization of |f |r .

3.11. Lemma. Let (K, |·|) be a complete non-archimedean field, let r > 0 and LEMMA
0 6= f ∈ K[x]. Then |·|r as
|f |r = sup{|f (α)| : α ∈ K̄, |α| ≤ r} . supremum
norm

Proof. Let f = a0 + a1 x + . . . + an xn . If |α| ≤ r, then

|f (α)| = |a0 + a1 α + . . . + an αn | ≤ max{|a0 |, |a1 |r, . . . , |an |rn } = |f |r .
Conversely, for any ε > 0 there is r − ε < ρ ≤ r such that ρ ∈ |K̄ × | and c log ρ
is not a slope of the Newton polygon of f (recall that |K̄ × | is dense in R>0 ; there
are only finitely many slopes). Then for any α ∈ K̄ such that |α| = ρ, there is a
unique term aj αj of maximal absolute value and therefore we have |f (α)| = |f |ρ .
Letting ρ tend to r, we find |f |r ≤ sup{|f (α)| : α ∈ K̄, |α| ≤ r}. q
 § 4. Multiplicative seminorms and Berkovich spaces 19

4. Multiplicative seminorms and Berkovich spaces

The absolute values |·|r on the polynomial ring K[x] that we have studied in the
last section are examples of ‘multiplicative seminorms’.

4.1. Definition. Let K be a field with absolute value |·| and let A be a K-algebra DEF
(i.e., A is a ring with a ring homomorphism K → A that gives A a compatible multiplicative
structure as a K-vector space). A multiplicative seminorm on A is a map seminorm
A −→ R≥0 , a 7−→ kak Banach
algebra
such that
(1) k·k restricts to |·| on K;
(2) ka + bk ≤ kak + kbk for all a, b ∈ A;
(3) kabk = kak · kbk for all a, b ∈ A.
If in addition a = 0 is the only element with kak = 0, then k·k is a multiplicative
norm (which is the same as an absolute value on A extending |·|). In general, we
call ker k·k = {a ∈ A : kak = 0} the kernel of k·k; it is a prime ideal in A.
A K-algebra A with a fixed multiplicative norm such that A is complete with
respect to this norm is a Banach algebra over K. ♦

A seminorm on A only needs to satisfy kabk ≤ kak · kbk with equality for a ∈ K;
similarly for a norm.
If |·| is non-archimedean, then k·k also satisfies the ultrametric triangle inequality,
since then
n  
n n
X n j n−j
ka + bk = k(a + b) k = ab
j=0
j
n   n
X n j n−j
X
≤ kak kbk ≤ kakj kbkn−j ≤ (n + 1)(max{kak, kbk})n
j=0
j j=0
√
and n n + 1 → 1 as n → ∞.
We introduce the following notation.

4.2. Definition. Let (K, |·|) be a field with absolute value, let a ∈ K and r ≥ 0. DEF
Then D(a, r)
D(a, r) := DK (a, r) := {ξ ∈ K : |ξ − a| ≤ r}
is the closed disk of radius r around a. ♦

4.3. Examples. Let (K, |·|) be a complete non-archimedean field. EXAMPLES

multiplicative
(1) For any a ∈ K, the map seminorms
f 7−→ kf ka,0 := |f (a)| on K[x]
is a multiplicative seminorm on K[x].
(2) For any a ∈ K and any r > 0, the map
f 7−→ kf ka,r := |f (x + a)|r
is a multiplicative (semi)norm on K[x].
 § 4. Multiplicative seminorms and Berkovich spaces 20

(3) Let (an ) be a sequence in K and (rn ) a strictly decreasing sequence in R>0
such that D(an+1 , rn+1 ) ⊂ D(an , rn ) for all n. Then
f 7−→ kf k = lim kf kan ,rn
n→∞

is a multiplicative seminorm on K[x].

(1) is clear and (2) follows from the properties of |·|r . For (3) note that by
Lemma 3.11,
kf kan+1 ,rn+1 = sup{|f (α)| : α ∈ K̄, |α − an+1 | ≤ rn+1 }
≤ sup{|f (α)| : α ∈ K̄, |α − an | ≤ rn } = kf kan ,rn ,
so that the sequence (kf kan ,rn ) decreases, hence must have a limit. Properties (2)
and (3) from Definition 4.1 then follow by taking the limit in the corresponding
relations for the k·kan ,rn . ♣

In fact, this is essentially the full story, at least when K is algebraically closed and
complete.

4.4. Theorem. Let (K, |·|) be a complete and algebraically closed non-archimedean THM
field and let k·k be a multiplicative seminorm on K[x]. Then there is a decreasing classification
nested sequence of disks D(an , rn ) such that of mult.
seminorms
kf k = lim kf kan ,rn
n→∞ on K[x]
for all f ∈ K[x].

Proof. Let D = {D(a, r) : a ∈ K, r > 0, k·k ≤ k·ka,r } be the set of all closed
disks such that k·k is bounded above by the corresponding seminorm. Then for
all a ∈ K we have D(a, kx − ak) ∈ D: if f (x + a) = a0 + a1 x + . . . + an xn , then
kf k = ka0 + a1 (x − a) + . . . + an (x − a)n k
≤ max{|a0 |, |a1 |kx − ak, . . . , |an |kx − akn }
= |f (x + a)|kx−ak = kf ka,kx−ak .
In particular, D is non-empty. Conversely, we have |x − a|a,r = r, which implies
that if D(a, r) ∈ D, then r ≥ kx − ak. So
D = {D(a, r) : a ∈ K, r ≥ kx − ak} .

We claim that D does not contain two disjoint disks: assume that D1 = D(a1 , r1 )
and D2 = D(a2 , r2 ) are in D. Then we have kx − a1 k ≤ r1 and kx − a2 k ≤ r2 , so
that
|a1 − a2 | = k(x − a2 ) − (x − a1 )k ≤ max{r1 , r2 } .
If r1 ≤ r2 , then this implies that a1 ∈ D2 and then that D1 ⊂ D2 ; if r2 ≤ r1 , then
we see in the same way that D2 ⊂ D1 .
Now let ρ = inf{r > 0 : ∃a ∈ K : D(a, r) ∈ D} and choose sequences (an ) in K
and (rn ) in R>0 such that (rn ) is strictly decreasing with rn → ρ and D(an , rn ) ∈ D
for all n. Then we have D(an+1 , rn+1 ) ⊂ D(an , rn ). This implies
kf k ≤ lim kf kan ,rn .
n→∞
 § 4. Multiplicative seminorms and Berkovich spaces 21

We still have to show the reverse inequality. Since every nonzero f is a product
of a constant and polynomials of the form x − a, it suffices to prove this for the
latter. If |a − an0 | > rn0 for some n0 , then
kx − ak = k(x − an0 ) − (a − an0 )k = |a − an0 |
= lim max{|a − an |, rn } = lim kx − akan ,rn ,
n→∞ n→∞

since kx − an0 kT≤ rn0 < |a − an0 |. Otherwise, we have |a − an | ≤ rn for all n, which
says that a ∈ n D(an , rn ), so that D(an , rn ) = D(a, rn ) and we have
lim kx − akan ,rn = lim kx − aka,rn = kx − aka,ρ = ρ .
n→∞ n→∞

Since D(a, kx − ak) ∈ D and ρ is the smallest possible radius of a disk in D, we

get kx − ak ≥ ρ as required. q
T
Note that if a ∈ n D(an , rn ) in the proof above, then D(a, ρ) ∈ D, since
kf k ≤ lim kf kan ,rn = lim kf ka,rn = kf ka,ρ
n→∞ n→∞

for all f ∈ K[x]. Since ρ ≤ r for any radiusT r of a disk in D and no two disks
in D are disjoint, it follows that D(a, ρ) = D. We can therefore distinguish the
following four types of multiplicative seminorms on K[x].

4.5. Definition. Let k·k be a multiplicative seminorm on K[x], where K is a DEF

complete and algebraically closed non-archimedean field. Let D be as in the proof types of
of Theorem 4.4. mult.
T seminorms
(1) k·k is of type 1 if D = {a} for some a ∈ K. Then kf k = |f (a)| and ker k·k
is the kernel of the evaluation map f 7→ f (a).
(2) k·k is of type 2 if D = D(a, r) for some a ∈ K and r > 0 such that r ∈ |K × |.
T
Then kf k = |f |a,r .
/ |K × |.
T
(3) k·k is of type 3 if D = D(a, r) for some a ∈ K and r > 0 such that r ∈
Then kf k = |f |a,r .
T
(4) Finally, k·k is of type 4 if D = ∅.
In the last three cases, k·k is actually a norm. ♦

We note that if ρ = 0 in the proof of TheoremT 4.4, then the completeness of K

implies that D = {a} for some a ∈ K. If D 6= ∅ for every decreasing nested
sequence of disks in K, then K is said to be spherically complete. (Then no DEF
multiplicative seminorms of type 4 exist.) For example, Qp is spherically complete, spherically
but Cp is not (Exercise). complete
For types 2, 3 and 4, k·k defines an absolute value on K[x]. We extend it to an
absolute value on the field of fractions K(x), which we can then complete to obtain
Hk·k , and we obtain a K-algebra homomorphism K[x] → Hk·k . For example, if
k·k = |·|r is of type 2 or 3 with a = 0, then the corresponding completion of K[x]
is the ring
nX ∞ o
−1 j j
Khr xi := aj x ∈ K[[X]] : lim |aj |r = 0
j→∞
j=0
of power series converging on D(0, r) (Exercise), so H|·|r is the field of fractions
of Khr−1 xi.
If k·k is of type 1, then the evaluation map at a gives us a K-algebra homo-
morphism K[x] → K =: Hk·k . In each case, we obtain k·k as the pull-back of
 § 4. Multiplicative seminorms and Berkovich spaces 22

the absolute value of a complete field H that is a K-Banach algebra via a K-

algebra homomorphism K[x] → H. Conversely, if we have such a homomorphism
K[x] → H, then pulling back the absolute value of H to K[x] will give us a
multiplicative seminorm on K[x].

4.6. Definition. Let K be a complete and algebraically closed non-archimedean DEF

field and let A be a finitely generated K-algebra, so that A is the coordinate ring Berkovich
of the affine K-variety X = Spec A. Then the Berkovich space associated to A space
or X is
Berk A := X an := {k·k : A → R multiplicative seminorm on A} ,
the set of multiplicative seminorms on A. The topology on X an is the weakest
topology that makes the maps X an → R, k·k 7→ kf k, continuous for all f ∈ A.
(Concretely, this means that any open set is a union of finite intersections of sets
of the form Uf,a,b = {k·k ∈ X an : a < kf k < b}.) ♦

We will usually call elements of X an points and denote them by ξ or similar. In

this context, the corresponding multiplicative seminorm will be written k·kξ and
the K-Banach algebra obtained by completion, Hξ .

4.7. Lemma. The topological space X an as defined in Definition 4.6 is Hausdorff. LEMMA
X an is
Hausdorff
Proof. Let ξ, ξ 0 ∈ X an be distinct. We must show that there are disjoint open sets
U, U 0 ⊂ X an with ξ ∈ U and ξ 0 ∈ U 0 . Since ξ 6= ξ 0 , there must be f ∈ A such
that kf kξ 6= kf kξ0 . Assume without loss of generality that kf kξ < kf kξ0 and let
kf kξ < a < kf kξ0 . Then we can take U = Uf,−∞,a and U 0 = Uf,a,∞ . q

Note that we can alternatively define Berk A as the set of all K-algebra homomor-
phisms A → H into complete K-Banach algebras that are fields, up to an obvious
equivalence. The topology is then that of pointwise convergence. This can be
seen as similar to the definition of Spec A as the set of all K-algebra homomor-
phisms into fields, up to an obvious equivalence. Here the topology is again that of
pointwise convergence, but with the cofinal topology on the target fields (so that
basic open sets are defined by relations f 6= 0). Since a Banach algebra has more
structure than just a K-algebra, there are fewer equivalences and therefore more
points in Berk A than in Spec A. This is made precise by the following statements.

4.8. Lemma. Let K, A and X be as in Definition 4.6. Then there is a canonical LEMMA
inclusion of X(K) into X an . The inclusion is continuous when X(K) is given the X(K) ,→ X an
topology induced by the absolute value on K.

Proof. Let ξ ∈ X(K). Then ξ gives rise to a K-algebra homomorphism A → K,

f 7→ f (ξ). We define ξ˜ ∈ X an to correspond to the multiplicative seminorm
f 7→ |f (ξ)|. Its kernel is the ideal of A consisting of functions vanishing at ξ and
so determines ξ. This gives us the desired inclusion.
To show that ξ 7→ ξ˜ is continuous, consider a basic open set Uf,a,b in X an . Its
preimage is {ξ ∈ X(K) : a < |f (ξ)| < b}. Since every f ∈ A is continuous
as a map X(K) → K (with respect to the K-topology) and |·| : K → R is also
continuous, this set is open. q
 § 4. Multiplicative seminorms and Berkovich spaces 23

It is in fact the case that the image of X(K) in X an is dense. We will see this
later when X is the affine line.
There is also a natural map in the other direction.

4.9. Lemma. Let K, A and X be as in Definition 4.6. Then there is a canonical LEMMA
continuous map Berk A → Spec A (where Spec A has the Zariski topology). Berk A
→ Spec A
Proof. Let ξ ∈ Berk A. Then k·kξ is the pull-back of the absolute value from some
K-Banach algebra H that is a field under a homomorphism φ : A → H. Since H
is a field, the kernel of φ must be a prime ideal of A and so defines an element
of Spec A (or we just take H as a field, forgetting the absolute value).
It remains to show that the map Berk A → Spec A obtained in this way is continu-
ous. Let Uf = {ξ ∈ Spec A : f (ξ) 6= 0} be a basic open set in Spec A. Its pull-back
to Berk A is {ξ ∈ Berk A : kf kξ 6= 0} = Uf,0,∞ and therefore open, too. q

The composition of the two maps X(K) → X an → Spec A is the inclusion of X(K)
(the set of maximal ideals of A) into Spec A (the set of prime ideals of A).
The map is actually surjective: let ξ ∈ Spec A; this gives us a K-algebra homo-
morphism A → R into an integral domain finitely generated over K. One can
show that one can always define an absolute value on such an R that extends the
absolute value on K. Pulling back to A, we obtain a multiplicative seminorm; this
is a preimage of ξ.
It is also true that X an is (even path-)connected when X is connected as an
algebraic variety and that X an is locally compact. We will see this concretely for
X = A1 in the next section.
 § 5. The Berkovich affine and projective line 24

5. The Berkovich affine and projective line

In the following, we will take a closer look at the Berkovich affine line over Cp ,
AC1,an
p
= Berk Cp [x]. According to the classification of Definition 4.5, we have four
types of points in A1,an 1
Cp . The type 1 points recover the points in A (Cp ) = Cp as in
Lemma 4.8. The type 2 and 3 points correspond to closed disks D(a, r) with r > 0
in, respectively, not in, the value group pQ of Cp . We will reserve the notation ζa,r
or ζD where D = D(a, r) for these points. We will frequently identify a with ζa,0 ,
however. Finally, the type 4 points correspond to nested sequences of disks with
empty intersection; these points are somewhat annoying, but don’t usually give
us problems. They are necessary to make the space locally compact. Two nested
sequences of disks (D(an , rn ))r and (D(a0n , rn0 ))n with empty intersection define
the same type 4 point if and only if D(an , rn ) ∩ D(a0n , rn0 ) 6= ∅ for all n.
Now fix ξ ∈ Cp and consider real numbers 0 ≤ r0 < r1 . Then there is a map

[r0 , r1 ] −→ AC1,an
p
, r 7−→ k·kξ,r = ζξ,r .

This map is continuous, since for any f ∈ Cp [x], the map

r 7−→ kf kξ,r = max{|aj |p rj : 0 ≤ j ≤ n}

is continuous, where f (x + ξ) = a0 + a1 x + . . . + an xn . The map is also clearly

injective. Since [r0 , r1 ] is compact, the map is actually a homeomorphism onto its
image. We write [ζξ,r0 , ζξ,r1 ] for this image.
Now consider two points ξ, η ∈ Cp and let δ = |ξ − η|p their distance. Then
D(ξ, δ) = D(η, δ) (“every point in a disk is a center”). Define

ξ ∨ η := ζξ,δ = ζη,δ .

Then
(
ζξ,r if 0 ≤ r ≤ δ,
γξ,η : [0, 2δ] −→ A1,an
Cp , r−
7 →
ζη,2δ−r if δ ≤ r ≤ 2δ

is a continuous path in A1,an

Cp joining ξ and η, whose image is [ξ, ξ ∨ η] ∪ [η, ξ ∨ η].

We can extend the definition of ξ ∨ η to arbitrary points in AC1,an

p
. We first observe
that
(
kx − ηkη,r = r if η ∈ D(ξ, r),
kx − ηkξ,r =
|ξ − η| > r if η ∈
/ D(ξ, r).

This shows that D(ξ, r) is uniquely determined by k·kξ,r and also shows that
D1 ⊂ D2 holds for two closed disks if and only if k·kD1 ≤ k·kD2 (the ‘only if’ part
follows also from the characterization of k·kξ,r as the sup norm on D(ξ, r)). This
prompts us to define

ξ ≤ ξ 0 ⇐⇒ ∀f ∈ Cp [x] : kf kξ ≤ kf kξ0

for arbitrary points ξ, ξ 0 ∈ A1,an 0

Cp . Furthermore, if ξ ≤ ξ , then we write [ξ, ξ ] for
0

the set of points ξ 00 such that ξ ≤ ξ 00 ≤ ξ 0 . If ξ and ξ 0 are points of type 1, 2 or 3,

then it is clear by the above that [ξ, ξ 0 ] is homeomorphic to a closed interval in R.
Regarding type 4 points, we have the following.
 § 5. The Berkovich affine and projective line 25

5.1. Lemma. Let ξ ∈ A1,an Cp be a type 4 point, represented by the nested sequence LEMMA
of disks D(an , rn ) with empty intersection. If ξ 0 ∈ A1,an 0 comparison
Cp with ξ ≤ ξ , then either
ξ 0 = ξ, or else ξ 0 is a point of type 2 or 3, corresponding to a disk D(a, r) such with
that D(an , rn ) ⊂ D(a, r) for all sufficiently large n. type 4 points

Let r∞ = limn→∞ rn . If ξ 0 corresponds to D(a, r), then

(
ζan ,ρ if ρ ≥ rn ,
γξ,ξ0 : [r∞ , r] −→ A1,an
Cp , ρ 7−→
ξ if ρ = r∞
is a continuous map with image [ξ, ξ 0 ].

Proof. Let ξ 0 ≥ ξ and assume first that ξ 0 is of type 2 or 3, so ξ 0 = ζa,r for some
a ∈ Cp and r > 0. We claim that D(a, r) ∩ D(an , rn ) 6= ∅ for all n. Otherwise, we
would have empty intersection for all sufficiently large n, and then
kx − akξ0 = r < lim |a − an | = lim kx − akan ,rn = kx − akξ ,
n→∞ n→∞
0
which contradicts ξ ≤ ξ . If n is large enough so that rn ≤ r, then we must
therefore have D(an , rn ) ⊂ D(a, r).
Now assume that ξ 0 is of type 1 or 4. If ξ 0 is of type 1, then kx − ηkξ0 is zero when
η = ξ 0 , whereas kx−ηkξ ≥ r∞ > 0 for all η, so this is not possible. If ξ 0 is of type 4,
say represented by the nested sequence of disks D(a0n , rn0 ), then ξ ≤ ξ 0 ≤ ζa0n ,rn0 for
all n. By the argument above, it follows that D(an , rn ) ∩ D(a0m , rm 0
) 6= ∅ for all
m, n. But this exactly means that ξ = ξ 0 .
It remains to show that γξ,ξ0 is well-defined and continuous. For the first, note
that ζan ,ρ = ζan+1 ,ρ when ρ ≥ rn , since D(an , ρ) = D(an+1 , ρ). For the second,
consider f ∈ Cp [x]. We have to show that
[r∞ , r] 3 ρ 7−→ kf kγξ,ξ0 (ρ)
is continuous. This is clear on (r∞ , r], and it follows for the left endpoint by the
definition of k·kξ , which implies that
kf kξ = lim kf kan(ρ) ,ρ
ρ&r∞

(with n(ρ) such that ρ ≥ rn ). q

For ξ 0 ≤ ξ, we define the path γξ,ξ0 as the reversal of the path γξ0 ,ξ .

5.2. Theorem. Any two points ξ, ξ 0 ∈ A1,an 0

Cp have a least upper bound ξ ∨ ξ (with THM
1,an
respect to the ordering introduced above). The path γξ,ξ∨ξ0 + γξ∨ξ0 ,ξ0 connects the ACp is path-
two points. connected

Proof. This is clear if neither ξ nor ξ 0 are of type 4: if ξ = ζa,r and ξ 0 = ζa0 ,r0 ,
then ξ ∨ ξ 0 = ζa,ρ = ζa0 ,ρ , where ρ = max{r, r0 , |a − a0 |}. If ξ ≤ ξ 0 or ξ 0 ≤ ξ, then
the statement is also clear (with ξ ∨ ξ 0 = ξ or ξ 0 ). Otherwise, we can represent
ξ and ξ 0 by nested sequences of disks D(an , rn ) and D(a0n , rn0 ), respectively, such
that D(an , rn ) ∩ D(a0n , rn0 ) = ∅ for n large enough. For all such n, ζan ,rn ∨ ζa0n ,rn0 is
the same and therefore agrees with ξ ∨ ξ 0 .
The statement on path-connectedness is then clear. q

We define analogues of closed disks in A1,an

Cp .
 § 5. The Berkovich affine and projective line 26

5.3. Definition. Let a ∈ Cp and r ≥ 0. We set DEF

closed disk
D(a, r) = {ξ ∈ A1,an : ξ ≤ ζa,r } = {ξ ∈ A1,an : kx − akξ ≤ r} .
Cp Cp in A1,an
Cp
and call this a closed disk in A1,an
Cp . We also define the corresponding open disk in
1,an
ACp to be
D(a, r)− = {ξ ∈ A1,an
Cp : kx − akξ < r} = Ux−a,−∞,r . ♦

The equality in the definition of D(a, r) can be seen as follows. ‘⊂’ is clear. To
show ‘⊃’, assume that kx − akξ ≤ r. Then for any b ∈ Cp , we have
kx−bkξ = k(x−a)+(a−b)kξ ≤ max{kx−akξ , |a−b|} ≤ max{r, |a−b|} = kx−bka,r .
Since every f ∈ Cp [x] is a constant times a product of such terms, it follows that
ξ ≤ ζa,r .

5.4. Lemma. LEMMA

sub-basis of
(1) Let f = (x − α1 ) · · · (x − αn ) ∈ Cp [x] be non-constant and let a ∈ R. Then
topology
there are r1 , . . . , rn ≥ 0 such that
{ξ ∈ A1,an
Cp : kf kξ ≤ a} = D(α1 , r1 ) ∪ . . . ∪ D(αn , rn )

and
{ξ ∈ A1,an −
Cp : kf kξ < a} = D(α1 , r1 ) ∪ . . . ∪ D(αn , rn )
−

(2) The open disks D(a, r)− and the complements of closed disks D(a, r) generate
the topology of A1,an
Cp (i.e., every open set is a union of finite intersections of
such sets).
(3) Two (open or closed) disks in A1,an
Cp are either disjoint or one is contained in
the other.

Proof.
(1) Exercise.
(2) We have Uf,a,b = Uf,−∞,b ∩ Uf,a,∞ . By part (1), we can write Uf,−∞,b as a union
of open disks, and we can write Uf,a,∞ , which is the complement of the set of ξ
such that kf kξ ≤ a, as the complement of a finite union of closed disks, which
is the same as a finite intersection of complements of closed disks. So each
basic open set Uf,a,b is a finite intersection of open disks and complements of
closed disks; this implies the claim.
(3) Assume that D(a, r) and D(a0 , r0 ) are not disjoint, so that there is ξ with
kx − akξ ≤ r and kx − a0 kξ ≤ r0 , w.l.o.g. such that r ≥ r0 . Then
|a − a0 | ≤ max{kx − akξ , kx − a0 kξ } ≤ r .
Now let ξ 0 ∈ D(a0 , r0 ) be arbitrary. Then
kx − akξ0 ≤ max{kx − a0 kξ0 , |a − a0 |} ≤ max{r0 , r} = r ,
so ξ 0 ∈ D(a, r). Hence D(a0 , r0 ) ⊂ D(a, r). The case when one or both disks
are open is similar. q

Together with (2) (and the fact that A1,an

Cp is a union of open disks), statement (3)
says that the open disks, from which a finite number of (pairwise disjoint) closed
disks is removed, form a basis of the topology: every open set is a union of such
sets.
 § 5. The Berkovich affine and projective line 27

5.5. Lemma. Let a ∈ Cp and r ≥ 0. LEMMA

D(a, r) is
(1) The set of type 1 points in D(a, r) is D(a, r), and D(a, r) is dense in D(a, r). compact
(2) D(a, r) is compact.

Proof. The first claim in (1) is clear (and holds in a similar way for open disks). For
the second claim, consider the intersection of a basic open set as above with D(a, r).
The set of type 1 points contained in this intersection is the intersection of D(a, r)
with the open disk in Cp corresponding to the open disk, minus the union of the
closed disks in Cp corresponding to the closed disks that were removed. Any such
set is non-empty (unless the whole disk D(a, r) is removed, but then the basic
open set has empty intersection with D(a, r)).
For (2), we can assume without loss of generality that a = 0. We know that
ξ ∈ D(0, r) ⇐⇒ k·kξ ≤ |·|r , so the image of D(0, r) in RCp [x] under the map
Φ : A1,an
Cp −→ R
Cp [x]
, ξ 7−→ (kf kξ )f ∈Cp [x]
Q
is the intersection of im(Φ) with the product C = f [0, |f |r ]. The definition of the
topology on A1,an
Cp is equivalent to saying that it is the subspace topology induced
by the map Φ above (with the product topology on RCp [x] ). The product C of
compact intervals is itself compact by Tychonoff’s Theorem. On the other hand,
the conditions defining a multiplicative seminorm are closed conditions, so the
image of Φ is closed. Now Φ(D(0, r)) is the intersection of a closed set and a
compact set, so it is itself compact. Since Φ is a homeomorphism onto its image,
it follows that D(0, r) is compact as well. q

5.6. Corollary. The space A1,an

Cp is locally compact. COR
A1,an
Cp is locally
compact
Proof. We must show that for every point ξ ∈ AC1,an p
and every open subset U
1,an
of ACp containing ξ, there is a compact neighborhood V of ξ contained in U .
Since U is open, U contains an open set of the form D(a, r)− \ j D(aj , rj ) that
S

in turn contains ξ. This means that kx − akξ < r and kx − aj kξ > rj . For
r0 < r sufficiently close to r and for rj0 > rj sufficiently close to rj , we still have
kx − akξ < r0 and kx − aj kξ > rj0 . Then
[
ξ ∈ V := D(a, r0 ) \ D(aj , rj0 )− ⊂ U .
j
0
S set D(a,0 r ) with a closed set) and
Now V is compact (we intersect the compact
0 −
contains the open neighborhood D(a, r ) \ j D(aj , rj ) of ξ. q

We now look a bit closer at the tree structure of AC1,an

p
.

5.7. Definition. The diameter of a point ξ ∈ A1,an

Cp is DEF
diameter
diam(ξ) := inf{kx − akξ : a ∈ Cp } . ♦

Then diam(ζa,r ) = r, and for a type 4 point ξ represented by a nested sequence of

disks D(an , rn ), we have diam(ξ) = limn→∞ rn > 0.
We use this notion to define two metrics.
 § 5. The Berkovich affine and projective line 28

5.8. Definition. For ξ, ξ 0 ∈ A1,an

Cp , we define the small metric by DEF
small and
d(ξ, ξ 0 ) = 2 diam(ξ ∨ ξ 0 ) − diam(ξ) − diam(ξ 0 )
big metric
= (diam(ξ ∨ ξ 0 ) − diam(ξ)) + (diam(ξ ∨ ξ 0 ) − diam(ξ 0 ))
and in the case that ξ, ξ 0 are both not of type 1, the big metric by
ρ(ξ, ξ 0 ) = c(2 log diam(ξ ∨ ξ 0 ) − log diam(ξ) − log diam(ξ 0 )) . ♦

So the small metric gives the total change of diameter as we move along the path
joining ξ to ξ 0 , whereas the big metric does the same for the change in (additive)
valuation (= −c log diam). Both are indeed metrics, d on all of AC1,an
p
and ρ on the
1,an
‘hyperbolic part’ of ACp , which consists of the points of types 2, 3 and 4. The
former has the advantage that it is also defined on the points of type 1, but the
latter is in some sense more natural. It can be seen as analogous to the hyperbolic
metric on the upper half plane in C (which is given by ds2 = (dx2 + dy 2 )/y 2 when
x and y are the real and imaginary parts), which is invariant under the action
of PSL(2, R) by Möbius transformations. This analogous property will turn out
to be satisfied (with respect to PGL(2, Cp )) by the big metric.
Now we introduce the kind of tree structure that shows up here. (For more
information, consult [BR, Appendix B].)

5.9. Definition. An R-tree is a metric space (T, d) such that for any two points DEF
x, y ∈ T , there is a unique path [x, y] in T joining x to y, which is a geodesic R-tree
segment (this means that the map γ : [a, b] → T giving the path can be chosen so
that d(γ(u), γ(v)) = |u − v| for all u, v ∈ [a, b]).
A point x ∈ T is a branch point if T \ {x} has at least three connected components
(in the metric topology). x is an endpoint if T \ {x} is connected. If T \ {x} has
exactly two connected components, then x is said to be ordinary. The R-tree T
is said to be finite, if it is compact and has only finitely many branch points and
endpoints.
The strong topology on the R-tree is the metric topology on T . To define the
weak topology, we define a tangent direction at x ∈ T to be an equivalence class
of paths [x, y] with y 6= x, where two paths are equivalent when they share an
initial segment. The tangent directions at x are in one-to-one correspondence
with the connected components of T \ {x}. For a tangent direction v at x, we set
Bx,v = {y ∈ T : y 6= x, [x, y] ∈ v} (which is the connected component of T \ {x}
corresponding to v). Then the weak topology on T is the topology generated by
the sets Bx,v . ♦

The set Bx,v can be interpreted as the set of points of T that can be ‘seen’ from x
when looking into direction v. This is why the weak topology is also called the
‘observer’s topology’.
Now the following is a fact.

5.10. Theorem. THM

A1,an
Cp as
(1) The small metric d turns A1,an 1,an
Cp into an R-tree. The Berkovich topology on ACp R-tree
is the weak topology on this R-tree.
(2) The big metric ρ turns A1,an
Cp \ Cp into an R-tree. The subspace topology
1,an
on ACp \ Cp is again the weak topology on this R-tree.
 § 5. The Berkovich affine and projective line 29

Proof. See [BR, Section 1.4], where the analogous statements are shown for D(0, 1).
For A1,an
Cp the argument is the same (except that there is no natural way to pick a
root). q

Compactifying the R-tree by adding a point at infinity (see below) we obtain

the universal dendrite that was first constructed by Ważewski in his thesis in
1923. Recent work by Hrushovski, Loeser and Poonen1 shows that V an can be
embedded in R2d+1 for any quasi-projective d-dimensional Cp -variety V (and P1,an
Cp
can be embedded in R2 ).
We can classify the points in the R-tree structure.
(1) Points of types 1 and 4 are endpoints.
(2) Points of type 2 are branch points, and the set of tangent directions corre-
sponds (canonically up to an automorphism of P1F̄p ) to P1 (F̄p ).
(3) Points of type 3 are ordinary.

Let ξ ∈ A1,an 0
Cp be any point. If ξ is another point, then there are three possibilities:
ξ > ξ 0 , ξ < ξ 0 , or neither. In the last two cases, the path from ξ to ξ 0 starts ‘going
up’ to ξ ∨ ξ 0 (= ξ 0 in the second case). So the points ξ 0 6= ξ such that ξ 6> ξ 0
constitute the set Bξ,up . If ξ is of type 1 or 4, then there are no ξ 0 with ξ > ξ 0 , so
‘up’ is the only direction.
Now consider ξ = ζa,r of type 3. If ξ 0 < ξ, then there is some ξ 0 ≤ ζa0 ,r0 < ξ, and
we have D(a0 , r0 ) ⊂ D(a, r), so |a0 − a|p ≤ r. Since r ∈ / |C×
p |p , we can choose r
0

such that |a0 − a|p < r0 < r, but then ζa0 ,r0 = ζa,r0 , so the path from ξ to ξ 0 shares
an initial segment with [ζa,r , ζa,0 ]. This shows that other than ‘up’, there is exactly
one tangent direction ‘down’.
Finally, when ξ = ζa,r is of type 2, then by a similar argument, for any ξ 0 < ξ
there is a0 ∈ D(a, r) such that ξ 0 ∈ D(a0 , r)− . If two points ξ 0 , ξ 00 are in the same
open disk, then the paths from ξ to these two points share an initial segment;
otherwise ξ 0 ∨ ξ 00 = ξ and they do not. So the directions other than ‘up’ at ξ
are in one-to-one correspondence with open disks of radius r contained in D(a, r).
To show that the set of such disks corresponds to F̄p , we assume that a = 0 and
r = 1 (we can shift and scale to reduce to this case). Two disks D(a0 , 1)− and
D(a00 , 1)− contained in D(0, 1) are equal if and only if |a0 − a00 |p < 1, which means
that a0 , a00 ∈ R (the valuation ring of Cp ) have the same image in the residue class
field F̄p . So we see that the ‘downward’ directions at a type 2 point correspond to
the elements of F̄p ; together with the ‘up’ direction, which we can let correspond
to ∞, we get P1 (F̄p ). (For ζ0,1 the correspondence is canonical. For other points
it is so only up to an automorphism of P1F̄p , since there is a choice involved in the
shifting and scaling.)

If ξ = ζa,r is a point of type 2 or 3, then the basic open set Bξ,v of the weak topology
is the complement of D(a, r) when v = ‘up’ and is D(a0 , r)− with a0 as above when
v is a ‘downward’ direction. For ξ of type 1 or 4, Bξ,v is just AC1,an p
\ {ξ}. This
shows that the weak topology of the R-tree agrees with the Berkovich topology.
Here is a rough sketch of D(0, 1):
1E. Hrushovski, F. Loeser and B. Poonen, Berkovich spaces embed in Euclidean spaces,
L’Enseignement Math. 60 (2014), no. 3–4, 273–292.
 § 5. The Berkovich affine and projective line 30

ζ0,1

ζ0,r

ζ0,|p|1/2

ζ1,|p| ζ−1,|p| ζ0,|p|

ζ0,|p|2

ξ
ζ1,0 ζ1+p,0 ζ−1,0 ζ0,0 ζp2,0 ζ−p,0 ζp,0

Points of type 1, 2, 3 and 4 are green, red, purple and blue, respectively.
Let T be an R-tree and let S ⊂ T be nonempty and finite. Then one can consider
the convex hull Γ of S; this is the union of the paths joining the points in S.
Then Γ is a finite R-tree and there is a natural deformation retraction τΓ : T → Γ
that sends any point x ∈ T to the point of Γ that is hit first by the unique path
from x to any fixed point in Γ. It can be shown that the weak topology on T is
the weakest topology that makes all the maps τΓ continuous (where we take, say,
the metric topology on Γ). (This is an exercise.)

Now we want to introduce the Berkovich projective line, P1,an

Cp . There are several
ways of constructing it.
(1) We set P1,an 1,an
Cp = ACp ∪ {∞}, where ∞ is a point of type 1, whose open neigh-
borhoods are the complements of compact subsets of AC1,an
p
together with the
point ∞. (This is the one-point compactification from general topology.)
(2) We observe that the map Cp → Cp , z 7→ z −1 induces a continuous involution φ
on A1,an
Cp \ {0} (that is given on points not of type 4 by ζa,r 7→ ζa−1 ,r/|a|2 when
|a| > r and ζ0,r 7→ ζ0,1/r for r > 0). One can then ‘glue’ two copies of A1,an Cp
1,an
along ACp \ {0} via this map. This identifies a neighborhood of ∞ with a
neighborhood of 0 (and so shows in particular that ∞ is in no way special).
One can also show that the big metric is invariant under φ and also under
affine maps z 7→ az + b (Exercise).
(3) One obtains the same result by gluing two copies of D(0, 1) along the ‘spheres’
D(0, 1) \ D(0, 1)− via φ. Since D(0, 1) is compact, this shows that PC1,anp
is
compact.
(4) Similar to the Proj construction in algebraic geometry, there is a ‘Berkovich
Proj’ taking a graded finitely generated algebra A over a complete non-archimedean
field as input and having a Berkovich type space as output, whose points corre-
spond to equivalence classes of multiplicative seminorms on A whose kernel doe
 § 5. The Berkovich affine and projective line 31

not contain the irrelevant ideal. Applying this to the polynomial ring Cp [x, y]
with the standard grading, this results in P1,an
Cp . We will discuss this in more
detail below. An advantage of this approach is that it makes it easy to see
that a morphism P1Cp → P1Cp induces a continuous map PC1,anp
→ P1,an
Cp .

Recall
Lthat a graded ring is a ring R coming with a direct sum decomposition
R = n≥0 Rn as an additive group such that for all f ∈ Rm and g ∈ Rn we have
f g ∈ Rm+n . The elements of Rn are said to be homogeneous of degree n. An
ideal I of R is said to be homogeneous if it is generated by homogenous elements.
If K is a field and the decomposition is a direct sum of K-vector spaces, then R is
a graded K-algebra. In algebraic geometry, Proj R is the L set of all homogeneous
prime ideals of R that do not contain the irrelevant ideal n≥1 Rn , with the Zariski
topology. For example, PnK = Proj K[X0 , X1 , . . . , Xn ] with the usual grading of
the polynomial ring. The ‘Berkovich Proj’ construction does something similar
for Berkovich spaces.

5.11. Definition. Let A be a finitely generated graded K-algebra with A0 = DEF

K, where K is a complete and algebraically closed non-archimedean field. We Berkovich
consider the set of multiplicative seminorms on A whose kernel does not contain Proj
the irrelevant ideal. We declare that two seminorms k·k and k·k0 are equivalent if
there is C > 0 such that kF k0 = C d kF k for all F ∈ Ad (i.e., F ∈ A homogeneous
of degree d). Then the projective Berkovich space associated to A, PBerk A, is the
set of equivalence classes of these seminorms.
Let a1 , . . . , am be homogeneous generators of the irrelevant ideal of A. Then in
each equivalence class, there are normalized seminorms k·k with the property that
maxj kaj k = 1, and all equivalent normalized seminorms agree on all homoge-
neous elements.
 We define the topology on PBerk A to be the weakest one such
that k·k 7→ kF k is continuous for every   homogeneous F ∈ A, where k·k is a
normalized representative of the class k·k .
If V = Proj A is the projective K-variety defined by A, then we also write V an for
PBerk A. ♦

The notation PBerk A is not standard.

In analogy with the usual construction in algebraic geometry, we can then define
the Berkovich projective line to be DEF
P1,an
P1,an
Cp = PBerk C p [X, Y ] , Cp

where the polynomial ring Cp [X, Y ] has its usual grading.

We find the usual two embeddings of A1,an 1,an
Cp into PCp by pulling back seminorms
under the two maps Cp [X, Y ] → Cp [x], F (X, Y ) 7→ F (x, 1) and F (X, Y ) 7→
F (1, x). The equivalence class we do not obtain under the first map satisfies
kY k = 0; it corresponds to the type 1 point 0 under the second map and represents
the point at infinity.
It is then natural to consider sets of the form P1,an −
Cp \ D(a, r) as closed and sets of
the form P1,an 1,an
Cp \ D(a, r) as open Berkovich disks (‘around infinity’) in PCp . Then
the open Berkovich disks generate the topology of PC1,an
p
, and a basis of the topology
is given by the open disks (including the whole space) minus finitely many closed
disks.
 § 5. The Berkovich affine and projective line 32

Since P1,an 1,an

Cp without the points of type 1 is the same as ACp without the points
of type 1, which is an R-tree with respect to the big metric, it follows by the
results mentioned in (2) above that Aut(P1Cp ) = PGL(2, Cp ) acts on this R-tree by
isometries. This action is transitive on the set of type 2 points, and the stabilizer
of the point ζ0,1 is PGL(2, R) (where R is the valuation ring of Cp ), with the
tangent directions at ζ0,1 being permuted transitively (by PGL(2, F̄p ) = Aut(P1F̄p )
via the canonical map PGL(2, R) → PGL(2, F̄p )).
A type 2 point ζ of P1,anCp can then be interpreted as corresponding to a reduction
1 1
map P (Cp ) → P (F̄p ), up to an automorphism of the target. The map is given
by associating to z ∈ P1 (Cp ) the tangent direction at ζ in which the type 1 point
corresponding to z can be seen. For ζ = ζ0,1 we get the usual reduction map. Such
a reduction map comes from a ‘model’ of P1 over R of the form P1R ; the reduction
map depends on how we identify the generic fiber of P1R with P1Cp . More generally,
we call any finite R-subtree of P1,an 1
Cp \ P (Cp ) with the big metric that is the convex
hull of finitely many points of type 2 a skeleton of P1,anCp . Any skeleton corresponds DEF
to a model of PCp over R that is usually more complicated than P1R in that its Skeleton
1

special fiber (i.e., the base change to F̄p ) is a configuration of several P1 ’s arranged
in tree form. (Possibly we return to that later.)
If we modify the definition of the diameter by defining it on PC1,an
p
as
(
diam(ξ) if ξ ∈ D(0, 1)
diam0 (ξ) =
diam(φ(ξ)) if ξ ∈ PC1,an
p
\ D(0, 1),
where φ is the involution induced by z 7→ 1/z, and set ξ ∨0 η to be the point
where [ξ, ζ0,1 ] and [η, ζ0,1 ] first meet, then we can define a small metric d0 on PC1,an
p

in the same way as before. Then P1,an Cp can be identified with the R-tree given by
the small metric. However, the small metric is not invariant under automorphisms
in general (the point ζ0,1 plays a special role in its definition; any automorphism
that moves ζ0,1 will change the metric).
To conclude this section, we give another interpretation of the seminorm associated
to a point in P1,anCp that (contrary to the interpretation as the supremum norm on
the associated disk when applied to polynomials) also works for rational functions.
Consider first a point ξ = ζa,r of type 2. If f is a nonzero polynomial with roots
α1 , . . . , αm in D(a, r), then for all α in the closed set D(a, r) \ m −
S
j=1 D(α j , r) we
have |f (α)|p = kf kξ . To see this, we can assume that a = 0 and f is monic. Write
f = (x − α1 ) · · · (x − αm )(x − αm+1 ) · · · (x − αn ) ,
where αm+1 , . . . , αn are the roots of f outside of D(0, r). Then
Yn Yn
kf kξ = kx − αj kξ = |α − αj |p = |f (α)|p ,
j=1 j=1

since for j ≤ m, we have kx−αj kξ = r = |α−αj |p (here we use that α ∈/ D(αj , r)− ),
and for j > m, we have kx − αj kξ = |αj |p = |α − αj |p . So |f |p is constant and
equal to kf kξ on D(a, r) outside a finite union of open disks contained in D(a, r).
We can apply this separately to the numerator and the denominator of a rational
function f = g/h ∈ Cp (x), which gives the same statement for such an f . So for
a type 2 point ξ = ζa,r , the induced seminorm k·kξ on Cp (x) is the absolute value
taken by any given function f ∈ Cp (x) on D(a, r) outside finitely many smaller
open disks (the set of disks to be removed depends on f ); this is also the same as
 § 5. The Berkovich affine and projective line 33

the absolute value taken on P1,an −

Cp \ D(a, r) outside a finite union of smaller open
disks (one of which is the ‘exterior’ PC1,an
p
\ D(a, r)). Note that when considering
rational functions f as quotients of two homogeneous polynomials in Cp [X, Y ] of
the same degree, then kf kξ is independent of the choice of representative of the
equivalence class of seminorms corresponding to ξ (the scaling factor C d cancels).
For a point ξ = a ∈ Cp of type 1, we have of course kf kξ = |f (a)|p unless a is a
pole of f (this also makes sense when a = ∞, with f (∞) interpreted as the value
at zero of f (1/x)).
If ξ = ζa,r is of type 3, then we have to use a ‘limit version’ of the above. Since
r∈ / |C× j
p |, no two nonzero terms of the form r |aj |p with aj ∈ Cp can be equal and
so we have kf kξ = maxj rj |aj |p when f (x + a) = a0 + a1 x + a2 x2 + . . . + an xn is
a nonzero polynomial. If s ∈ |C× j
p | is sufficiently close to r, then all terms s |aj |p
will still be pairwise distinct, hence for α ∈ Cp such that |α − a|p = s we will have
|f (α)|p = maxj sj |aj |p , which is close to kf kξ . So for any given ε > 0 there is a
closed annulus A = D(a, r + δ) \ D(a, r − δ)− such that |f (α)|p − kf kξ ≤ ε for
all α ∈ A.
The general statement is as follows (see Section 2.4 in [BR]).

5.12. Lemma. Let ξ ∈ P1,an Cp and denote the corresponding multiplicative semi- LEMMA
norm on Cp (x) as usual by k·kξ . Then for every f ∈ Cp (x) (that does not have a characteri-
pole at ξ when ξ is of type 1) we have that kf kξ is the unique ν ∈ R≥0 such that for zation
every ε > 0 there is a closed neighborhood U of ξ in P1,an
Cp such that |f (α)|p −ν ≤ ε
of k·kξ
1
for all α ∈ P (Cp )∩U . The set U can be taken to be a closed Berkovich disk minus
a finite union of open Berkovich disks.

Proof. Fix f . Since η 7→ kf kη is continuous by definition of the Berkovich topology,

the set U = {η : |kf kη − kf kξ | ≤ ε} is a closed neighborhood of ξ. It contains
an open neighborhood of ξ (for example, the set obtained by replacing ‘≤ ε’ by
‘< ε’), which contains an open neighborhood V that is an open Berkovich disk
minus finitely many closed Berkovich disks. Then U contains the closure of V ,
which is the corresponding closed Berkovich disk minus the corresponding open
Berkovich disks. We can always make U smaller, so we can assume that U has
this form. Then |f (α)|p −kf kξ ≤ ε for α ∈ P1 (Cp )∩U follows from the definition
of U . Since P1 (Cp ) ∩ U is dense in U , this property determines kf kξ to within ε.
Since ε > 0 was arbitrary, kf kξ is characterized by it. q

In a similar way as for Berkovich spaces associated to affine Cp -varieties, mor-

phisms between projective Cp -varieties induce continuous maps between the as-
sociated Berkovich spaces. For example, any rational function ϕ ∈ Cp (x) can be
considered as a morphism P1Cp → P1Cp and therefore induces a map P1,an 1,an
Cp → PCp .
If ϕ is not constant (otherwise it is quite clear what happens), then it follows
from the discussion above that ϕ(ξ) = η for type S 2 points−ξ = ζa,r and η = ζa0 ,r0
if and only if ϕ maps a suitable set D(a, r) \ j D(aj , r) to a set of the form
D(a0 , r0 ) \ j D(a0j , r0 ). This can be helpful when one is trying to figure out what
S

ϕ is doing on P1,an
Cp .
 § 6. Analytic spaces and functions 34

6. Analytic spaces and functions

In this section, we follow roughly part of Brian Conrad’s text [AWS, Chapter 1].
Recall that a K-Banach algebra (for a complete field K with absolute value) is
a K-algebra A with a (submultiplicative) norm |·|A that restricts to the absolute
value on K and such that A is complete with respect to this norm. So we have
|a|A 6= 0 for a 6= 0, |a + b|A ≤ |a|A + |b|A and |ab|A ≤ |a|A · |b|A .
If K is non-archimedean, then |·|A also satisfies the ultrametric triangle inequality.

6.1. Definition. Let A be a K-Banach algebra; we write the norm on A as |·|A . DEF
A multiplicative seminorm k·k on A is said to be bounded, if kf k ≤ |f |A for all bounded
f ∈ A. ♦ multiplicative
seminorm
We note that it is sufficient to require kf k ≤ C|f |A for some constant C > 0, since
this implies
kf k = (kf n k)1/n ≤ (C|f n |A )1/n ≤ C 1/n |f |A ;
we obtain kf k ≤ |f |A by letting n tend to infinity. So the notion of ‘bounded
multiplicative seminorm’ does not change when we replace the norm on A by
an equivalent one (i.e., one that is bounded below and above by some positive
multiple of |·|A ).

6.2. Definition. Let A be a K-Banach algebra. The Berkovich space Berk A DEF
associated to A is the set of all bounded multiplicative seminorms on A, with the Berkovich
weakest topology that makes the maps k·k 7→ kf k continuous for all f ∈ A. ♦ space
associated
In a similar way as we did earlier, one shows the following. to A

6.3. Theorem. Let A be a K-Banach algebra. Then Berk A is a compact Haus- THM
dorff space, which is empty if and only if A = {0}. Berk A is
compact and
We note that one can define Berk A for any (commutative) ring A that is complete Hausdorff
with respect to some absolute value (a Banach ring); it does not have to be an
algebra over some complete field. One could take A = Z with the usual absolute
value, for example (completeness follows trivially, since every Cauchy sequence
must be eventually constant), or any ring with the trivial absolute value. It is an
instructive exercise to work out the structure of Berk Z!
From now on, K will be, as usual, a complete and algebraically closed non-
archimedean field, for example, K = Cp . There is an important class of K-Banach
algebras.

6.4. Definition. Let n ∈ Z>0 and r1 , . . . , rn > 0. We define DEF

nX o Tate
Khr1−1 x1 , . . . , rn−1 xn i := ai xi : ri |ai | → 0 as |i| → ∞ ⊂ K[[x1 , . . . , xn ]] . algebra
i∈Zn
≥0

Here we use multi-index notation: for i = (i1 , . . . , in ) we set xi = xi11 · · · xinn and
ri = r1i1 · · · rnin ; also |i| = i1 + . . . + in . Then
X
ai xi := max ri |ai |
i
i
 § 6. Analytic spaces and functions 35

defines a multiplicative norm on A = Khr1−1 x1 , . . . , rn−1 xn i turning A into a K-

Banach algebra. We call A a Tate algebra over K when all rj ∈ |K × |, otherwise
A is a generalized Tate algebra over K.
More generally, if A is any K-Banach algebra, then we can define (generalized)
relative Tate algebras Ahr1−1 x1 , . . . , rn−1 xn i in the same way. ♦

Note that by scaling the variables suitably, any Tate algebra in n variables is
isomorphic to the standard Tate algebra
Khx1 , . . . , xn i := Kh1−1 x1 , . . . , 1−1 xn i .
This is not true for generalized Tate algebras.

6.5. Example. We have Berk Cp hxi ∼ = D(0, 1). The map is induced by the EXAMPLE
inclusion Cp [x] ,→ Cp hxi; the image consists of all seminorms on Cp [x] that are D(0, 1)
bounded by the restriction of the norm on the Tate algebra to the polynomial
ring, which is exactly |·|1 = k·k0,1 . By definition, this set is D(0, 1). It also follows
easily from the definitions that the map induces a homeomorphism.
More generally, the same argument shows that Berk Cp hr−1 xi ∼ = D(0, r), for every
r > 0. We see that there is some kind of correspondence between type 2 points and
Tate algebras, and type 3 points and generalized Tate algebras. Quite explicitly, we
obtain Cp hr−1 xi as the completion of Cp [x] with respect to the norm corresponding
to ζ0,r . ♣

We state the following facts without proof.

6.6. Lemma. The Tate algebras over K are noetherian (every ideal is finitely LEMMA
generated) unique factorization domains. They are also Jacobson rings (every Properties
prime ideal is the intersection of the maximal ideals it is contained in). Every of Tate
ideal is closed. algebras
The norm is intrinsically defined as |f | = maxφ |φ(f )|, where φ runs over all
continuous K-algebra homomorphisms to K.
The Tate algebra T = Khr1−1 x1 , . . . , rn−1 xn i has the following universal property.
Given any K-Banach algebra A and elements a1 , . . . , an ∈ A with |aj |A ≤ rj , there
is a unique continuous K-algebra homomorphism φ : T → A with φ(xj ) = aj .

Let A be a Tate algebra and I ⊂ A an ideal. Then by the above, I is closed.

Define a real-valued function on A/I by
|a + I|A/I = inf{|a + b|A : b ∈ I} .

6.7. Lemma. |·|A/I is a norm on A/I turning A/I into a Banach algebra such LEMMA
that the canonical homomorphism A → A/I is continuous. induced
norm
Proof. We have to show a number of things.
(1) |a + I|A/I = 0 only for a ∈ I.
If |a + I|A/I = 0, then there is a sequence (bn ) in I such that |a + bn |A → 0,
hence −bn → a. Since I is closed, this implies a ∈ I.
(2) |(a + a0 ) + I|A/I ≤ max{|a + I|A/I , |a0 + I|A/I }.
This follows easily from the definition.
 § 6. Analytic spaces and functions 36

(3) |aa0 + I|A/I ≤ |a + I|A/I · |a0 + I|A/I .

For b, b0 ∈ I we have |aa0 + I|A/I ≤ |(a + b)(a0 + b0 )|A ≤ |a + b|A · |a0 + b0 |A ;
taking the infimum over all b and b0 yields the claim.
(4) A/I is complete.
Consider a Cauchy sequence (an + I) in A/I, so that |an+1 − an + I|A/I → 0.
Use the definition of |·|A/I to pick recursively a sequence (bn ) in I with b0 = 0
such that |(an+1 + bn+1 ) − (an − bn )|A → 0. Then (an + bn ) is a Cauchy
sequence in A; since A is complete, this sequence has a limit a. But then
|an − a + I|A/I ≤ |(an + bn ) − a|A → 0, so (an + I) converges to a + I in A/I.
(5) A → A/I is continuous.
This follows from |(a + I) − (b + I)|A/I ≤ |a − b|A . q

6.8. Definition. A K-Banach algebra A is a K-affinoid algebra if there is a DEF

generalized Tate algebra T and a surjective K-algebra homomorphism φ : T → A affinoid
such that the norm of A is (equivalent to) the induced norm on T / ker(φ) ∼
= A. If algebra
T can be taken to be a Tate algebra, then A is a strict K-affinoid algebra. and domain
The space Berk A associated to a (strict) affinoid algebra is a (strict) affinoid
domain. ♦

Simple examples of affinoid domains are the closed Berkovich disks D(0, r), where
A is itself a (generalized) Tate algebra.

6.9. Example. Let 0 < r ≤ s. Then we can consider the algebra EXAMPLE
annuli
−1 −1 Khs−1 X, rY i
A = Khs X, rX i :=
hXY − 1i
with its induced norm. Using the relation XY = 1, its elements can be written
uniquely as the image of a series of the form
∞
X ∞
X
a−m Y m + a0 + an X n
m=1 n=1

with sn |an | → 0 as n → ∞ and rm |am | → 0 as m → −∞. We can interpret this

(via Y = X −1 in A) as the Laurent series
∞
X
an X n ;
n=−∞

the conditions on the coefficients then are equivalent to saying that this series
converges whenever we plug in some element α such that r ≤ |α| ≤ s. For K = Cp ,
we see that via the map Cp [x] → A sending x to X, Berk A gets identified with
A(r, s) := D(0, s) \ D(0, r)− , which is a closed Berkovich annulus. ♣

Most of the structure of affinoid algebras is intrinsic. We state the following

without proof.
 § 6. Analytic spaces and functions 37

6.10. Theorem. Let A be a K-affinoid algebra. THM

Properties
(1) Any two K-Banach algebra norms on A are equivalent. In particular, the of affinoid
topology on A induced by the norm is intrinsic, and the same is true for notions algebras
of boundedness.
(2) Any K-algebra morphism A0 → A from another K-affinoid algebra is contin-
uous (and hence a morphism of K-affinoid algebras).

There are more properties similar to those of algebras of finite type in algebraic
geometry. For example, if A is K-affinoid and A0 is a K-algebra that is a finitely
generated A-module, then A0 is also K-affinoid. Any K-affinoid algebra is a finitely
generated module over some Tate algebra (this corresponds to the Noether nor-
malization theorem).
One can then construct more general Berkovich K-analytic spaces by gluing affi-
noid domains. There are some technical issues here, because the gluing is along
compact ‘sub-affinoid domains’ and not along open sets as one does in algebraic
or differential geometry.
For example, for fixed s > 0 one can glue the closed disks of radius r < s via
the natural inclusions D(0, r) ⊂ D(0, r0 ) whenever r < r0 < s to obtain the
open disk D(0, s)− . In a similar way, one gets open annuli A(r, s)− (for r < s).
Considering the affine line as ‘the open disk of radius ∞’, we obtain AC1,an
p
. We get
1,an
the projective line PCp by gluing two closed disks D(0, 1) along the annulus A(1, 1)
as discussed in the last section.
We also have the following (see the exercises). As usual, R denotes the valuation
ring of K and k the residue field.

6.11. Theorem. Let A be a K-Banach algebra. Then THM

Reduction
A◦ = a ∈ A : {|an |A : n ≥ 0} is bounded

map
is a closed subalgebra of A that contains (the image of ) R and depends on the norm
of A only up to equivalence (and so is completely intrinsic when A is affinoid).
The set A◦◦ = {a ∈ A : an → 0} is an ideal of A◦ , and the quotient à = A◦ /A◦◦
is a k-algebra functorially associated to A. If A is strictly K-affinoid, then à is a
finitely generated k-algebra. If k·k ∈ Berk A, then the set {a ∈ A◦ : kak < 1} is
a prime ideal of A◦ containing A◦◦ ; in this way we obtain a canonical reduction
map ρA : Berk A → Spec Ã. If A is strictly K-affinoid, then ρA is anti-continuous:
preimages of closed sets are open and vice versa.

For example, taking A = Cp hxi, we find à = F̄p [x], and the reduction map sends
ζ0,1 to the generic point of A1F̄p = Spec à and the points in the open Berkovich
disk D(α, 1)− (for α ∈ D(0, 1)) are mapped to the closed point ᾱ ∈ A1 (F̄p ). In
particular, we see that the preimage of a closed point is an open Berkovich disk,
and the preimage of the open set A1F̄p \ S, where S is a finite set of closed points,
is the closed set D(0, 1) \ s∈S D(αs , 1)− , where αs ∈ D(0, 1) has image s.
S

6.12. Example. Let 0 < r < s be in |K × | and consider as before the algebra EXAMPLE
A = Khs−1 X, rX −1 i. The induced norm on A is given by annuli
X
an X n = max {rn |an | : n ≤ 0} ∪ {sn |an | : n ≥ 0}


A
n∈Z
 § 6. Analytic spaces and functions 38

(this is because |XY | = s/r > 1, so the ‘smallest’ representative of 1 ∈ A =

KhX, Y i/hXY − 1i is 1). By choosing c, C ∈ K such that |c| = r and |C| = s and
scaling X and Y , we can write A as the quotient
A∼= KhX, Y i/hXY − cC −1 i
of the standard two-variable Tate algebra T . From this it is fairly easy to see that
à = T̃ /hXY i = k[X, Y ]/hXY i ,
where k is the residue field. So Spec à is a union of two affine lines (corresponding
to X = 0 and Y = 0) meeting transversally in one point (X = Y = 0). Let
Γ = [ζ0,r , ζ0,s ] ⊂ A(r, s). One can check that ρA maps ζ0,r ∈ A(r, s) to the generic
point of the line X = 0, ζ0,s to the generic point of the line Y = 0, points retracting
under τΓ to ζ0,r to the point on X = 0 corresponding to their tangent direction and
similarly for points retracting to ζ0,s , and all points retracting to any other point
of Γ (in the ‘open interval’ between the endpoints) to the point of intersection of
the two lines. ♣

The anti-continuity of the reduction map ρA implies that the gluing together of
strict affinoids along closed sub-affinoids induces a corresponding gluing of the
Spec Ã’s along open sets. For example, gluing two copies of D(0, 1) along A(1, 1)
to obtain P1,an 1 1
Cp corresponds to gluing to copies of AF̄p along A \{0} (which is Spec Ã
for the annulus A(1, 1)), resulting in P1F̄p in the usual way. The reduction maps
also glue, which in the example gives the map identifying the tangent directions
at ζ0,1 ∈ P1,an 1 1
Cp with P (F̄p ) and sending ζ0,1 to the generic point of PF̄p .

A given space can usually be obtained in many different ways by gluing. This can
result in different ‘special fibers’ (the glued spectra of the algebras Ã) and therefore
different reduction maps. For example, we can obtain PC1,an p
also by gluing D(0, 1)
(via the inversion map) and D(0, 1/p) to A(1/p, 1) along the sub-annuli A(1, 1)
and A(1/p, 1/p). The special fiber of this object consists of two copies of P1F̄p
meeting transversally in one point. It is obtained from the special fiber of the
annulus — two A1 ’s meeting in one point — by gluing an A1 to each of the lines
(along the complement of the intersection point). In a similar way, we can glue
together also more general affinoids (the most general of which is a closed disk
minus a finite union of open ones) to obtain a ‘model’ of P1,an Cp . The skeleton
associated to such a model is the convex hull of the points mapping to generic
points of the special fiber; in the preceding example, this would be the interval
[ζ0,1/p , ζ0,1 ]. In general, it will be a finite sub-tree of P1,an
Cp with endpoints of type 2.

There is a converse to this, valid for general smooth projective irreducible curves C
over Cp : There is a bijective correspondence between finite subgraphs of Berk C
with vertices of type 2 and semistable models of C over the valuation ring R of Cp
(up to isomorphism), such that the vertices correspond to the generic points of
the special fiber of the semistable model under the reduction map.2

2M. Baker, S. Payne, J. Rabinoff, On the structure of non-Archimedean analytic curves.

Tropical and non-Archimedean geometry, 93–121, Contemp. Math., 605, Amer. Math. Soc.,
Providence, RI, 2013; arXiv: 1404.0279.
 § 7. Berkovich spaces of curves 39

7. Berkovich spaces of curves

In this section we will take a closer look at Berkovich spaces of (smooth projective
irreducible) curves over Cp . We begin with some lemmas. Let Tr = Cp hr−1 xi be
the univariate (generalized) Tate algebra with parameter r > 0.

7.1. Lemma. Let LEMMA

roots of
F (x, y) = Fn (x)y n + Fn−1 (x)y n−1 + . . . + F0 (x) ∈ Tr [y]
polynomials
be a polynomial such that there is δ < 1 with |F0 |r ≤ δ, |1 + F1 |r ≤ δ and over Tate
|Fj |r ≤ δ 2−j for all j ≥ 2. Then there is h(x) ∈ Tr such that |h(0)|p < 1 and algebras
F (x, h(x)) = 0.

Proof. The equation F (x, y) = 0 is equivalent to the fixed point equation

y = Ψ(y) := F0 (x) + (1 + F1 (x))y + F2 (x)y 2 + . . . + Fn (x)y n .
For |y1 |, |y2 | ≤ δ we have
|Ψ(y2 ) − Ψ(y1 )|r
= |y2 − y1 |r
· |(1 + F1 (x)) + F2 (x)(y2 + y1 ) + . . . + Fn (x)(y2n−1 + y2n−2 y1 + . . . + y1n−1 )|r
≤ max{|1 + F1 |r , |F2 |r δ, . . . , |Fn |r δ n−1 } · |y2 − y1 |r
≤ δ · |y2 − y1 |r .
Also,
|Ψ(y)|r ≤ max{|F0 |r , |1 + F1 |r δ, |F2 |r δ 2 , . . .} ≤ δ
when |y|r ≤ δ. So Ψ defines a contracting map on {y ∈ Tr : |y|r ≤ δ}. The Banach
Fixed Point Theorem then gives us a unique solution. q

7.2. Lemma. Let F (x, y) ∈ Tr [y] be monic of degree n such that F (0, y) ∈ Cp [y] LEMMA
has no multiple roots. Then there is 0 < r0 ≤ r and there are h1 , . . . , hn ∈ Tr0 such étale
that F (x, y) = (y − h1 (x)) · · · (y − hn (x)). covering
of disk
Proof. Let η1 , . . . , ηn ∈ Cp be the n distinct roots of F (0, y). By assumption,
βj := F 0 (0, ηj ) 6= 0 (where F 0 denotes the derivative of F as a polynomial in y).
Let F̃j (x, y) := −βj−1 F (x, y + ηj ) Note that F̃j has no constant term and that
the coefficient of y is −1. So, writing F̃j (x, y) = f0 (x) + f1 (x)y + . . . and setting
γ = max{|f0 |r , |1 + f1 |r , |f2 |r , . . .}, we have for every 0 < r0 ≤ r that
|f0 |r0 ≤ γr0 /r , |1 + f1 (x)|r0 ≤ γr0 /r and |fm |r0 ≤ γ for m ≥ 2.
Now pick λ ∈ C×
p with |λ|p ≤ γ
−1
and set

F̃˜j (x, y) = λ−1 F̃j (x, λy) = λ−1 f0 (x) + f1 (x)y + λf2 (x)y 2 + . . . ;
we have |λ−1 f0 |r0 ≤ |λ|−1 0 0
p γr /r, |1+f1 |r0 ≤ γr /r and |λ
m−1
fm |r0 ≤ |λ|m−1
p γ ≤ γ 2−m
for m ≥ 2. Fix 0 < δ < 1 such that δ ≤ γ −1 . With r0 = r min{1, |λ|p γ −1 δ, γ −1 δ}
the assumptions of Lemma 7.1 are then satisfied for F̃˜ , so (translating back to
j
the original coordinates) there is hj ∈ Tr0 with hj (0) = ηj and F (x, hj (x)) = 0.
We can take r0 so that it works for all j; we then obtain n distinct roots of F
in Tr0 . q
 § 7. Berkovich spaces of curves 40

I think that one can take any r0 ≤ r here such that F (ξ, y) has no multiple roots
for all ξ ∈ Cp with |ξ|p ≤ r0 , at least when p > n, but so far I have found no proof
for this more precise statement. We will see below that it is true when n = 2 and
p > 2.

7.3. Example. The example F (x, y) = y 2 − y + x over C2 shows that the EXAMPLE
condition p > n is necessary in general. The two roots of F are necessity
√ ∞   of ‘p > n’
1 − 1 − 4x 2 3 4 5
X 1 2n − 2
n
h1 = = x + x + 2x + 5x + 14x + . . . = x
2 n=1
n n−1
n
and h2 = 1 − h1 . The coefficients are integers, but the coefficient of x2 is odd
for all n, so the coefficients of the series are bounded but do not tend to zero
2-adically. This means that h1 , h2 ∈ Tr for all r < 1, but not for r = 1. On
the other hand, the discriminant of F is 1 − 4x, which does not vanish even for
|x|2 < 4. This seems to be some issue related to ‘wild ramification’. ♣

7.4. Theorem. Let C be a smooth projective irreducible curve over Cp and let THM
φ : C → P1 be a morphism of degree n. Then the induced map φ∗ : C an → P1,an coverings
has fibers of size at most n. of P1

Proof. Let ζ ∈ P1,an be some point; we can assume without loss of generality that
ζ 6= ∞. The statement is clear when ζ is a type 1 point, so we consider ζ not
of type 1; in particular, the seminorm on Cp [x] associated to ζ is a norm. The
point ζ also corresponds to a homomorphism Cp [x] → H into a complete field H
with absolute value such that the absolute value pulls back to the seminorm given
by ζ and the image of Cp [x] in H generates a dense subfield. Let K be the
function field of C; we have an inclusion Cp (x) ⊂ K, which is a field extension
of degree n. Then H ⊗Cp (x) K is an algebra over H of degree n, which splits as
a direct product of finite field extensions of H (whose degrees sum to n). There
is a unique extension of the absolute value of H to each of these fields, which by
pulling back to K (and then restricting to the affine coordinate ring A of φ−1 (A1 ))
give rise to the various multiplicative (semi)norms on A extending k·kζ . So the
fiber φ−1
∗ (ζ) has as many points as there are fields in the splitting of H ⊗Cp (x) K;
this number is at most n. q

This basically shows that we can glue together C an from n copies of P1,an . If the
statement above on the choice of r0 is true and p > n, then it follows that φ∗ can
be branched (i.e., have fibers of size < n) only above points in the convex hull of
the type 1 points corresponding to branch points of φ. This might be true even
when this statement is wrong; in any case it would mean that C an deformation
retracts to a finite graph that is glued from n copies of a tree (= the convex hull
of the branch points).
We now consider hyperelliptic curves in more detail. We assume that p > 2. First
we need two elementary lemmas.

7.5. Lemma. Let f (x) ∈ Tr with |f |r < 1. Then there is h(x) ∈ Tr with |h|r < 1 LEMMA
such that 1 + f (x) = (1 + h(x))2 . square
roots
More generally, the same result applies to any complete ring R in place of Tr as
long as |2|R = 1 and 2 ∈ R× .
 § 7. Berkovich spaces of curves 41

Proof. We want a root h of F (x, y) = − 21 y 2 − y + 21 f (x) with |h − 1|r < 1. Since

|2|p = 1 and |f |r < 1, we can directly apply Lemma 7.1. The more general
statement follows in the same way. q

7.6. Lemma. Let H be the completion of the field of fractions of Tr . Then x is LEMMA
not a square in H. x not a
square
Proof. First consider the case that r ∈ |C× p |. Then we can assume r = 1, since
Tr is isomorphic to T1 and the isomorphism only scales x. Assume that x = z 2
is a square in H. Since the field of fractions of T1 is dense in H, there must be
a, b ∈ T1 such that |z − a/b| is small, which implies that |a(x)2 − xb(x)2 |  |b(x)|2 .
We can scale a and b so that |b| = 1; then also |a| ≤ 1. Looking at the reductions,
we must have that a(x)2 − xb(x)2 reduces to zero in F̄p [x], but this is impossible,
since the first term has even degree and the second term is nonzero and has odd
degree.
To deal with arbitrary r, we can enlarge Cp to obtain a field K (that is again
complete and algebraically closed) such that r ∈ |K × |. Then the argument above
shows that x is not even a square in a larger field. q

The preceding lemma can be generalized: if f ∈ T1◦ has reduction in F̄p [x] that is
not a square, then f is not a square in H. Conversely (assuming p is odd as we
do here), one can show using Lemma 7.5 that when the reduction of a polynomial
f ∈ Cp [x] ∩ T1◦ is a nonzero square, then f is a square in H. See the exercises.
A hyperelliptic curve C of genus g over Cp can be glued together from two affine
pieces, the plane curves
y 2 = f2g+2 x2g+2 + f2g+1 x2g+1 + . . . + f1 x + f0
and
Y 2 = f0 X 2g+2 + f1 X 2g+1 + . . . + f2g+1 X + f2g+2
with thePidentifications xX = 1 and Y = X g+1 y; we assume that the polynomial
f (x) = j fj xj has no multiple roots and degree at least 2g + 1. If we assume
that f2g+2 6= 0, then C → P1x is unbranched above infinity, so we can restrict our
attention to the first affine patch. We denote the double cover C → P1 given by
the x-coordinate by π, and we denote by Θ the set of branch points in P1 (Cp ) (i.e.,
the roots of f ), which we also consider as type 1 points of P1,an Cp .

7.7. Lemma. Let ζ = ζ0,r ∈ A1,an be of type 2 or 3. We have a natural partition LEMMA
of Θ into finitely many nonempty subsets Θv (indexed by the tangent direction v branching
in which the points in Θv are visible). Then π∗−1 (ζ0,r ) consists of two points if all of π
subsets Θv have even cardinality, and of one point otherwise.

/ |C×
Proof. If r ∈ p | (then ζ is of type 3), then there are just two tangent directions,
which we denote 0 and ∞. Otherwise we select one point a ∈ P1 (Cp ) representing
each occurring tangent direction different from ‘up’ and write Θa for the corre-
sponding subset (and Θ∞ for the subset of roots ξ such that |ξ|p > r). Then we
can write
Y x  Y #Θa
Y ξ − a 
f (x) = c 1− · (x − a) 1−
ξ∈Θ
ξ a ξ∈Θ
x−a
∞ a
 § 7. Berkovich spaces of curves 42

with some c ∈ C× p . By Lemma 7.5 (note that |x/ξ|r < 1 when |ξ|p > r) each factor
1 − x/ξ in the first product is a square in Tr . Now let H be the completion of the
field of fractions of Tr . Then for ξ ∈ Θa , we have |ξ − a|p < r and |x − a|r = r,
so (ξ − a)/(x − a) ∈ H has absolute value < 1, so 1 − (ξ − a)/(x − a) is also a
square in H using Lemma 7.5 again. Since Cp is Qalgebraically closed, c is of course
also a square. So f (x) is a square in H times a (x − a)#Θa . By Lemma 7.6 and
the remark following it, it follows that the latter and hence f (x) is a square in H
if and only if all exponents #Θa are even (if ζ is of type 3, then we only need
that x is not a square; if ζ is of type 2, we can assume that r = 1). On the other
hand, the degree 2 algebra over H induced by π is H[y]/hy 2 − f (x)i. This is a
field if and only if f (x) is not a square in H; otherwise it splits as a product of
two copies of H (by the Chinese Remainder Theorem). Since the number of fields
in the product decomposition of this algebra is the number of points in the fiber
of π∗ above ζ, this completes the proof. q

7.8. Corollary. Let p > 2 and let C : y 2 = f (x) be a hyperelliptic curve COR
over Cp with hyperelliptic double cover π : C → P1 . Then the induced double hyperelliptic
cover π∗ : C an → P1,an is branched exactly above points ζ such that there is at least Berkovich
one tangent direction at ζ that points to an odd number of roots of f . In particular, curve
branching only occurs along the convex hull T of the branch points of π (i.e., the
roots of f ) considered as points of type 1, and C an can be obtained by gluing two
copies of P1,an along parts of T .

Proof. Let ζ ∈ P1,an . If ζ ∈ / T , then there is some Berkovich disk D(a, r) con-
taining ζ such that D(a, r) does not contain branch points of π. The proof of
Lemma 7.7 in the case that Θ = Θ∞ (or also Lemma 7.5 directly) then shows that
f (x + a) is a square in Cp hr−1 xi, which implies that π∗ is unramified above D(a, r)
and in particular above ζ. So branching can only occur along T ; in particular, no
branching occurs at points of type 4 (which is consistent with the statement, since
there is only one tangent directions in which all 2g + 2 branch points can be seen).
For points ζa,r ∈ T not of type 1, the statement on branching is Lemma 7.7 (after
shifting x by a). For the points ζ of type 1, the statement is clear — note that
when ζ is a branch point of π, then one can see the other 2g + 1 branch points in
the unique tangent direction; otherwise one sees all 2g + 2 branch points in this
direction. q

Since P1,an deformation retracts to T , we see that C an will deformation retract to

what is obtained by gluing two copies of the tree T along the set of points that
partition the set of branch points (which are the leaves of T ) into subsets at least
one of which has an odd number of elements. For simplicity, let us say that a
vertex or edge of T is even, if the partition of the leaves it induces results in sets
with even cardinality, and odd otherwise. We will consider leaves as odd vertices.
Then the edge connecting a leaf to its adjacent vertex is always odd, so (as long
as there is at least one non-leaf vertex of T ) at the cost of another deformation
retraction, we can remove these edges from the graph obtained by gluing the trees.
We can continue this process as long as there are terminal odd edges (and at least
one vertex remains). In particular, we see that C an can be contracted to a point
(and so is simply connected) if and only if all edges are odd (equivalently, the
branch locus is all of T ). Otherwise we can find two odd vertices connected by a
chain of even edges and vertices, and the gluing produces a circle, so the resulting
space is not simply connected.
 § 7. Berkovich spaces of curves 43

7.9. Examples. The simplest case is when g = 0 (strictly speaking, a double EXAMPLES
cover of P1 is a hyperelliptic curve only when g ≥ 2, but we can consider them double
also for g = 0 or g = 1). Up to a change of coordinates, we can take f (x) = x; covers of P1,an
then the branch locus in P1 (Cp ) consists of the two points 0 and ∞. Their convex
hull T is simply the arc connecting the two (it consists of all points ζ0,r for r ≥ 0
together with ∞). So as a tree, T has two leaves connected by an odd edge. By
the remarks above, C an is simply connected, which is no surprise, since C ∼ = P1 .
1 1,an
In the sketches below, P (Cp ) is symbolized by the circle and P by the disk it
encloses; odd vertices and edges of T are red and even ones (which do not exist
here) are green.

The next case is when g = 1; it is a bit more interesting. We have four branch
points; by a change of coordinates, we can move three of them to 0, 1, ∞. Then
f (x) = x(x − 1)(x − λ) for some λ ∈ Cp \ {0, 1}. There are four automorphisms
of P1 that fix a set of four points, but there are 24 ways of selecting the three
points that are mapped to 0, 1, ∞, so each curve C arises from (in general) six
different choices of λ. If λ is one of them, then the others are

1 1 1 λ
1 − λ, , 1− , and .
λ λ 1−λ λ−1
In particular, we can assume that |λ|p ≤ 1. There are now two cases.
If |λ|p = |1 − λ|p = 1, then the four branch points lie in four different tangent
directions at ζ0,1 , since the reduction λ̄ ∈ P1 (F̄p ) is distinct from (the reductions of)
0, 1, ∞. So the tree T has one inner vertex of degree 4 and four edges connecting
it to the four leaves. In particular, all vertices and edges are odd, and C an is
simply connected. In this case C is an elliptic curve with good reduction.
In the second case, we can assume (perhaps after replacing λ by 1 − λ) that
|λ|p < 1. Then at ζ0,1 , we see one branch point each when looking to ∞ or 1,
but we see two (namely 0 and λ) when looking to 0. We get a similar partition
at ζ0,|λ|p , which is now {0}, {λ}, {1, ∞}. So T has the two (odd) inner vertices ζ0,1
and ζ0,|λ|p , each of degree 3, there is one (even) edge connecting them, and each of
the two vertices has two more (odd) edges connecting them to two of the branch
points (1 and ∞, resp., 0 and λ). Gluing two copies of T along the odd edges and
vertices then gives a graph that looks like T with the central edge doubled. This
double edge forms a circle to which everything else can be retracted. Its length in
the big metric is twice the length of the edge, so it is 2vp (λ), which is the same as
 § 7. Berkovich spaces of curves 44

−vp (j(C)), where j(C) is the j-invariant of the elliptic curve C, given by
(λ2 − λ + 1)3
j(C) = 28 .
λ2 (1 − λ)2
In any case, C an is not simply connected; its fundamental group is that of the circle,
so ∼= Z. Taking the universal covering space ‘unwraps’ the circle and replaces
it by a line; the universal covering C̃ an of C an is Gan m = P
1,an
\ {0, ∞}. (Gm
denotes the multiplicative group; its affine coordinate ring is Cp [x, x−1 ] and its
Cp -points are Gm (Cp ) = C× p .) Tate has shown (using different language) that the
covering map Gan m → C an
is actually a group homomorphism and that the group of
deck transformations is generated by ‘multiplication by q’, where q ∈ C× p satisfies
|q|p < 1. Pulling back the coordinate functions to Gm gives an explicit analytic
uniformization C(Cp ) ∼ = C× p /q
Z
that is similar in spirit to the uniformization
over the complex numbers: C(C) ∼ = C/(Z + Zτ ) ∼ = C× /q Z , where the second
2πiz 2πiτ
isomorphism is induced by z 7→ e and q = e , with τ ∈ C in the upper
half-plane (so |q| < 1).

∞ ∞

1 1
ζ 0,1 ζ 0,1

ζ 0,|λ|
λ
0 λ 0

Left: |λ|p = |1 − λ|p = 1, right: |λ|p < 1. ♣

 § 8. Integration 45

8. Integration

In this final section we want to consider integration on Berkovich spaces. The

goal is to have a ‘reasonable’ way of associating to a closed 1-form ω (meaning
that dω = 0) on an analytic space X and a path γ : [0, 1] → X whose endpoints
are ‘type 1 points’ (i.e., they correspond to seminorms whose kernel is a maximal
ideal, so functions can be evaluated at such points to give a value in Cp ) an integral
Z
ω ∈ Cp .
γ

By ‘reasonable’, we mean that this integral has the properties we expect from such
a notion, for example:
Z Z Z
(1) Linearity in the 1-form. (α1 ω1 + α2 ω2 ) = α1 ω1 + α2 ω2
γ γ γ
for α1 , α2 ∈ Cp and ω1 , ω2 closed 1-forms on X.
Z Z Z
(2) Additivity on paths. ω= ω+ ω
γ1 +γ2 γ1 γ2
when γ1 (1) = γ2 (0) and γ1 + γ2 is the path obtained by traversing first γ1 and
then γ2 .
Z
(3) Fundamental theorem of calculus. df = f (γ(1)) − f (γ(0))
γ
when f is a function on X.
Z Z
∗
(4) Change of variables. φω= ω
γ φ◦γ
when φ : Y → X is a morphism, γ is a path in Y and ω is a closed 1-form
on X.
Z Z
(5) Homotopy invariance. ω= ω
γ1 γ2
when γ1 and γ2 are homotopic with fixed endpoints. (This includes invariance
under orientation-preserving re-parameterization of the path.)
We note that Property (3) uniquely determines the integral when X = D(0, 1)−
is an open
P disk,n since then any 1-form on X has the form ω = f (t) dt where
f (t) = ∞ a
n=0 n t converges on D(0, 1) −
, and
∞
X an−1 n
f (t) dt = d t ,
n=1
n
where the series on the right converges, too (since for 0 ≤ r < 1 we have that
rn /|n|p ≤ nrn → 0). So every 1-form is exact and Property (3) applies. The same
is true for open Berkovich ‘poly-disks’ (which are the analytic spaces associated
to a product of open disks in Cp ), when we deal with higher-dimensional spaces.
As a more interesting example, let us consider the logarithm, which should be
defined on X = Gan
m = P
1,an
\ {0, ∞} by the integral
Zx
dt
log x = .
t
1
 § 8. Integration 46

Recall that Gan

m is simply connected, so by Property (5) the integral depends only
on the endpoints and not on the path (which is basically unique here anyway). If
|x − 1|p < 1, then we can apply Property (3) using Property (4) for the inclusion
D(1, 1)− ⊂ Ganm ; this leads to the usual expression
∞
z2 z3 X zn
log(1 + z) = z − + ∓ ... = (−1)n−1 for |z|p < 1.
2 3 n=1
n

We can do the same within any open disk not containing zero; using Property (2)
this gives us
for a ∈ C×


log a(1 + z) = log a + log(1 + z) p , |z|p < 1.

The question then is, how to fix the values of log a when |a − 1|p ≥ 1?
If we use another important property of the logarithm, namely its functional equa-
tion
log(xy) = log x + log y ,
then this determines log on the subgroup R× = {a ∈ C× ×
p : |a|p = 1}: If a ∈ R ,
×
then there is a root of unity ζ ∈ R such that |a−ζ|p < 1. The functional equation
forces us to set log ζ = 0, so
log a = log ζ + log(ζ −1 a) = log(1 + z)
with |z|p = |ζ −1 a − 1|p < 1. Note that the functional equation follows from the
properties an integral should have:
Zxy Zx Zxy Zx Zy
dt (2) dt dt (4) dt dt
log(xy) = = + = + = log x + log y ,
t t t t t
1 1 x 1 1

where we use that dt/t is invariant under the scaling map t 7→ xt.
This does not tell us, however, how to define log p, say. In fact, it turns out that
we can give log p an arbitrary value λ, but then log is indeed uniquely determined
as
log(upv ) = log u + vλ for |u|p = 1, v ∈ Q.
We can take λ = 0, or (at the other extreme) treat λ as an indeterminate and
consider our integrals to have values in Cp [λ]. We write logλ for the ‘branch of the
logarithm’ with logλ p = λ.
As it turns out, fixing λ already fixes the integral. The following was proved by
Berkovich [Ber, Thm. 9.1.1].

8.1. Theorem. If we fix λ, then there is a unique integral satisfying properties THM
(1) to (5) and Existence
and
Zp uniqueness
dt
(6) = λ. of integral
t
1

(In fact, Berkovich does more: he defines sheaves of functions on X that allow for
iterated integration.)
 § 8. Integration 47

8.2. Example. Consider the open Berkovich annulus A = A(r, R)− . In a similar EXAMPLE
way as for an open disk, the analytic functions on A can be written as Laurent Integral
series on an
∞
X annulus
−3 −2 −1 2
f (t) = . . . + a−3 t + a−2 t + a−1 t + a0 + a1 t + a2 t + . . . = an tn
n=−∞

that converge for all τ ∈ Cp such that r < |τ |p < R. Then f (t) dt = dg(t)+a−1 dt/t
where X an−1
g(t) = tn
n6=0
n
is again an analytic function on A. Recall that A is simply connected, so integrals
depend only on the endpoints. If τ1 , τ2 ∈ Cp satisfy r < |τj |p < R, then we obtain
Zτ2 Zτ2 Zτ2 Zτ2
dt  (1) dt
f (t) dt = dg(t) + a−1 = dg(t) + a−1
t t
τ1 τ1 τ1 τ1
(3,4,6) τ2
= g(τ2 ) − g(τ1 ) + a−1 logλ .
τ1
Here we use the change of variables formula for the map A → Gan m that multiplies
−1
by τ1 and (6). So if we want the integral to be well-defined (i.e., independent of
the choice of λ), we need a−1 = 0, so that f (t) dt = dg(t) is exact. ♣

8.3. Example. Consider an elliptic curve E over Cp (say, p > 2, but this is EXAMPLE
not really necessary) such that vp (j(E)) < 0. Then (as Tate has shown) there is Integral
q ∈ C× ×
p such that |q|p < 1 and E(Cp ) ' Cp /q ; the latter extends to a covering
Z on Tate
map π : Gan an
m → Gm /q ' E
Z an
that exhibits Ganm as the universal covering of E
an ell. curve
an 1
(recall that Gm is simply connected). Let 0 6= ω ∈ Ω (E) be a regular (and
therefore invariant) differential on E and let γ : [0, 1] → E an be the closed path
∗
that is the image under π of the (unique) path from 1 to q in Gan m . Then π ω is
an invariant differential on Gm , so π ∗ ω = α dt/t for some α ∈ C×p . Then
Z Zq Zq
(4) dt (1) dt (6)
ω= α =α = α logλ q .
t t
γ 1 1
λ
There is a unique λ ∈ Cp such that log q = 0; if we choose this value, then
integrals over ω on E will not depend on the path. But this will not work for
two Tate curves simultaneously when their q parameters are not multiplicatively
dependent. So we cannot avoid the path-dependence of the integral. In particular,
there will not be a function f (in the sense of Berkovich’s integration theory) on
all of E an such that df = ω. This is somewhat analogous with the situation over C,
where there is no holomorphic function f on C× such that df = dz/z. ♣

There is a different kind of integral that one can define on elliptic curves, and more
generally, abelian varieties over Cp . We will do this here for regular (= invariant)
1-forms, but the theory can be extended to arbitrary algebraic 1-forms on analytic
spaces associated to algebraic varieties, see [Col].
This is similar to the standard logarithm, which is related to the group structure
on the multiplicative group Gm . Let ω be an invariant 1-form on an abelian
variety A over Cp (for example, an elliptic curve). There is an open (and closed)
neighborhood U of the origin 0 ∈ A(Cp ) that is a subgroup and analytically
isomorphic to an open polydisk D. Every closed 1-form is exact on an open
 § 8. Integration 48
Rx
polydisk, so we obtain a function logω : U → Cp that satisfies logω x = 0 ω
for x ∈ U . By the same argument as before, logω is a homomorphism on U .
Contrary to Gm , A is projective, which implies that A(Cp )/U is a torsion group
(every element has finite order), so in this case there is a unique extension of logω
to all of A(Cp ) as a group homomorphism logω : A(Cp ) → Cp : if x ∈ A(Cp ) is
arbitrary, then there is some n ∈ Z≥1 such that nx ∈ U ; we must then have
logω x = (logω (nx))/n (and this does not depend on the choice of n with nx ∈ U ).
We define, for ω as above and x, y ∈ A(Cp ),
Zy
Ab
ω = logω y − logω x .
x

Then this integral satisfies Properties (1), (2) and (4) for abelian varieties and in-
variant differentials, where in (4) we only consider morphisms as varieties (which
are morphisms as abelian varieties composed with a translation). This integral
depends only on the endpoints and does not involve a choice of path, so Prop-
erty (5) is trivially satisfied. We can extend it to (the Berkovich spaces associated
to) arbitrary smooth projective varieties and closed regular 1-forms as follows. Let
V be such a variety and let A be its Albanese variety; write [y − x] for the image
of (y, x) under the Albanese map V × V → A. Fixing x0 ∈ V , set α : V → A,
x 7→ [x − x0 ]. Then if ω is a closed regular 1-form on V , there is an invariant
1-form ωA on A such that α∗ ωA = ω (this does not depend on the choice of x0 ); in
fact, α∗ gives a canonical isomorphism of the space of invariant 1-forms on A with
the space of closed regular 1-forms on V . We then define the ‘abelian integral’ as
Zy [y−x]
Z
Ab Ab
ω= ωA = logωA [y − x] .
x 0

It satisfies (1), (2), and (4) for morphisms of varieties. If one extends it to arbitrary
algebraic closed 1-forms as in [Col], then it also satisfies (3) (for a suitable class of
functions) and (6). Also this more general integral depends only on the endpoints,
not on a path between them. This makes it clear that it differs from Berkovich’s
integral, which in general does depend on the chosen path.
(This does not contradict the uniqueness statement in Theorem 8.1, because
the abelian integral is restricted to algebraic objects and morphisms, whereas
Berkovich’s theorem talks about integration for analytic 1-forms, spaces and maps.)
The abelian integral has some useful applications, which derive from the following
easy result. The rank of an abelian group G is defined as rk G = dimQ G ⊗Z Q, or DEF
equivalently, the maximal number of elements of G that are linearly independent rank
over Z (rk G := ∞ if this number is unbounded).

8.4. Theorem. Let A be an abelian variety over Cp and let Γ ⊂ A(Cp ) be a THM
subgroup of rank rk Γ = r < g = dim A. Then there are at least g − r linearly 1-forms
independent (over Cp ) invariant 1-forms ω such that logω γ = 0 for all γ ∈ Γ. killing Γ

Proof. Write Ω1 (A) for the Cp -vector space of invariant 1-forms on A. This space
has dimension g = dim A. The map
Ω1 (A) × A(Cp ) −→ Cp , (ω, x) 7−→ logω x
 § 8. Integration 49

is Cp -linear in ω and Z-linear in x. Let γ1 , . . . , γr ∈ Γ be Z-linearly independent;

then for each γ ∈ Γ there are integers m > 0 and m1 , . . . , mr such that
(8.1) mγ = m1 γ1 + . . . + mr γr .
Now let V ⊂ Ω1 (A) be the subspace given by
V = {ω ∈ Ω1 (A) : logω γj = 0 for j = 1, . . . , r} .
Since we are imposing r linear conditions, we have dimCp V ≥ g − r. The rela-
tion (8.1) then implies that logω γ = 0 for all γ ∈ Γ. So any Cp -basis of V does
what is required. q

8.5. Corollary. Let C be a curve of genus g > 0 defined over a number field K, COR
with Jacobian variety J. Let P0 ∈ C(K); denote by i the associated embedding killing
C → J sending P0 to the origin. Let Kv be the completion of K at some p- rational
adic place v. If rk J(K) = r < g, then there are at least g − r > 0 Kv -linearly points on
independent 1-forms ω on C over Kv such that a curve
ZP
Ab
ω=0 for all P ∈ C(K).
P0

Proof. We apply Theorem 8.4 and its proof to Γ = J(K) ⊂ J(Kv ) ⊂ J(Cv ). If we
restrict to 1-forms defined over Kv , the proof goes through as before, working with
vector spaces over Kv instead of over Cp . This gives us ω1 , . . . , ωg−r ∈ Ω1 (J/Kv )
such that logωj x = 0 for all j and all x ∈ J(K). Pulling back 1-forms induces an
isomorphism between Ω1 (J/Kv ) and Ω1 (C/Kv ), and we have for any j and any
P ∈ C(K)
ZP i(P )
Z
Ab Ab
i∗ ωj = ωj = logωj i(P ) = 0 ,
P0 0

so i∗ ω1 , . . . , i∗ ωg−r do what we want. q

We would like to obtain a uniform bound for the number of points in C(K). The
following result allows us to reduce to disks and annuli.

8.6. Lemma. Let C and Kv be as above. Then there is a finite collection of LEMMA
analytic maps φm : D(0, 1)− → C an and ψn : A(ρn , 1)− → C an with ρn ∈ |Kv |× partition
(where C an is defined with respect to Kv ⊂ Cp ) defined over Kv such that of C(/Kv )
[
[
φm D(0, 1)− ∩ Kv ∪ ψn A(ρn , 1)− ∩ Kv ,


C(Kv ) =
m n

and the number of the φm and the ψn each are bounded in terms of g and Kv only.

Proof. See [Sto, Prop. 5.3]. The numer of annuli is ≤ 3g − 3, and the number of
disks is in O(qg), where q is the size of the residue field of Kv . q
 § 8. Integration 50

− −
Write Dm = φm (D(0, 1)
∩ Kv ) and An =
ψn (A(ρn , 1) ∩ Kv ). It then suffices to
bound # C(K) ∩ Dm and C(K) ∩ An for each m and n. Let V ⊂ Ω1 (C/Kv )
be the space of regular 1-forms that kill C(K) as in Corollary 8.5. Then

ZP
Ab
n o
C(K) ∩ Dm ⊂ P ∈ Dm : ω=0
P0

for every ω ∈ V , and similarly for An . For a disk Dm we have for τ ∈ Kv with
|τ |p < 1
φZ
m (τ ) φZ
m (0) φZ
m (τ ) φZ
m (0) Zτ
Ab Ab Ab Ab
ω= ω+ ω= ω+ φ∗m ω =: h(τ )
P0 P0 φm (0) P0 0

(on a disk there is no difference between the two integrals; this comes from the
definition of logω on U by formal integration of power series on the polydisk D).
Let π be a uniformizer of Kv (i.e., |π|p < 1 and generates the value group of Kv ).
Then the number of zeros of h in Kv ∩ D(0, 1)− ⊂ D(0, |π|p ) is bounded in terms
of the number of zeros of φ∗m ω in D(0, 1)− and |π|p (see Exercises). The number of
zeros of φ∗m ω in D(0, 1)− is the same as the number of zeros of ω in φm (D(0, 1)− ),
which is at most the total number of zeros of ω, which in turn is 2g − 2. This gives
us a bound for #(C(K) ∩ Dm ) that only depends on g and Kv (via |π|p ). Since
the total number of disks Dm is bounded in terms of g and Kv , we get a bound
on the total number of K-points of C contained in some Dm that only depends
on g and Kv .
For annuli, the situation is more complicated. There is the following theorem that
compares the two integrals on an annulus.

8.7. Theorem. Fix a Kv -defined analytic map ψ : A(ρ, 1)− → J an . There is a THM
linear form α : Ω1 (J/Kv ) → Kv such that for any x, y ∈ A(ρ, 1)− we have comparison
of integrals
Zy ψ(y)
Ab
Z on annuli
∗


ψ ω− ω = α(ω) vp (y) − vp (x)
x ψ(x)

for all ω ∈ Ω1 (J/Kv ).

Proof. See [Sto, Prop. 7.3] or [KRZB, Prop. 3.29]. q

If r ≤ g − 3, then for each annulus An there will be at least one 0 6= ω ∈ V such

that ψn∗ ω is exact and α(ω) = 0 (note that α depends on the annulus), since this
imposes two linear conditions. If ψn∗ ω = dh is exact, then we can again bound the
number of zeros of h in Kv ∩ A(ρn , 1)− ⊂ A(ρn /|π|p , |π|p ) in terms of g and Kv ,
under some additional assumption (the ‘relevant’ range of exponents in the Laurent
series representation of h includes 0), which can be shown to always hold (this is
done in [Sto, Cor. 6.7] for hyperelliptic curves and in [KRZB, Lemma 4.15] in
general). This finally leads to the following result.
 § 8. Integration 51

8.8. Theorem. Fix d ≥ 1 and g ≥ 3. There is some B = B(d, g) depending only THM
on d and g such that for any number field K with [K : Q] ≤ d and any curve C of uniform
genus g over K, with Jacobian J such that rk J(K) ≤ g −3, we have #C(K) ≤ B. bound on
rational
Proof. We can assume that C(K) 6= ∅; let P0 ∈ C(K) and use it as the base-point points
for an embedding i : C → J as in Corollary 8.5. Fix some prime p; there are then
only finitely many possibilities (up to isomorphism) for the completion Kv at a
p-adic place v. Let B(d, g) be the maximum of the bounds for C(K) obtained by
the reasoning above for g and each Kv . The claim follows. q

For K = Q, one can write down explicit bounds. For hyperelliptic curves (which
allow for quite explicit computations) a possible bound is
(
1 if r = 0 (and g ≥ 3),
#C(Q) ≤ 33(g − 1) +
8rg − 1 if 1 ≤ r ≤ g − 3,
where r = rk J(Q), see [Sto, Thm. 9.1] (this works with Q3 ). In the general case,
one obtains a bound of the shape O(g 2 ), see [KRZB, Thm. 5.1]. For arbitrary d,
one still has a bound of the shape Od (g 2 ), but the implied constant depends
exponentially on d.
 Literatur 52

References
[BR] Matthew Baker, Robert Rumely: Potential theory and dynamics on the Berkovich
projective line. Mathematical Surveys and Monographs 159. American Mathematical
Society, Providence, RI, 2010.
[AWS] Matthew Baker, Brian Conrad, Samit Dasgupta, Kiran S. Kedlaya, Jeremy
Teitelbaum: p-adic geometry. Lectures from the 10th Arizona Winter School held at
the University of Arizona, Tucson, AZ, March 10–14, 2007. Edited by David Savitt
and Dinesh S. Thakur. University Lecture Series, 45. American Mathematical Society,
Providence, RI, 2008. xii+203 pages.
[Ber] Vladimir G. Berkovich: Integration of one-forms on p-adic analytic spaces. Annals
of Mathematics Studies 162. Princeton University Press, Princeton, NJ, 2007. vi+156
pages.
[Col] Pierre Colmez: Intégration sur les variétés p-adiques. Astérisque 248, Société
Mathématique de France, Paris, 1998. viii+155 pages.
[KRZB] Eric Katz, Joseph Rabinoff and David Zureick-Brown: Uniform bounds for
the number of rational points on curves of small Mordell-Weil rank, Preprint (2015),
arXiv:1504.00694v2 [math.NT].
[Sto] Michael Stoll: Uniform bounds for the number of rational points on hyperelliptic
curves of small Mordell-Weil rank, Preprint (2015), arXiv:1307.1773v6 [math.NT].