MAXIMALLY COMPLETE FIELDS

BJORN POONEN

This article has been published in L’Enseign. Math. 39 (1993), 87–

106.

Abstract. Kaplansky proved in 1942 that among all fields with

a valuation having a given divisible value group G, a given alge-
braically closed residue field R, and a given restriction to the min-
imal subfield (either the trivial valuation on Q or Fp , or the p-adic
valuation on Q), there is one that is maximal in the strong sense
that every other can be embedded in it. In this paper, we construct
this field explicitly and use the explicit form to give a new proof of
Kaplansky’s result. The field turns out to be a Mal’cev-Neumann
ring or a p-adic version of a Mal’cev-Neumann ring in which the
elements are formal series of the form g∈S αg pg where S is a well-
P
ordered subset of G and the αg ’s are residue class representatives.
We conclude with some remarks on the p-adic Mal’cev-Neumann
field containing Q̄p .

1. Introduction
It is well known that if k is an algebraically closed field of charac-
teristic zero, then the algebraic closure of the field of Laurent series
k((t)) is obtained by adjoining t1/n for each integer n ≥ 1, and that
the expansion of a solution to a polynomial equation over k((t)) can be
obtained by the method of successive approximation. (For example, to
find a square root of 1 + t, one solves for the coefficients of 1, t, t2 , . . . in
turn.) But if k is algebraically closed of characteristic p, ∞ 1/n
S
n=1 k((t ))
is no longer an algebraic closure of k((t)). In particular, the Artin-
Schreier equation xp − x = t−1 has no solution in ∞ 1/n
S
n=1 k((t )). (See
p. 64 of Chevalley [3].) If one attempts nevertheless to successively ap-
proximate a solution, one obtains the expansion (due to Abhyankar [1])
2 3
x = t−1/p + t−1/p + t−1/p + · · · ,
in which the exponents do not tend to ∞, as they should if the series
were to converge with respect to a valuation in the usual sense. How-
ever, one checks (using the linearity of the Frobenius automorphism)

Date: March 4, 1992.

1
2 BJORN POONEN

that this series does formally satisfy our polynomial equation! (The
other solutions are obtained by adding elements of Fp to this one.)
It is natural to seek a context in which series such P as these make
sense. If one tries to define a field containing all series q∈Q αq tq , one
fails for the reason that multiplication is not well defined. But then
one notices that a sequence of exponents coming from a transfinite
successive approximation process must be well-ordered. If one considers
only series in which the set of exponents is a well-ordered subset of Q,
one does indeed obtain a field.
Such fields are commonly known as Mal’cev-Neumann rings. (We
will review their construction in Section 3.) They were introduced by
Hahn in 1908, and studied in terms of valuations by Krull [8] in 1932.
(Mal’cev [11] in 1948 and Neumann [12] in 1949 showed that the same
construction could be performed for exponents in a non-abelian group
to produce a division ring.)
If one tries to find p-adic expansions of elements algebraic over Qp ,
one encounters a similar situation. One is therefore led to construct
p-adic analogues of the Mal’cev-Neumann rings. (See Section 4.) This
construction is apparently new, except that Lampert [9] in 1986 de-
scribed the special case of value group Q and residue field F̄p without
giving details of a construction. (We will discuss this special case in
detail in Section 7.)
In Section 5 we prove our main theorems. A corollary of our The-
orem 2 is that a Mal’cev-Neumann ring (standard or p-adic) with di-
visible value group G and algebraically closed residue field R has the
amazing property that every other valued field with the same value
group, the same residue field, and the same restriction to the minimal
subfield (either the trivial valuation on Q or Fp , or the p-adic valuation
on Q) can be embedded in the Mal’cev-Neumann ring. (We assume
implicitly in the minimal subfield assumption that in the p-adic case
the valuation of p must be the same element of G for the two fields.)
Kaplansky [5] proved the existence of a field with this property using
a different method. He also knew that it was a Mal’cev-Neumann ring
when the restriction of the valuation to the minimal subfield is trivial,
but was apparently unaware of its structure in the p-adic case.

2. Preliminaries
All ordered groups G in this paper are assumed to be abelian, and
we write the group law additively. We call G divisible if for every g ∈ G
and positive integer n, the equation nx = g has a solution in G. Every
ordered group can be embedded in a divisible one, namely its injective
 MAXIMALLY COMPLETE FIELDS 3

hull. Since an ordered group G is necessarily torsion-free, its injective

hull G̃ can be identified with the set of quotients g/m with g ∈ G,
m a positive integer, modulo the equivalence relation g/m ∼ h/n iff
ng = mh in G. We make G̃ an ordered group by setting g/m ≥ h/n
iff ng ≥ mh in G. (One can check that this is the unique extension to
G̃ of the ordered group structure on G.)
If G is an ordered group, let G∞ = G ∪ {∞} be the ordered monoid
containing G in which g + ∞ = ∞ + g = ∞ for all g ∈ G∞ and g < ∞
for all g ∈ G. As usual, a valuation v on a field F is a function from F
to G∞ satisfying for all x, y ∈ F
(1) v(x) = ∞ iff x = 0
(2) v(xy) = v(x) + v(y)
(3) v(x + y) ≥ min{v(x), v(y)}.
The value group is G. The valuation ring A is { x ∈ F | v(x) ≥ 0 }.
This is a local ring with maximal ideal M = { x ∈ F | v(x) > 0 }. The
residue field is A/M. We refer to the pair (F, v) (or sometimes simply
F ) as a valued field.

3. Mal’cev-Neumann rings
This section serves not only as review, but also as preparation for the
construction of the next section. Mal’cev-Neumann rings are general-
izations of Laurent series rings. For any ring R (all our rings are com-
mutative with 1), and any ordered group G, the Mal’cev-Neumann ring
R((G)) is defined as the set of formal sums α = g∈G αg tg in an inde-
P

terminate t with αg ∈ R such that the set Supp α = { g ∈ G | αg 6= 0 }

is a well-ordered subset of G (under the given order of G). (Often
authors
P
suppress the indeterminate and write the sums in the form
αg g, as in a group ring. We use the indeterminate in orderPto make
clear the analogy with the fields of the next section.) If α = g∈G αg tg
and β = g∈G βg tg are elements of R((G)), then α + β is defined as
P
P g P j
g∈G (αg + βPg )t , and αβ is defined by a “distributive law” as j∈G γj t
where γj = g+h=j αg βh .
Lemma 1. Let A, B be well-ordered subsets of an ordered group G.
Then
(1) If x ∈ G, then A ∩ (−B + x) is finite. (We define −B + x =
{ −b + x | b ∈ B }.)
(2) The set A + B = { a + b | a ∈ A, b ∈ B} is well-ordered.
(3) The set A ∪ B is well-ordered.
Proof. See [13]. 
4 BJORN POONEN

The lemma above easily implies that the sum defining γj is always
finite, and that Supp (α + β) and Supp (αβ) are well-ordered. Once
one knows that the operations are defined, it’s clear that they make
R((G)) a ring.
Define v : R((G)) → G∞ by v(0) = ∞ and v(α) = min Supp α for
α 6= 0. (This makes sense since Supp α is well-ordered.) If α ∈ R((G))
is nonzero and v(α) = g, we call αg tg the leading term of α and αg the
leading coefficient. If R is a field, then v is a valuation on R((G)), since
the leading term of a product is the product of the leading terms.
Lemma 2. If α ∈ R((G)) satisfies v(α) > 0, then 1 − α is a unit in
R((G)).
Proof. One way of proving this is to show that for each g ∈ G, the
coefficients of tg in 1, α, α2 , . . . are eventually zero, so 1+α+α2 +· · · can
be defined termwise. Then one needs to check that its support is well-
ordered, and that it’s an inverse for 1 − α. See [13] for this. An easier
way [15] is to obtain an inverse of 1−α by successive approximation. 
Corollary 1. If the leading coefficient of α ∈ R((G)) is a unit of R,
then α is a unit of R((G)).
Proof. Let rtg be the leading term of α. Then α is the product of rtg ,
which is a unit in R((G)) with inverse r−1 t−g , and (rtg )−1 α, which is a
unit by the preceding lemma. 
Corollary 2. If R is a field, then R((G)) is a field.
So in this case, if we set K = R((G)), (K, v) is a valued field. Clearly
the value group is all of G, and the residue field is R. Note that
char K = char R, since in fact, R can be identified with a subfield of
K. (We will refer to these fields as being the “equal characteristic”
case, in contrast with the p-adic fields of the next section in which the
fields have characteristic different from that of their residue fields.) For
example, if G = Z, then R((G)) is the usual field of formal Laurent
series.

4. p-adic Mal’cev-Neumann fields

To construct analogous examples of characteristic zero whose residue
field has nonzero characteristic requires a more complicated construc-
tion. First we recall two results about complete discrete valuation rings.
For proofs, see [17], pp. 32–34.
A valued field (F, v) is called discrete if v(F ) = Z.
 MAXIMALLY COMPLETE FIELDS 5

Proposition 1. If R is a perfect field of characteristic p > 0, then there

exists a unique field R0 of characteristic 0 with a discrete valuation v
such that the residue field is R, v(p) = 1 ∈ Z, and R0 is complete with
respect to v. (The valuation ring A of R0 is called the ring of Witt
vectors with coefficients in R.)
For example, if R = Fp , then R0 = Qp with the p-adic valuation.
Proposition 2. Suppose F is field with a discrete valuation v, and
t ∈ F satisfies v(t) = 1. Let S ⊂ F be a set of representatives for the
residue classes with 0 ∈ S. Then every element x ∈ F can be written
uniquely as m∈Z xm tm , where xm ∈ S for each m, and xm = 0 for all
P

sufficiently negative m. Conversely, if F is complete, every such series

defines an element of F .
Now for the construction. Let R be a perfect field of characteris-
tic p, and let G be an ordered group containing Z as a subgroup, or
equivalently with a distinguished postive element. (When we eventu-
ally define our valuation v, this element 1 ∈ G will be v(p).) Let A be
the valuation ring of the valued field (R0 , v 0 ) given by Proposition 1.
What we want is to have the indeterminate t stand for p in elements
of A((G)), so we get elements of the form g∈G αg pg . The problem is
P

that some elements of A((G)), like −p + t1 , “should be” zero. So what

we do is to take a quotient A((G))/N where N ⊂ A((G)) is the ideal
of elements that “should be” zero.
We say that α = g αg tg ∈ A((G)) is a null series if for all g ∈ G,
P
P n 0
n∈Z αg+n p = 0 in R . (Recall that we fixed a copy of Z in G.) Note
that αg+n = 0 for sufficiently negative n, since otherwise Supp α would
0
not be well-ordered. Also, v (αg+n p ) ≥ n, so n∈Z αg+n pn always
n P

converges in R0 . Let N be the set of null series.

Proposition 3. N is an ideal of A((G)).
Proof. Clearly N is an additive subgroup. LetPG0 ⊂ G be a set of
coset representatives for G/Z. Suppose α = g∈G αg tg ∈ A((G)),
β = h∈G βh th ∈ N , and αβ = j∈G γj tj . Then for each j ∈ G,
P P

γj+n pn = αg βh pn
X X

n∈Z
g+h=j+n
n∈Z

(αj−h0 +l pl )(βh0 +m pm )
X
=
h0 ∈G0
l,m∈Z

(We write h = h0 + m with h0 ∈ G0 and let l = n − m.)

6 BJORN POONEN

Since β ∈ N , m∈Z βh0 +m pm = 0 for each h0 ∈ G, so we get

P
P n 0
n∈Z γj+n p = 0. (These infinite series manipulations in R are valid,
because for each i ∈ Z, only finitely many terms have valuation less
than i, since each γj+n is a finite sum of products αg βh .) Hence N is
an ideal. 
Define the p-adic Mal’cev-Neumann field L as A((G))/N .
Proposition 4. Let S ⊂ A be a set of representatives for the residue
classes of A, with 0 ∈ S. Then any element α = g∈G αg tg ∈ A((G))
P

is equivalent modulo N to a element β = g∈G βg tg with each βg in S.

P

Moreover, β is unique.
Proof. Let G0 ⊂ G be a set of coset representatives for G/Z. For each
g ∈ G0 , we may write
αg+n pn = βg+n pn
X X

n∈Z n∈Z

R0 is complete
with βg+n ∈ S, by Proposition 2. (This is possiblePsince P
with respect to its discrete valuation.) Then β = g∈G0 n∈Z βg+n tn is
a well-defined element of A((G)), since Supp (β) ⊆ (Supp α)+N, which
is well-ordered by part 2 of Lemma 1. Finally α − β ∈ N , by definition
of N . The uniqueness follows from the uniqueness in Proposition 2. 
Corollary 3. L=A((G))/N is a field.
Proof. The previous proposition shows that any α ∈ A((G)) is equiva-
lent modulo N to 0 or an element which is a unit in A((G)) by Corol-
lary 1. 
Proposition 4 allows us to P
write an element of L uniquely (and
somewhat carelessly) as β = g∈G βg pg , with βg ∈ S. Thus given
S, we can speak of Supp (β) for β ∈ L. Define v : L → G∞ by
v(β) = min Supp β.
Proposition 5. The map v is a valuation on L, and is independent of
the choice of S. The value group is G and the residue field is R.
αg tg ∈ A((G)), define
P
Proof. For α = g∈G
 
 
w(α) = min g + v 0 ( αg+n pn ) .
X
g∈G  
n∈Z

The elements in the “min” belong to (Supp (α) + N) ∪ {∞}, which

is well-ordered by part 2 of Lemma 1, so this is well defined. It’s
clearly unchanged if an element of N is added to α. In particular,
if we do so to get an element α0 ∈ A((G)) with coefficients in S, we
 MAXIMALLY COMPLETE FIELDS 7

find w(α) = w(α0 ) = min Supp α0 . Thus if β is the image of α in

L, v(α) = w(β). Since w is independent of the choice of S, so is
v. If α0 , β 0 are the representatives in A((G)) with coefficients in S
of elements α, β ∈ L, then it is clear that w(α0 β 0 ) = w(α0 ) + w(β 0 )
(because the leading coefficient of α0 β 0 has valuation 0 under v 0 ) and
that w(α0 + β 0 ) ≥ min{w(α0 ), w(β 0 )}. Thus v is a valuation.
The value group of v is all of G, since v(pg ) = g for any g ∈ G.
The natural inclusion A ⊂ A((G)) composed with the quotient map
A((G)) → L maps A into the valuation ring of L, which consists of
series g≥0 αg pg , so it induces a map φ from A to the residue field of
P

L. The residue class of g≥0 αg pg equals φ(α0 ) ∈ A (since the maximal

P

ideal for L consists of series g>0 αg pg ). Thus φ is surjective. Its kernel

P

is the maximal ideal of A, so φ induces an isomorphism from the residue

class field of A to that of L. 
For example, if R is any perfect field of characteristic p, and G =
−1
k Z for some k ≥ 1 (with its copy of Z as a subgroup of index k),
√
then L = R0 ( k p) with the p-adic valuation.
Lemma 3. If α = g∈G αg pg and β = g∈G βg pg with αg , βg ∈ S are
P P

two elements of L, then v(α − β) = min{ g ∈ G | αg 6= βg }. (The

corresponding fact for the usual Mal’cev-Neumann fields is obvious.)
Proof. LetPw be the map used P in the proof of the previous proposition.
0 0
Let α = g∈G αg t and β = g∈G βg tg in A((G)). Then v(α − β) =
g

w(α0 − β 0 ). If g0 = min{ g ∈ G | αg 6= βg }, then the leading term of

α0 − β 0 is (αg0 − βg0 )tg0 , and the leading coefficient here has valuation 0
under v 0 , since αg0 , βg0 represent distinct residue classes, so w(α0 −β 0 ) =
g0 , as desired. 
Remarks . Since the construction of A from R is functorial (the Witt
functor), it is clear that the construction of L from R is functorial as
well (for each G). However, whereas the Witt functor is fully faithful on
perfect fields of characteristic p, this new functor is not. For example,
Proposition 11 (to be proved in Section 7) shows L can have many
continuous (i.e. valuation-preserving) automorphisms not arising from
automorphisms of R.
Our construction could be done starting from a non-abelian value
group to produce p-adic Mal’cev-Neumann division rings, but we will
not be interested in such objects.
5. Maximality of Mal’cev-Neumann fields
A valued field (E, w) is an immediate extension of another valued
field (F, v) if
8 BJORN POONEN

(1) E is a field extension of F , and w|F = v.

(2) (E, w) and (F, v) have the same value groups and residue fields.
A valued field (F, v) is maximally complete if it has no immediate ex-
tensions other than (F, v) itself. (These definitions are due to F. K.
Schmidt, but were first published by Krull [8].) For example, an easy
argument shows that any field F with the trivial valuation, or with a
discrete valuation making it complete, is maximally complete.
Proposition 6. Let (F, v) be a maximally complete valued field with
value group G and residue field R. Then
(1) F is complete.
(2) If R is algebraically closed and G is divisible, then F is alge-
braically closed.
Proof. (1) The completion F̂ of F is an immediate extension of F
(see Proposition 5 in Chapter VI, §5, no. 3 of [2]), so F̂ = F .
(2) The algebraic closure F̄ of F is in this case an immediate ex-
tension of F (see Proposition 6 in Chapter VI, §3, no. 3 and
Proposition 1 in Chapter VI, §8, no. 1 of [2]), so F̄ = F .
(This delightful trick is due to MacLane [10].) 
Proposition 7. Any continuous endomorphism of a maximally com-
plete field F which induces the identity on the residue field is automat-
ically an automorphism (i.e., surjective).
Proof. The field F is an immediate extension of the image of the endo-
morphism, which is maximally complete since it’s isomorphic to F . 

From now on, when we refer to Mal’cev-Neumann fields, we mean

one of the two fields K or L from the previous two sections. Let these
have valuation v with value group G and residue field R. From now
on, the proofs for the equal characteristic case K will be the same as
(or easier than) those for the p-adic case L, so we will only give proofs
for L. (To get a proof for K, simply replace pg with tg , and replace the
set S of representatives with R.)
We will use the following lemma to show K and L are maximally
complete.
Lemma 4. Let (F, v) be a valued field with value group G. Suppose
we have an arbitrary system of inequalities of the form v(x − aσ ) ≥ gσ ,
with aσ ∈ F and gσ ∈ G for all σ in some index set I. Then
(1) If the system has a solution x ∈ F , then v(aσ1 −aσ2 ) ≥ min{gσ1 , gσ2 }
for all σ1 , σ2 ∈ I.
 MAXIMALLY COMPLETE FIELDS 9

(2) Suppose in addition that F = L (or K) is one of the Mal’cev-

Neumann fields. Then the converse is true; i.e., if v(aσ1 −aσ2 ) ≥
min{gσ1 , gσ2 } for all σ1 , σ2 ∈ I, then the system has a solution.

Proof. (1) This is simply a consequence of the triangle inequality.

(2) Suppose v(aσ1 − aσ2 ) ≥ min{gσ1 , gσ2 } for all σ1 , σ2 ∈ I. For each
g ∈ G, let xg be the coefficient of pg in aσ for any σ for which gσ > g,
and let xg = 0 if no such σ exists. We claim xg is uniquely defined. For
if gσ1 , gσ2 > g, then v(aσ1 − aσ2 ) > g, so by Lemma 3 the coefficients of
pg in aσ1 , aσ2 must be the same.
Define x = g∈G xg pg . To show x ∈ L, we must check that Supp x
P

is well-ordered. Suppose h1 , h2 , . . . is a strictly descending sequence

within Supp x. Then by definition of xg , h1 < gσ for some σ ∈ I, and
hn ∈ Supp aσ for all n ≥ 1. This is a contradiction, since Supp aσ is
well-ordered. Thus x ∈ L.
By definition of xg , the coefficients of pg in x and aσ agree for g < gσ .
From Lemma 3 it follows that v(x − aσ ) ≥ gσ . 

Theorem 1 (Krull [8]). The Mal’cev-Neumann fields K and L are

maximally complete. (Actually, Krull proved this only for the equal
characteristic case (K), but his proof applies equally well to the p-adic
fields L.)

Proof. (As usual, we treat only the p-adic case.) Suppose (M, w) is a
proper immediate extension of (L, v). Fix µ ∈ M \ L. Consider the
system of inequalites w(x − aσ ) ≥ gσ , where aσ ranges over all elements
of L and gσ = w(µ − aσ ). Obviously µ is a solution (in M ), so by
part 1 of Lemma 4, w(aσ1 − aσ2 ) ≥ min{gσ1 , gσ2 } for all σ1 , σ2 . Now
v(aσ1 − aσ2 ) = w(aσ1 − aσ2 ) ≥ min{gσ1 , gσ2 }, so we may apply part 2 of
Lemma 4 to deduce that the system of inequalites v(x − aσ ) ≥ gσ has
a solution λ ∈ L.
The idea is that λ is a best approximation in L to µ. We will
contradict this by adding the “leading term” of the difference µ − λ
to λ to get a better one. Since µ 6∈ L, µ − λ 6= 0, so we can let
g = w(µ − λ) ∈ G. (Here we are using that L and M have the same
value group.) Then w(p−g (µ − λ)) = 0, so there exists a unique repre-
sentative s ∈ S for the (nonzero) residue class containing p−g (µ − λ).
(Here we are using that L and M have the same residue field.) Then
w(p−g (µ − λ) − s) > 0, so w(µ − λ − sp−g ) > g. On the other hand,
g = v(−spg ) = v(λ − (λ + spg )) ≥ w(µ − (λ + spg )), by the definition
of λ, using aσ = λ + spg . This contradiction proves L is maximally
complete. 
10 BJORN POONEN

Remark . It is true in general that F is maximally complete iff part 2 of

Lemma 4 is true for F . See Kaplansky’s discussion of pseudolimits [5],
and Theorem 5 in Chapter I of [4].
Corollary 4. Any Mal’cev-Neumann field is complete. A Mal’cev-
Neumann field with divisible value group and algebraically closed residue
field is itself algebraically closed.
Proof. Combine the previous theorem with Proposition 6. 
Remark . In practice, to find solutions to a polynomial equation over
a Mal’cev-Neumann field, one can use successive approximation. This
method could be used to give another (much messier) proof that these
Mal’cev-Neumann fields are algebraically closed.
We will show that the Mal’cev-Neumann fields K and L are maximal
in a sense much stronger than Theorem 1 implies. This will be made
precise in Corollary 5.
Theorem 2. Suppose L (or K) is a Mal’cev-Neumann field with valu-
ation v having divisible value group G and algebraically closed residue
field R. Suppose E is a subfield of L, and that (F, w) is a valued
field extension of (E, v), with value group contained in G and residue
field contained in R. Then there exists an embedding of valued fields
φ : F → L which extends the inclusion E ,→ L.
Proof. Since G is divisible and R is algebraically closed, we can extend
the valuation on F to a valuation on F̄ with value group in G and
residue field in R, by Proposition 6 in Chapter VI, §3, no. 3 and
Proposition 1 in Chapter VI, §8, no. 1 of [2]. If we could find an
embedding of F̄ into L, we would get an embedding of F into L. Thus
we may assume that F is algebraically closed.
Let C be the collection of pairs (E 0 , φ) such that E 0 is a field between
E and F and φ : E 0 → L is an embedding of valued fields. Define
a partial order on C by saying (E20 , φ2 ) is above (E10 , φ1 ) if E20 ⊇ E10
and φ2 extends φ1 . By Zorn’s Lemma, we can find a maximal element
(E 0 , φ) of C. By relabeling elements, we can assume E 0 ⊆ L, and we
may as well rename E 0 as E.
We claim this E is algebraically closed. Both F and L are alge-
braically closed. (For L, this follows from Corollary 4.) So we have an
algebraic closure of E in F and in L, each with a valuation extending
the valuation on E. By Corollary 1 in Chapter VI, §8, no. 6 of [2], two
such valuations can differ only by an automorphism of Ē over E; i.e.,
there exists a continuous embedding of the algebraic closure of E in F
into L. By maximality of (E, φ) in C, E must be algebraically closed
already.
 MAXIMALLY COMPLETE FIELDS 11

If E = F , we are done, so assume there is some element µ ∈ F \ E.

We will define a corresponding element µ0 ∈ L.

Case 1: There exists a best approximation e0 ∈ E to µ; i.e. there

exists e0 ∈ E such that w(µ − e) ≤ w(µ − e0 ) for all e ∈ E. Let
g = w(µ − e0 ) ∈ G.
Case 1a: g 6∈ v(E). Then define µ0 = e0 + pg .
Case 1b: g = v(δ) for some δ ∈ E. Then w(δ −1 (µ − e0 )) = 0,
so we let s ∈ S be the representative of the (nonzero) residue class
corresponding to δ −1 (µ − e0 ) ∈ F , and define µ0 = e0 + sδ.

Note that in these cases, v(µ0 − e0 ) = g, so for all e ∈ E,

v(µ0 − e) ≥ min{v(µ0 − e0 ), v(e − e0 )} (the triangle inequality)
= min{g, v(e − e0 )}
= min{w(µ − e0 ), w(e − e0 )} (since v and w agree on E)
≥ min{w(µ − e0 ), w(µ − e), w(µ − e0 )} (the triangle inequality)
= w(µ − e) (by definition of e0 ).

Case 2: For every e ∈ E, there exists e0 ∈ E with w(µ−e0 ) > w(µ−e).

Consider the system of inequalities w(x − eσ ) ≥ gσ , where eσ ranges
over all elements of E and gσ = w(µ − eσ ). Since µ is a solution (in
F ), w(eσ1 − eσ2 ) ≥ min{gσ1 , gσ2 } by part 1 of Lemma 4. We have
v(eσ1 − eσ2 ) = w(eσ1 − eσ2 ) ≥ min{gσ1 , gσ2 },
so by part 2 of Lemma 4, the system of inequalities v(x − eσ ) ≥ gσ has
a solution µ0 in L.

Claim: In all cases, w(µ − e) = v(µ0 − e) for all e ∈ E.

Proof: From the remarks at the end of Case 1, and by the definition
of µ0 in Case 2, we have w(µ − e) ≤ v(µ0 − e) for all e ∈ E.
First suppose e is not a best approximation to µ, so w(µ − e0 ) >
w(µ−e), for some e0 ∈ E. Then equality holds in the triangle inequality,
w(e − e0 ) = w((µ − e0 ) − (µ − e)) = w(µ − e)
so
v(e − e0 ) = w(e − e0 ) = w(µ − e) < w(µ − e0 ) ≤ v(µ0 − e0 ).
Again equality holds in the triangle inequality, so we get
v(µ0 − e) = v((µ0 − e0 ) − (e − e0 )) = v(e − e0 ) = w(µ − e)
which proves the claim in this case.
12 BJORN POONEN

Thus we are left with the case in which w(µ − e0 ) ≤ w(µ − e) for all
0
e ∈ E. Then Case 1 holds and w(µ − e) = w(µ − e0 ) = g. Suppose
v(µ0 − e) > g. Then applying the triangle equality to e − e0 = (µ0 −
e0 )−(µ0 −e) and using v(µ0 −e0 ) from our remarks at the end of Case 1,
we get v(e − e0 ) = v(µ0 − e0 ) = g. Thus g ∈ v(E) so we must be in
Case 1b. Moreover
 
v δ −1 (µ0 − e0 ) − δ −1 (e − e0 ) = v(δ −1 ) + v(µ0 − e) > −g + g = 0

so δ −1 (µ0 − e0 ) and δ −1 (e − e0 ) have the same image in the residue field

R. But by definition of µ0 in Case 1b, δ −1 (µ0 − e0 ) has the same image
in R as δ −1 (µ − e0 ). Combining these facts gives us
 
w δ −1 (µ − e0 ) − δ −1 (e − e0 ) > 0
so w(µ − e) > w(δ) = v(δ) = g, contradicting the definitions of g
and e0 . Thus we cannot have v(µ0 − e) > g. But we know v(µ0 − e) ≥
w(µ−e) = g, so we must have v(µ0 −e) = w(µ−e) = g. This completes
the proof of the claim.

Since µ 6∈ E, v(µ0 − e) = w(µ − e) 6= ∞ for all e ∈ E. Hence µ0 6∈ E.

But E is algebraically closed, so µ and µ0 are transcendental over E,
and we have an isomorphism of fields Φ : E(µ) → E(µ0 ) over E which
maps µ to µ0 .
We claim that Φ preserves the valuation. (The valuations on E(µ),
E(µ0 ) are the restrictions of w, v respectively.) Since E is algebraically
closed, any element ρ ∈ E(µ) can be written
ρ = 0 (µ − 1 )n1 (µ − 2 )n2 · · · (µ − k )nk ,
for some  ∈ E and ni ∈ Z. By the Claim above, and the fact that v
and w agree on E, it follows that w(ρ) = v(Φ(ρ)), as desired.
But (E(µ), Φ) contradicts the maximality of (E, φ) in C. Thus we
must have had E = F , so we are done. 
Corollary 5. Let (F, v) be a valued field with value group contained in a
divisible ordered group G, and residue field contained in an algebraically
closed field R. Define K and L as usual as the Mal’cev-Neumann fields
with value group G and residue field R. (Define the p-adic Mal’cev-
Neumann field L only if char R > 0.) Then there exists an embedding
of valued fields φ : F → K or φ : F → L, depending on if the restriction
of v to the minimal subfield of F is the trivial valuation (on Q or Fp )
or the p-adic valuation on Q.
Proof. Apply Theorem 2 with E as the minimal subfield. 
 MAXIMALLY COMPLETE FIELDS 13

Corollary 6. Every valued field F has at least one immediate extension

which is maximally complete. If the value group G is divisible and the
residue field R is algebraically closed, then there is only one (up to
isomorphism).
Proof. Embed F in a Mal’cev-Neumann field L (or K) with value group
G̃ and residue field R̄, according to the previous corollary. Let C be
the collection of valued subfields of L which are immediate extensions
of F . By Zorn’s Lemma, C has a maximal element M . If M had an
immediate extension M 0 , then by Theorem 2, we could embed M 0 in
L. This would contradict the maximality of M .
If G is divisible and the R is algebraically closed, then any maximally
complete immediate extension M of F can be embedded in L, and L
is an immediate extension of M , so L = M . 
Remarks . Krull [8] was the first to prove that every valued field F had
a maximal extension. His proof involves showing directly that there
exists a bound on the cardinality of a valued field with given value
group and residue field. Then Zorn’s Lemma is applied.
Kaplansky [5] has investigated in detail the question of when the
maximally complete immediate extension is unique. He has found
weaker conditions on the value group and residue field which guar-
antee this extension is unique. If char R = 0, the extension is unique.
If char R = p > 0, the extension is unique if the following pair of
conditions is satisfied:
(1) Any equation of the form
n n−1
x p + a1 x p + · · · + an−1 xp + an x + an+1 = 0
with coefficients in R has a root in R.
(2) The value group G satisfies G = pG.
Also if G is discrete of arbitrary rank and char F = char R, then the
extension in unique [6]. But Kaplansky gives examples where the ex-
tension is not unique. The exact conditions under which the extension
is unique are not known.

6. Applications
One application of Theorem 2 is to the problem of “glueing” two
valued fields. (This result can also be proved directly without the
use of Mal’cev-Neumann fields; it is equivalent to Exercise 2 for §2 in
Chapter VI of [2]. Our method has the advantage of showing that the
value group of the composite field can be contained in any divisible
14 BJORN POONEN

value group large enough to contain the value groups of the fields to
be glued.)

Proposition 8. Suppose E, F, F 0 are valued fields and that we are

given embeddings of valued fields φ : E → F , φ0 : E → F 0 . Then there
exist a Mal’cev-Neumann field L (or K) and embeddings of valued fields
Φ : F → L, Φ0 : F 0 → L such that Φ ◦ φ = Φ0 ◦ φ0 .

Proof. By the glueing theorem for ordered groups [14], we can assume
the value groups of F and F 0 are contained in a single ordered group G.
Also we can assume that their residue fields are contained in a field R.
Moreover, we may assume G is divisible and R is algebraically closed.
Then E can be embedded as a valued subfield of a power series field L
(or K) with value group G and residue field R, by Corollary 5. Finally,
Theorem 2 gives us the desired embeddings Φ, Φ0 . 

Remark . Transfinite induction can be used to prove the analogous

result for glueing an arbitrary collection of valued fields.

Since a non-archimedean absolute value on a field can be interpreted

as a valuation with value group contained in R, we can specialize the
results of Section 5 to get results about fields with non-archimedean
absolute values. For example, Corollary 5 implies the following, which
may be considered the non-archimedean analogue of Ostrowski’s theo-
rem that any field with an archimedean absolute value can be embedded
in C with its usual absolute value (or one equivalent).

Proposition 9. Let (F, | |) be a field with a non-archimedean absolute

value, and suppose the residue field is contained in the algebraically
closed field R. Define K and L as the Mal’cev-Neumann fields with
value group R and residue field R. (Define the p-adic Mal’cev-Neumann
field L only if char R > 0.) The valuations on K and L induce
corresponding absolute values. Then there exists an absolute value-
preserving embedding of fields φ : F → K or φ : F → L, depending
on if the restriction of | | to the minimal subfield of F is the trivial
absolute value (on Q or Fp ) or the p-adic absolute value on Q.

Similarly, Proposition 8 above gives a glueing proposition for non-

archimedean absolute values. In fact, this result holds for archimedean
absolute values as well, in light of Ostrowski’s theorem.
 MAXIMALLY COMPLETE FIELDS 15

7. Example: the maximally complete immediate extension

of Q̄p
For this section, (L, v) will denote the p-adic Mal’cev-Neumann field
having value group Q and residue field F̄p . We have a natural embed-
ding of Qp into L. By Corollary 4, L is algebraically closed, so this
embedding extends to an embedding of Q̄p into L (which is unique
up to automorphisms of Q̄p over Qp .) In fact this embedding is con-
tinuous, since there is a unique valuation on Q̄p extending the p-adic
valuation on Qp . Since Q̄p has value group Q and residue field F̄p , L is
an immediate extension of Q̄p . By Corollary 6, L is in fact the unique
maximally complete immediate extension of Q̄p . Also, any valued field
(F, w) of characteristic 0 satisfying
(1) The restriction of w to Q is the p-adic valuation.
(2) The value group is contained in Q.
(3) The residue field is contained in F̄p .
can be embedded in L, by Corollary 5. For example, the completion
Cp of Q̄p can be embedded in L. (This could also be proved by noting
that L is complete by Corollary 4.)
We will always use as the set S of representatives for F̄p the primitive
th
k roots of 1, for all k not divisible by p, and 0. Then the elements of
L have the form g αg pg for some primitive k th roots αg of 1, where the
P

exponents form a well-ordered subset of Q. In particular, the elements

of Q̄p can be expressed in this form. This was first discovered by
Lampert [9].

Example: (similar to those in [9]) Let p be an odd prime. The pth

roots of 1 − p in Q̄p have the expansion
2 2 +1/p3
1 − p1/p + p1/p+1/p − p1/p+1/p + ···
1/(p−1)
+ζp + (higher order terms),
where ζ is any one of the p solutions to ζ p = −ζ in Q̄p .

Proposition 10. The fields L and Qp have cardinality 2ℵ0 (and hence
so do all intermediate fields).
Proof. Each series in L defines a distinct function Q → F̄p by sending
q to the residue class of the coefficient of pq . The number of such
functions is ℵ0 ℵ0 = 2ℵ0 , so |L| ≤ 2ℵ0 . On the other hand, as is well
known, |Qp | = 2ℵ0 already, so the result follows. 
Since L and Cp are both complete algebraically closed fields of car-
dinality 2ℵ0 , it is natural to ask if L = Cp . That L strictly contains
16 BJORN POONEN

Cp follows from Lampert’s remark that the support of the series of

an element of Q̄p is contained in N1 Z[1/p] for some N , and that the
residue classes of the coefficients in the series lie in Fq for some q. (For
example, p−1 + p−1/2 + p−1/3 + · · · is an element of L which cannot be
approached by elements of Q̄p .) In fact, we can show that the set of
series with these properties forms an algebraically closed field, using
the following lemma, which is of interest in its own right, and which
we can apply also toward the computation of the algebraic closure of
Laurent series fields.
Lemma 5. Suppose E is an algebraically closed field, and S ⊆ Aut(E).
Let F be the set of elements e ∈ E whose orbit { σ(e) | σ ∈ S } under
S is finite. Then F is an algebraically closed subfield of E.
Proof. Let Orb(x) denote the orbit of x under S. If x, y ∈ F , then
Orb(x + y) ⊆ Orb(x) + Orb(y) which is finite, so x + y ∈ F . Similar
considerations complete the proof that F is a subfield.
Given p(x) ∈ F [x], let c be a zero of p in E. Then the orbit of p(x)
under S is finite (since each coefficient has finite orbit), and Orb(c)
consists of zeros of polynomials in the orbit of p(x) (to be specific, σ(c)
is a zero of σ(p)), so c ∈ F . Hence F is algebraically closed. 
The characteristic p case of the following corollary was proved by
Rayner [16] using a different method.
Corollary 7. If k is an algebraically closed field of characteristic 0,
¯ = S∞ k((t1/n )). If k is an algebraically closed field of
then k((t)) n=1
characteristic p, then the set of series in k((Q)) with support in N1 Z[1/p]
for some N (depending on the series) is an algebraically closed field
containing k((t)).
Proof. If ζ is a homomorphism from Q/Z to the group of all roots
of unity in k, then we get an automorphismPof the algebraically closed
Mal’cev-Neumann ring k((Q)) by mapping q∈Q αq t to q∈Q ζ(q)αq tq .
q P

Let E = k((Q)) and let S be the set of all such automorphisms. Then
the lemma shows that the set F of elements of E with finite orbit under
S is an algebraically closed field. If char k = 0, F = ∞ 1/n
S
n=1 k((t )),
and the desired result follows easily. If char k = p, F is the set of series
in k((Q)) with support in N1 Z[1/p] for some N (since ζ is necessarily
trivial on Z[1/p]/Z). 
Corollary 8. The set of series in L with support in N1 Z[1/p] for some
N such that the residue classes of the coefficients lie in Fq for some q
forms an algebraically closed field which contains Qp , hence also Q̄p .
 MAXIMALLY COMPLETE FIELDS 17

Proof. If µ denotes the group of all k th roots of 1 for all k relatively

prime to p, and ζ : Q/Z → µ is any group homomorphism, then we
get an automorphism of A((Q)) (using the notation of Section 4) by
sending g∈Q αg tg to g∈Q ζ(g)αg tg . This maps the ideal N into itself,
P P

so it induces an automorphism of L. We also get automorphisms of L

coming functorially from the automorphisms of F̄p .
Let E = L, and let S be the set of both types of automorphisms.
Then the elements of L with finite orbit under the first type of au-
tomorphisms are those with support in N1 Z[1/p] for some N , and the
elements with finite orbit under the second type of automorphisms are
those such that the residue classes of the coefficients lie in Fq for some q.
Hence the result follows from the lemma. (Obviously this field contains
Qp .) 

There are many automorphisms of L besides those used in the pre-

vious proof. In fact, L has an enormous number of continuous auto-
morphisms even over Cp .
Proposition 11. Given µ ∈ L \ Cp , let r = supe∈Cp v(µ − e) ∈ R.
Then for any µ0 ∈ L such that v(µ − µ0 ) ≥ r, there exists a continuous
automorphism of L over Cp taking µ to µ0 .
Proof. We will extend the inclusion Cp ,→ L to an embedding Cp (µ) →
L using the proof of Theorem 2 (instead of taking the obvious inclu-
sion). There is no best approximation to µ in Cp , since given any
approximation, we can find a better one by subtracting the leading
term of the series of the difference. So we are in Case 2 of the proof of
Theorem 2, and it follows that we may embed Cp (µ) in L by sending
µ to any solution µ0 ∈ L of the inequalities v(x − eσ ) ≥ gσ , where eσ
ranges over all elements of Cp and gσ = v(µ − eσ ). These are satisfied if
v(µ−µ0 ) ≥ r, by the triangle inequality. Finally, extend this embedding
Cp (µ) → L to a continuous endomorphism L → L using Theorem 2.
This endomorphism is an automorphism by Proposition 7. 

Lampert proved that Cp has transcendence degree 2ℵ0 over the com-
pletion Cunram
p of the maximal unramified extension Qunram
p of Qp , and
that Cpunram ℵ0
has transcendence degree 2 over Qp . We now extend this
chain of results by calculating the transcendence degree of L over Cp ,
using the following generalization of a proposition of Lampert’s.
Proposition 12. If V is a sub-Q-vector space of R containing Q, then
the set of elements in L of which all the accumulation values of the
exponents are in V form a complete algebraically closed field.
18 BJORN POONEN

Proof. The proof is exactly the same as Lampert’s proof for the special
case V = Q [9]. 
Corollary 9. L has transcendence degree 2ℵ0 over Cp .
Proof. Let B be a basis for R as a vector space over Q, with 1 ∈ B. For
each b ∈ B, b 6= 1, pick a strictly increasing sequence q1 , q2 , . . . in Q
with limit b, and define zb = pq1 + pq2 + · · · ∈ L. Let Kb be the field of
Proposition 12 with V the Q-vector space generated by all elements of
B except b. Then Kb contains Cp , since it contains Qp and is complete
and algebraically closed. If c ∈ B, zc ∈ Kb iff c 6= b. But each Kb
is algebraically closed, so no zb can be algebraically dependent on the
others over Cp . Thus the transcendence degree of L over Cp exceeds
the dimension of R over Q (it does not matter that we threw away one
basis element), which is 2ℵ0 . On the other hand the cardinality of L
ℵ0
is at most 2 , since the collection of all series g∈Q αg pg with αg ∈ S
P

has cardinality ℵ0 ℵ0 = 2ℵ0 . So the transcendence degree must equal

2ℵ0 . 
Traditionally, P
p-adic analysis has been done in Cp . But every power
series F (X) = ∞ n
n=0 an X with an ∈ Cp can be defined on L, and
the radius of convergence is the same in L as in Cp , because in either
field the series converges iff the valuation of its terms approach +∞.
(Remember that L is complete.) As an example, we state the following
proposition.
Proposition 13. There exists a unique function logp : L∗ → L such
that
(1) logp x = ∞ n+1
(x − 1)n /n, for v(x − 1) > 0.
P
n=1 (−1)
(2) logp xy = logp x + logp y, for all x, y ∈ L∗ .
(3) logp p = 0.
Proof. The proof for L is exactly the same as the proof for Cp . See
pp. 87–88 in [7]. 
Although we can extend any power series defined on Cp to L, it seems
that p-adic analysis rarely (if ever) would need to use properties of L
not true of Cp . All that seems important is that the field is a complete
algebraically closed immediate extension of Q̄p . It would be interesting
to investigate whether anything can be gained by doing p-adic analysis
in L instead of in Cp .
References
[1] Abhyankar, S. S., Two notes on formal power series, Proc. Amer. Soc. 7,
no. 5 (1956), 903–905.
 MAXIMALLY COMPLETE FIELDS 19

[2] Bourbaki, N., “Commutative Algebra,” Hermann, 1972.

[3] Chevalley, C., “Introduction to the theory of algebraic functions of one
variable,” Amer. Math. Soc. Surveys, 1951.
[4] Fuchs, L. and L. Sache, “Modules Over Valuation Domains,” Marcel
Dekker, 1985.
[5] Kaplansky, I., Maximal Fields with Valuations, Duke Math. J. 9 (1942),
313–321.
[6] —, Maximal Fields with Valuations II, Duke Math. J. 12 (1945), 243–248.
[7] Koblitz, N., “p-adic Numbers, p-adic Analysis, and Zeta-Functions,”
Springer-Verlag, 1984.
[8] Krull, W., Allgemeine Bewertungstheorie, J. für Mathematik 167 (1932),
160–196.
[9] Lampert, D., Algebraic p-adic expansions, J. of Number Theory 23 (1986),
279–284.
[10] MacLane, S., The Universality of Formal Power Series Fields, Bull. Amer.
Math. Soc. 45 (1939), 888–890.
[11] Mal’cev, A., On the Embedding of Group Algebras in Division Algebras,
Dokl. Akad. Nauk. SSSR 60 (1948), 1499–1501.
[12] Neumann, B., On Ordered Division Rings, Trans. of the A.M.S. 66 (1949),
202–252.
[13] Passman, D., “The Algebraic Structure of Group Rings,” John Wiley and
Sons, 1977, 598–602.
[14] Poonen, B., “Formal Power Series and Maximally Complete Fields,” under-
graduate thesis, Harvard University, 1989.
[15] —, Units in Mal’cev-Neumann Rings, unpublished manuscript, 1990.
[16] Rayner, F., An Algebraically Closed Field, Glasgow J. Math. 9 (1968), 146–
151.
[17] Serre, J.-P., “Local Fields,” Springer-Verlag, 1979.

Department of Mathematics, University of California at Berkeley,

Berkeley, CA 94720
E-mail address: [email protected]