Math 676.

Global extensions approximating local extensions

1. Motivation
A very useful tool in number theory is the ability to construct global extensions with specified local
behavior. This problem can arise in several different forms. For example, if F is a global field and v1 , . . . , vn
is a finite set of distinct places of F equipped with finite separable extensions Ki0 /Fvi , and if S is an auxiliary
finite set of non-archimedean places of F , does there exist a finite separable extension F 0 /F unramified at S
such that the completion of F 0 at some place vi0 over vi is Fvi -isomorphic to Ki0 ? Moreover, if all extensions
Ki0 /Fvi have degree d then can we arrange that [F 0 : F ] has degree d? If so, are there infinitely many such
F 0 (up to F -isomorphism)?
An alternative problem is the following: if the extensions Ki0 /Fvi are Galois (and so have solvable Galois
group, which is obvious when vi is archimedean and which we saw in class in the non-archimedean case via
the cyclicity of the Galois theory of the finite fields), then can we arrange for F 0 /F as above to also be
Galois with solvable Galois group? If the extensions Ki0 /Fvi are abelian (resp. cyclic) then can we arrange
for F 0 /F is have abelian (resp. cyclic) Galois group?
A typical example of much importance in practice is this: construct a totally real solvable extension L/Q
that is unramified over a specified finite set of primes S of Q and induces a specified extension of Qp for p in
another disjoint finite set of primes. Here we are prescribing the local behavior of L at a finite set of primes
p and also at the infinite place while insisting that L also be unramified at the auxiliary finite set S. In this
handout we provide an affirmative answer to all such construction problems, though we must note at the
outset that whereas the answer to the first problem (construction of F 0 /F with prescribed local behavior
but no Galois conditions) will only require the weak approximation theorem (as in §9 of the handout on
absolute values), the solution of the problems involving Galois groups will reduce to the abelian case whose
solution requires class field theory (and so we will give suitable references in the book of Artin–Tate).
To indicate the subtle nature of such construction problems, consider the cyclic case of the preceding
questions: if Ki0 /Fvi is cyclic with degree ni for all i, can we find a cyclic extension F 0 /F inducing Ki0 /Fvi
for each i? If such a construction can be done, then the cyclic group Gal(Ki0 /Fvi ) is the decomposition group
at vi in the cyclic group Gal(F 0 /F ), and so [F 0 : F ] must be divisible by the least common multiple n of the
ni ’s. Can one always construct such a cyclic extensions F 0 /F with this minimal possible degree? It turns
out that this is a subtle problem and in very special circumstances it has a negative answer, and then the
best that can be done is [F 0 : F ] = 2n. These special circumstances are exhaustively studied in Chapter X
of Artin–Tate, and they only arise when F is a number field and some of the vi ’s are non-archimedean with
residue characteristic 2. We simply give an explicit counterexample here: F = Q, {vi } = {v1 } = {2}, and
K10 /Q2 the unramified extension of degree 8. There does not exist a cyclic extension F 0 /Q of degree 8 in
which 2 is inert (which is to say that F 0 induces the degree-8 unramified extension of Q2 ). The proof of such
non-existence can be given using class field theory, via the product formula for the norm residue symbol in
cyclic extensions (see the discussion following Theorem 1 in §1, Chapter X in Artin–Tate). It can also be
given in an elementary way as a consequence of the determination of (2|p) for odd primes p; we leave this
latter approach to non-existence as an exercise.

2. Approximation with one local field

Before we attack the problem of finding global extensions that induce given local extensions, we address
an issue that we should have considered some time ago: is every local field K realized as a completion of a
global field F ? The archimedean case is obvious, and in the non-archimedean cases we treat characteristic 0
and positive characteristic separately, as follows. If K has positive characteristic then K ' k((t)) for a finite
field k, and so we may take F = k(t) and v to be the t-adic place of F . If K has characteristic 0 and residue
characteristic p then K is a finite (separable) extension of Qp , and since Q is the p-adic completion of Q we
may simply invoke the following theorem with F = Q and v the p-adic place. (Note that for this application
to p-adic fields we only need the separable aspect of the theorem below, and this aspect is handled near the
end of the proof by a simple self-contained application of Krasner’s lemma.)
1
2

Theorem 2.1. Let F be a global field and let v be a non-trivial non-archimedean place on F . Let K 0 be a
finite extension of the local field K = Fv , say with degree d. There exists a finite extension F 0 /F with degree
d and a place v 0 on F 0 over v such that Fv0 0 is isomorphic to K 0 over the identification Fv = K. If K 0 /K is
separable then F 0 /F must be separable.
If K 0 /K is Galois, then there exists a finite Galois extension F 0 /F with a place v 0 over v and an inter-
mediate field F0 whose completion under v 0 is identified with Fv = K such that Gal(F 0 /F0 ) is identified with
Gal(K 0 /K).
Note in particular that if K 0 /K is abelian then we can realize it as a completion of an abelian extension of
global fields with the same Galois group. However, this comes at a serious cost: we have to replace the initial
F with some unknown extension F0 . In the next section we will give a better solution to a generalization of
the refined problem with Galois groups, avoiding the interference of the auxiliary F0 , by using the full power
of class field theory.
Historically, the ability to realize an abelian extension of local fields as the completion of an abelian
extension of global fields (with the same Galois group, but with weak control on the global base field)
was important in the first approach to local class field theory via global class field theory. This is the
approach followed in Lang’s Algebraic Number Theory, but such an approach forces one to confront questions
concerning dependence of the resulting theory on the choice of “global model” for a given local abelian
extension. The approach to local class field theory via Galois cohomology (as in Serre’s Local fields) is
intrinsic to the local case and so avoids such hassles.
The reader will check that in the separable case, the proof below has nothing to do with global fields and
so the result in the separable case is really a theorem about fraction fields of arbitrary discrete valuation
rings.
Proof. Suppose we have such an F 0 . Let us show first that F 0 /F must be separable with v 0 the unique
place on F 0 over v if K 0 /K is separable, in which case when F 0 /F is Galois then Gal(F 0 /F ) = D(v 0 |v)
obviously maps isomorphically to Gal(K 0 /K). Since Fv0 0 = F 0 Fv = F 0 K, the natural map F 0 ⊗F Fv → Fv0 0
over Fv is surjective. Consideration of Fv -dimensions shows that this must be an isomorphism, and hence if
Fv0 0 /Fv is separable then the discriminant of the Fv -algebra F 0 ⊗F Fv is nonzero. Since the formation of the
discriminant is compatible with extension of the base field, it follows that the discriminant of the F -algebra
F 0 must be nonzero, and from the handout on étale algebras we know that this latter non-vanishing is
equivalent to separability of the field extension F 0 /F . Since F 0 /F is separable, F 0 ⊗F Fv is Fv -isomorphic
to the product of the completions of F 0 at the places over v on F 0 , so since this tensor product is identified
with Fv0 0 we conclude that v 0 is indeed the unique place on F 0 over v.
Now we turn to the existence problem. By working in towers, we may separately treat the cases when
K 0 /K is separable and (in characteristic p > 0) purely inseparable. We first dispose of the purely inseparable
case, say with characteristic p > 0. We may assume K 0 /K has degree p, so K 0 ⊆ K 1/p . Since K ' k((t))
with k perfect (even finite) of characteristic p, it is clear that K 1/p = k((t1/p )) = K(t1/p ) has degree p over
K, so K 0 ' K 1/p over K. In other words, K 0 /K is unique up to isomorphism. Let κ be the constant field of
F and use Riemann–Roch to exhibit F as a finite separable extension of the field F0 = κ(t) such that v lies
over the t-adic place of F0 . Note that t is not a pth power in F (as it is not a pth power in the subfield F0
and F/F0 is separable). Hence, F 0 = F (t1/p ) is a purely inseparable extension of degree p. There is a unique
place v 0 on F 0 over v (why?) and Fv0 0 contains a pth root of t. If we can show that t is not a pth power in
Fv then it follows that the extension Fv0 0 = Fv (t1/p ) is purely inseparable of degree p and so realizes K 0 /K.
Since v lies over the t-adic place on the subextension F0 = κ(t) over which F is finite separable, Fv is finite
separable over the t-adic completion κ((t)) of F0 . Hence, t is a pth power in Fv if and only if it is a pth power
in κ((t)), and so the obvious falsehood of this latter possibility shows that indeed t is not a pth power in Fv ,
as desired. This takes care of inseparability aspects. (Using stronger methods in commutative algebra, it
can be proved that for any global function field F with characteristic p and any non-trivial place v on F , the
non-algebraic extension of fields Fv /F is always separable in the sense of field theory, and so in particular
an element of F that is not a pth power cannot become a pth power in Fv . This is best understood via
Grothendieck’s theory of excellent rings. The point of mentioning it here is to clarify that the preceding ad
 3

hoc trick with Riemann–Roch is not really necessary if we grant ourselves better foundations in commutative
algebra.)
It remains to study the more interesting (and more important) case when K 0 /Fv is a finite separable
extension. By the primitive element theorem, we may write K 0 ' Fv (α) where α is a root of an irreducible
separable monic polynomial h ∈ Fv [T ] with degree d. By Krasner’s lemma, if h1 ∈ Fv [T ] is a monic
polynomial of degree d whose coefficients are sufficiently close to those of h, then h1 is separable and
irreducible with the field Fv [T ]/(h1 ) isomorphic to Fv [T ]/(h) ' Fv (α) over Fv . In other words, slightly
moving h as a monic polynomial of degree d does not affect its usefulness as a way to describe K 0 over
Fv . Hence, by denseness of F in Fv we may assume h ∈ F [T ]. Irreducibility and separability of h over Fv
forces the same over the subfield F , and thus F 0 = F [T ]/(h) is a finite separable extension of F with degree
deg h = d. Moreover, we have an Fv -algebra isomorphism Fv ⊗F F 0 ' Fv [T ]/(h) ' K 0 , yet since F 0 /F is a
0 0
Q
finite separable extension we have Fv ⊗F F ' v0 |v Fv0 . Hence, it follows that v admits a unique extension
to a place v 0 on F 0 , and that Fv0 0 is Fv -isomorphic to K 0 .
Finally, assume that K 0 /K is Galois. The finite separable extension F 0 /F as above is generally not Galois,
so let F 00 /F 0 be a Galois closure of F 0 over F . Since F 00 /F is a splitting field for h, upon choosing a place
v 00 on F 00 over v 0 we see that Fv0000 = F 00 Fv is a splitting field over Fv = K for h. However, the subfield
K 0 = Fv0 0 is such a splitting field since K 0 /K is Galois and is generated by a root of h, so Fv0000 = Fv0 0 . Taking
F0 to be the decomposition field for v 00 over F , and v0 to be the place on F0 under v 00 , the general theory
of the decomposition field shows that v 00 is the unique place on F 00 over v0 and D(v 00 |v) = D(v 00 |v0 ), with
Fv0000 = (F0 )v0 ⊗F0 F 00 , and the resulting natural map
Gal(K 0 /K) = Gal(Fv0000 /Fv ) = Gal(Fv000 /(F0 )v0 ) → Gal(F 00 /F0 ) = D(v 00 |v0 ) = D(v 00 |v)
an isomorphism. Hence, by completing F 00 /F0 at v 00 and v0 we recover K 0 /K without changing the Galois
group. 

3. Approximation with several local fields

Having shown that any local field may be obtained by completing a global field at a suitable place, we
can now turn to our main questions considered at the outset: can we simultaneously approximate extensions
of finitely many completions of a fixed global field F with completions of a single finite extension of F ? In
other words, can we construct global extensions with prescribed local behavior? And how about refinements
concerning the structure of local and global Galois groups? It is this simultaneous local approximation that
will make use of weak approximation, as well as class field theory for the Galois aspects.
Let v1 , . . . , vn be distinct places of F and let Ki0 /Fvi be a finite extension. Can we find a finite extension
F /F and a place vi0 on F 0 over vi such that Fv0 0 ' Ki0 over Fvi for every i, with F 0 /F even separable (resp.
0
i
Galois) when all Ki0 /Fvi are separable (resp. Galois)? This is a refinement of Theorem 2.1 in several respects:
we permit several valuations at once, we allow archimedean places, and we insist on a fixed global base field F .
If Ki0 /Fvi is separable for some i but inseparable for others then F 0 /F cannot be separable, so in such cases the
simultaneous approximation is impossible: it can be proved that non-trivial inseparability in an extension of
global function fields cannot be lost under passage to any completions. Likewise, if the inseparability degrees
of Ki0 /Fvi are not all the same, the construction cannot be done. Hence, it is only reasonable to consider
the cases when all Ki0 /Fvi are separable or all Ki0 /Fvi are inseparable with the same nontrivial degree of
inseparability. Since local fields K of characteristic p > 0 admit exactly one purely inseparable extension of
n
degree pn for each n > 0 (namely, K 1/p ), simultaneous approximation of inseparable local extensions with
a common inseparability degree (which never comes up in practice, as far as I am aware) is easily reduced
to the separable case. It is the separable case that is of fundamental importance, and so only this case will
be considered in what follows.

Theorem 3.1. Let F be a global field and let v1 , . . . , vn be inequivalent places on F . Let Ki = Fvi , and
let Ki0 /Ki be finite separable extensions. There exists a finite separable extension F 0 /F and places vi0 on F 0
over vi such that Fv0 i is Ki -isomorphic to Ki0 for every i, and if [Ki0 : Ki ] = d for all i then we can arrange
4

that [F 0 : F ] = d. If S is a finite set of non-archimedean places of F distinct from the vi ’s then F 0 /F may
be arranged to also be unramified over all places in S.
If moreover all Ki0 /Ki are Galois then we can simultaneously arrange that F 0 /F is a solvable Galois
extension. If in addition all Ki0 /Ki are abelian then F 0 /F can be arranged to be abelian, and likewise for
cyclic extensions.
Of course, in the Galois case it cannot be arranged that [F 0 : F ] = d if [Ki0 : Ki ] = d for all i except for
possibly if all of the Galois groups Gal(Ki0 /Ki ) are isomorphic (as each such decomposition group at vi0 must
fill up the global group Gal(F 0 /F ) for cardinality reasons). This is a very special situation that only seems
to be tractable in the important cyclic case, in which case it has a negative answer in special situations;
see Theorem 5 in §2, Chapter X in Artin–Tate for the full story concerning the degree aspects in the cyclic
case. Our treatment of the Galois assertions in Theorem 3.1 will ultimately reduce to problems in class field
theory that are solved in Chapter X of Artin–Tate.

Proof. Let hi ∈ F [T ] be a monic irreducible separable polynomial such that Ki0 ' Ki [T ]/(hi ) for each i; the
existence of such hi was shown via Kranser’s lemma in the proof of Theorem 2.1 when vi is non-archimedean,
and it is trivial when vi is archimedean (take hi = T if Ki0 = Ki and take hi = T 2 +1 if Ki = R and Ki0 ' C).
Let d = max[Ki0 : Ki ], so by multiplying hi against d − deg hi distinct monic linears T − αij for αij ∈ F
that are not roots of hi , we get monic separable multiples Hi ∈ F [T ] of hi with degree d such that Ki0 /Ki is
generated by a root of Hi , and in fact Ki0 /Ki is a splitting field of Hi when Ki0 /Ki is Galois.
Use weak approximation to find a monic polynomial h ∈ F [T ] that is vi -close to Hi for every i, so in
particular h is separable. Moreover, for some i we have Hi = hi , and hence Hi is irreducible in Fvi [T ] for such
an i. Thus, by using sufficiently close approximation at this place we ensure that h is irreducible over Fvi and
so is irreducible over F . (Strictly speaking, this irreducibility step supposes that the distinguished place vi is
non-archimedean, but the archimedean analogue is trivial: the only non-linear monic irreducible polynomials
in the archimedean case occur for monic quadratic polynomials in R[T ] with negative discriminant, and in
such cases a small perturbation of the lower-degree coefficients does not affect the sign of the discriminant
and so does not affect irreducibility over R.) Let F 0 = F [T ]/(h), so F 0 /F is finite separable with degree d.
For each i, we have
Y
Fv0 i ' Fvi ⊗F F 0 ' Fvi [T ]/(h).
vi0 |vi

Since h and Hi are separable over Fvi , by the theorem on continuity of roots (proved in the non-archimedean
case in the handout on algebraic closedness of completions of algebraic closures in the non-archimedean case,
and provable by elementary local compactness arguments in the archimedean case), it follows that the monic
factorization type of h over Fvi mirrors that of Hi in the sense that each monic irreducible factor of h is
close to a unique monic irreducible factor of Hi with the same degree, with such closeness made as small as
we please by taking h near enough to Hi in Fvi . Hence, Ki0 is a factor field of Fvi [T ]/(h) and thus we can
find a suitable vi0 |vi on F 0 for each i such that Fv0 0 is Fvi -isomorphic to Ki0 . Moreover, if [Ki0 : Ki ] = d for all
i
i then our construction yields [F 0 : F ] = d.
Now let S be a finite set of non-archimedean places away from the vi ’s, and we ask if F 0 /F can be chosen
to be unramified at all places in S. This is very easy: since h ∈ F [T ] is a monic polynomial of degree d that is
merely constrained by approximation conditions on its lower-degree coefficients when viewed in Fvi for each i,
weak approximation permits us to arrange for h to also be simultaneously near any desired monic polynomial
of degree d in Fv [T ] for each v ∈ S. There exists a monic irreducible polynomial hv ∈ OFv [T ] with degree d
and disc(hv ) ∈ OF×v for every v ∈ S, such as a monic lift of a separable irreducible polynomial of degree d
over the finite residue field at each such v, and so we can arrange that h ∈ F [T ] is v-integral and separable
irreducible over Fv with disc(h) ∈ F × a v-adic unit for all v ∈ S. It follows that Fv ⊗F F 0 ' Fv [T ]/(h) is an
unramified extension field of Fv for every v ∈ S, so each v ∈ S is unramified (and even inert) in F 0 .
Assume that Ki0 /Ki is Galois for each i. We seek to make F 0 /F such that F 0 /F is also Galois, with
Galois group that is even solvable (and moreover abelian when all Ki0 /Ki are abelian, and similarly in the
cyclic case). Moreover, we wish to carry out such refined constructions while maintaining unramifiedness
 5

conditions at a fixed finite set of non-archimedean places S disjoint from the vi ’s. By the Galois theory
of local fields (and the trivial archimedean case), the finite Galois extensions Ki0 /Ki are solvable. Suppose
we have solved the case when each Ki0 /Ki is abelian. If the solvable extension K10 /K1 is not abelian then
there exists a non-trivial intermediate field L, and by induction on field-degrees can find a solvable F 0 /F
that “works” for the local extensions L/K1 and Ki0 /Ki for all i > 1. Now we use induction again, with this
F 0 as the base field and with the local extensions K10 /L and the trivial extensions Ki0 /Ki0 for i > 1 (and S
replaced with the set of places over it in F 0 ) to get a further solvable extension F 00 /F 0 . The extension F 00 /F
is finite separable but probably not Galois. However, it is clear (check!) that the Galois closure of F 00 over
F is solvable and does the job.
It remains to consider the case when all Ki0 /Ki are abelian. By Theorem 4 in §2, Chapter X of Artin–Tate,
there exists a finite abelian extension F 0 /F equipped with places vi0 over vi such that Fv0 i is Fvi -isomorphic
to Ki0 for all i and such that F 0 is unramified over each v ∈ S. By Theorem 5 in §2, Chapter X of Artin–Tate,
the same holds in the cyclic case. 
We conclude with two important classes of typical examples, one Galois and one not.
Example 3.2. Let F be global field and let v1 , . . . , vn be a set of distinct places on F . Let Ki0 /Fvi be a
finite separable extension of degree d > 1. We claim that there exist infinitely many non-isomorphic finite
separable extensions F 0 /F with degree d such that vi is inert in F 0 with Fv0 i ' Ki0 over Fvi for all i. By
Theorem 3.1, we can use separable Eisenstein polynomials of degree d at non-archimedean places away from
the vi ’s to make separable extensions F 0 /F of degree d such that each vi is inert in F 0 with Fv0 i ' Ki0
over Fvi for every i and F 0 /F is totally ramified (resp. unramified) at any desired auxiliary finite set of
non-archimedean places of F . Thus, by ramification reasons alone we infer that up to F -isomorphism there
exist infinitely many F 0 /F that satisfy the requirements at the vi ’s because a totally ramified extension of
degree d > 1 cannot be unramified.
Example 3.3. Let F be a totally real number field (such as F = Q), and let {v1 , . . . , vn } be a finite set of
non-archimedean places of F . Let Ki0 /Fvi be a finite Galois extension. We claim that there exists a totally
real solvable extension F 0 /F inducing Ki0 /Fvi as the local extension at vi for each i, that such F 0 /F can be
chosen to be unramified (resp. totally ramified) at any auxiliary finite set of places, and that if all Ki0 /Fvi
are abelian (resp. cyclic) then F 0 /F can be taken to be abelian (resp. cyclic). Of course, as we vary the
auxiliary finite set of ramification conditions, we would be getting necessarily non-isomorphic extensions of
F.
Theorem 3.1 provides an affirmative answer to the preceding claim, as we merely insert the archimedean
(all real) places of F into the collection of vi ’s and at these places we impose the local condition that F 0 /F
induces the extension R/R at each of the real places of F . This ensures that F 0 is a totally real number
field. Of course, we could similarly arrange that F 0 is totally complex.