GALOIS THEORY AFTER GALOIS

TAMAS SZAMUELY

The editor of Lettera Matematica Pristem has kindly asked me to

write an informal survey of some modern aspects of Galois theory, in-
cluding innite Galois theory and Grothendiecks approach. Here is
my attempt at fullling his request. Of course, since I shall be dealing
with advanced subjects, Ill have to assume that readers are familiar
with some concepts in higher mathematics, including basic Galois the-
ory as presented in standard algebra textbooks. Later sections will also
require some familiarity with the fundamentals of point set topology
and complex analysis.
But let us begin at the beginning.

1. What is a Galois extension?

In the latter half of the 19th century eminent mathematicians such
as Liouville, Jordan, Dedekind and Weber devoted considerable eort
to understanding and developing the seminal work of Galois. During
this work of clarication the focus of Galois theory gradually shifted
from equations to eld extensions until it arrived at the form we use
nowadays. In the course of this process the concept of a Galois exten-
sion of elds was redened several times. Let us list four equivalent
formulations, in increasing order of abstraction.
The denition which is closest to the original approach of Galois is:
(1) A Galois extension is a eld extension L|K, where L is obtained
by adjoining to K all roots of an irreducible separable polynomial with
coecients in K.
Recall that an irreducible polynomial is separable if it has no multi-
ple roots. Correspondingly, a eld extension is called separable if the
minimal polynomials of all elements are separable.
As a next step, it was noticed that the above dening property can
be made independent of the choice of a polynomial.
(2) A Galois extension is a nite separable eld extension L|K having
the following property: if an irreducible polynomial with coecients in
K has a root in L, then all of its roots lie in L.
Here the emphasis is already on elds but polynomials still lurk
behind. A further step towards abstraction became possible when
Dedekind dened the Galois group as the automorphism group of a
Date: February 6, 2012.
1
2 TAMAS SZAMUELY

Galois eld extension. In previous approaches it was always dened as

a group permuting the roots of a dening polynomial.
(3) A Galois extension is a nite separable eld extension L|K where
every element of L not lying in K is moved by a eld automorphism of
L xing K.
The link with the previous denition is as follows: an automorphism
of L xing K must take an element L to another root of its minimal
polynomial over K. If is not in K, there is at least one other such
root.
The last step was taken by Emil Artin in the 1940s. It is the most
elegant denition of a nite Galois extension.
(4) A Galois extension is a eld extension L|K, where K arises as the
xed eld of a nite group G acting on L via eld automorphisms.
To relate this to the previous denition, one has to prove that a
eld extension as above is automatically separable. Afterwards the
dening property follows. The group G is called the Galois group of
the extension L|K.
It was also Artin who formulated the Galois correspondence as we
know it today. Let us recall the statement:
Let L|K be a nite Galois extension with Galois group G. There is a
one-to-one correspondence between extensions M |K contained in L and
subgroups H G. Galois extensions correspond to normal subgroups.
In this case the Galois group of M over K is isomorphic to G/H.
The correspondence is given by mapping M to the subgroup of G
whose elements x M elementwise, and conversely, by mapping a sub-
group H G to the subeld of L it xes.
This statement is purely in the spirit of the great German algebra
school of the early 20th century. It is still the one we teach and apply
today. There was, however, one further step to be taken: to relax the
condition of niteness.

2. The absolute Galois group

There have been several motivations for studying innite algebraic
extensions. One came from the investigation of special classes of ex-
tensions in algebraic number theory. For instance, cyclotomic elds,
which are obtained by adjoining to the eld Q of rational numbers
some root of unity , have always played a prominent role since the
discovery of their relevance to Fermats last theorem. Any n-th root
of is again a root of unity, so cyclotomic elds give naturally rise to
towers of eld extensions. For instance, one may x a prime number
p, and consider the tower Q(p ) Q(p2 ) Q(p3 ) of elds
obtained by adjoining p-th, p2 -th, p3 -th roots of unity, and so on. Their
union is a very interesting innite algebraic extension whose study gave
 GALOIS THEORY AFTER GALOIS 3

rise to one of the most important branches of present-day arithmetic,

Iwasawa theory.
Another motivation comes from a purely algebraic question: how to
nd elds over which every polynomial equation has a solution? That
this question is not purely academic goes back to a famous theorem of
Gauss, nowadays known as the fundamental theorem of algebra:
Over the eld C of complex numbers every polynomial equation f = 0
has a root C.
Moreover, all roots of the equation must lie in C, as one sees by
dividing f by factors of the form (x ).
But how to nd other examples? There is a natural idea: start
with a eld K and try to nd a means of adjoining to K all solutions
of polynomial equations with coecients in K. The resulting eld K
will then have the required property. Indeed, given a polynomial f
with coecients in K, its nitely many coecients lie in some nite
extension L of K by construction of K. But then a root of f lies in a
nite extension of L which is again a nite extension of K. As such, it
is generated by roots of polynomials with coecients in K, so must
lie in K.
For K = Q the construction of K is easy. It suces to view Q
as a subeld of C and take all complex numbers which are roots of
some polynomial with coecients in Q. One shows that these numbers
form a eld, the eld of algebraic numbers. Since we know by the
fundamental theorem of algebra that the roots of every polynomial
with coecients in Q lie in C, we have solved our problem.
For those elds that cannot be embedded in C (for instance those
of positive characteristic or transcendental extensions of C) the above
method breaks down. The general solution was found in a fundamental
paper of Steinitz [6] which exploited the then fairly new techniques of
set theory, in particular the axiom of choice. He showed:
Every eld K has an algebraic closure K which is an algebraic exten-
sion of K such that every polynomial equation with coecients in K has
a root in K. Moreover, K is unique up to a non-unique isomorphism.
It is this non-unicity which is crucial for developing innite Galois
theory. The fact that K is algebraic over K means that it is generated
by roots of polynomials with coecients in K. We thus see that K
indeed arises by adding all roots of such polynomials in a systematic
way.
Once we have an algebraic closure K at hand, we may dene a sep-
arable closure Ks of K as being those elements in K whose minimal
polynomial over K is separable. It can be checked that these elements
form a eld over which every separable polynomial has a root. Con-
sider now the group of eld automorphisms of Ks xing K. It is the
4 TAMAS SZAMUELY

absolute Galois group of K; we denote it by . It depends on the choice

of Ks but its isomorphism class does not.
The absolute Galois group has several important features. Firstly,
it has the property of denition (3) above: any element Ks not in
K is moved by some element of . The proof of this fact uses another
nontrivial theorem of Steinitz: if we map to another root of its

minimal polynomial, the resulting isomorphism K() K( ) can be
extended to an automorphism of Ks . Another important point is that
Ks is the union of all nite Galois extensions of K contained in K;
this is because one may embed any nite separable extension in a nite
Galois extension. Furthermore, the consideration of gives us a means
for giving yet another denition of a nite Galois extension:
(5) A nite separable extension L|K contained in Ks is Galois if (L) L
for all .
That L should be contained in Ks is not a serious restriction, as
any L separable over K can be embedded in Ks . We thus get another
insight into the basics of Galois theory: one can decide whether an
extension is Galois by using the absolute Galois group which is always
there. The next step is to use it to describe all subelds of Ks .

3. Infinite Galois Theory

The discussion of the previous two sections points to several equiva-
lent ways of dening a possibly innite Galois extension. We mention
two of them. The rst is a variant of denition (3): a Galois extension
is a separable algebraic extension L|K such that every element of L
not lying in K is moved by a eld automorphism of L xing K. The
second one is motivated by denition (5): an extension L|K contained
in Ks is Galois if (L) L for all in the absolute Galois group .
The most interesting innite Galois extension of K is of course the
separable closure Ks .
Whichever denition we adopt, there is only one way to dene the
Galois group Gal(L|K): it is the group of eld automorphisms of L
xing K. However, it is the second denition that gives us an easy
key to a fundamental property of the Galois group. Namely, given a
Galois extension M |K contained in L, one has a natural group homo-
morphism Gal(L|K) Gal(M |K) given by restricting automorphisms
of L to M . This homomorphism is moreover surjective by the theo-
rem of Steinitz already mentioned in the previous section: one may
extend any automorphism of M over K to an automorphism of Ks
which must then preserve L by denition. Thus Gal(M |K) arises as
a quotient of Gal(L|K); in particular this applies to all nite Galois
extensions contained in L.
The basic fact is now that Gal(L|K) is completely determined by
its nite quotients. This is not so surprising if we recall from the
 GALOIS THEORY AFTER GALOIS 5

previous section that L is the union of all the nite Galois extensions it
contains. However, giving a precise formulation requires a sophisticated
algebraic tool called the inverse limit: one says that Gal(L|K) is the
inverse limit of its nite quotients. A group that arises as an inverse
limit of nite groups is called a pronite group. We shall not give the
detailed denition of inverse limits and pronite groups here. What is
important to bear in mind is that pronite groups are determined by
their nite quotients.
Pronite groups carry an important additional structure: they are
topological groups. It is possible to introduce this topology by a gen-
eral method that starts by putting the discrete topology on nite
quotients. However, in the case of Galois groups there is a more di-
rect approach introduced by Krull in his groundbreaking paper [4] on
innite Galois theory. Consider the kernels of the homomorphisms
Gal(L|K) Gal(M |K) considered above for all nite Galois exten-
sions M |K contained in L, and declare these to be a system of open
neighbourhoods of the identity in Gal(L|K). For a general element
Gal(L|K) a system of open neighborhoods is given by the cosets
U , where U is an open neighbourhood of the identity. One checks that
these open sets form the basis of a topology on Gal(L|K). In honour
of its father it is called the Krull topology. It can be shown that with
the Krull topology Gal(L|K) becomes a compact Hausdor topological
group.
Any open subgroup in Gal(L|K) is also closed because its comple-
ment is a union of cosets which must be open as well. Hence any,
possibly innite, intersection of open subgroups is a closed subgroup.
On the other hand, it is not hard to show that every closed subgroup
is the intersection of the open subgroups containing it. This gives a
hint at how one should develop the Galois correspondence for innite
extensions: since open normal subgroups correspond to nite Galois
extensions, closed subgroups should correspond to arbitrary Galois ex-
tensions, again because the latter are unions of nite Galois extensions.
Indeed, Krull proved the following generalized Galois correspondence:

Let L|K be a Galois extension with Galois group G. There is a one-to-

one correspondence between extensions M |K contained in L and closed
subgroups H G. Finite extensions correspond to open subgroups and
Galois extensions to closed normal subgroups. In the latter case the
Galois group of M over K is isomorphic to G/H.

At this point the natural question arises whether there exist non-
closed subgroups in the Galois group or, in other words, whether there
exist subgroups that do not arise as the subgroup of elements of G
xing some eld extension. In fact, this question had been solved by
Dedekind [2] well before Krull set up his theory. He argued as follows:
6 TAMAS SZAMUELY

If L1 L2 L2 . . . is a strictly increasing chain of nite Galois ex-

tensions of K, then each automorphism in Gal(Li |K) can be extended
in at least two dierent ways to Li+1 by classical Galois theory. An
innite chain thus gives rise to an uncountable Galois group which has
uncountably many subgroups. But if the base eld K is countable,
e.g. K = Q, then there are only countably many nite eld extensions
of K because there are only countably many polynomials with coe-
cients in K. It should be remarked in passing that the aforementioned
work of Dedekind was the main inspiration for Krulls theory. In fact,
Dedekind already had the insight that innite Galois groups should
enjoy some continuity property.
In concrete situations it is easy to exhibit non-closed subgroups. For
instance, if Fq denotes the nite eld of order q, then the Frobenius
automorphism F : x 7 xq of Fq does not generate a closed subgroup.
In fact, it is easy to describe the Galois theory of the innite extension
Fq |Fq without mentioning pronite groups. As we know from the the-
ory of nite elds, for each integer r > 0 there is a unique subextension
Fqr |Fq of Fq |Fq which has degree r over Fq , and moreover it is a Galois
extension of Fq with group Z/rZ. It follows that the open subgroups
of = Gal(Fq |Fq ) are totally ordered by inclusion, and therefore ev-
ery system of open subgroups either has trivial intersection or has a
smallest element. Hence every nontrivial closed subgroup H is in
fact open; moreover, it is normal with /H cyclic. The innite cyclic
subgroup of generated by F is not open: its xed eld is Fq but it
does not equal .

4. Grothendiecks Reformulation
Alexander Grothendieck, whose inuence on mathematics in the lat-
ter half of the 20th century was comparable to that of Galois in the
19th, found a very useful reformulation of the main theorem of Galois
theory which can be generalized to many other settings as well. In
his seminar [3] he gave a general categorical formulation encompassing
several situations. We stick here to the already explored case of eld
extensions. From Grothendiecks viewpoint the aim of Galois theory
is to classify nite separable extensions of a given eld by means of
permutation representations.
To explain his idea, let K be a base eld and L|K a nite separable
extension. Fix a separable closure Ks of K. As we know from the
work of Galois himself, L is generated over K by a single element .
Let f be the minimal polynomial of over K, and 1 , . . . , n the
roots of f in Ks . The absolute Galois group := Gal(Ks |K) acts on
the nite set 1 , . . . , n via permutations: for each the n-tuple
(1 ), . . . , (n ) is just the system of the i listed in a possibly dierent
order. There are two important properties of this action. Firstly, it is
transitive, which means that for given i , j we may nd a with
 GALOIS THEORY AFTER GALOIS 7

(i ) = j . Secondly, it is continuous for the topology of . This

means that for each i the set of those for which (i ) = i is
an open subgroup of .
Grothendiecks reformulation now can be expressed as follows:
There is a one-to-one correspondence between isomorphism classes of
nite separable extensions of K and nite sets S equipped with a con-
tinuous transitive action of .
We have already seen how to associate a nite continuous -set to
a eld extension. To get a map in the reverse direction, one picks
an element i S and considers its stabilizer in , i.e. those
for which (i ) = i . It is an open subgroup in , so by innite
Galois theory it xes a nite separable extension L|K. If we choose the
stabilizer of another element, we arrive at the xed eld of a conjugate
open subgroup, which xes a eld extension isomorphic to L.
In a sense the above formulation of Galois theory is closer to Galois
original approach than Artins Galois correspondence because Galois
also considered the permutation representation of the Galois group on
the roots of the equation this aspect is somewhat obscured if one only
looks at the subgroups/subelds correspondence.
One can relax the condition of transitivity of the -action by con-
sidering not just nite separable extensions of K but also nite direct
products of these. Following Grothendieck they are usually called nite
etale k-algebras.

5. Monodromy Representations
Another situation where equations can be classied by means of
group representations that was already abundantly studied in the 19th
century is the monodromy theory of dierential equations. Consider a
linear dierential equation
(1) y (n) + a1 y (n1) + + an1 y + an y = 0
in the complex plane C, where the coecients ai are complex functions
that are holomorphic except in nitely many points x1 , . . . , xr where
they extend meromorphically. By a basic existence theorem of Cauchy
each point x = xi has an open neighbourhood U not containing any of
the xi where the equation has n local holomorphic solutions y1 , . . . , yn
that are linearly independent over C and moreover every local solution
over U is a linear combination of these. In other words, locally around
x the solutions of the equation form an n-dimensional C-vector space.
The trouble is that when we move the point x the solution space
may not remain the same. To see this, consider the simplest example
where r = 1 and the equation is of the form
y = f y
8 TAMAS SZAMUELY

with a meromorphic function f that is holomorphic outside x1 . As is

well known, all local solutions of the equation are constant multiples
of the function exp F , where F is a primitive of f . However, as
we learn in basic complex analysis, a primitive of F does not exist on
the whole of X \ {x1 }, only in X \ D, where D is a half-line starting
from x1 . For instance, if x1 = 0, then a well-dened primitive F
of f exists over U = C \ [0, i) and another primitive F+ over
U+ = C \ [0, i). The intersection U U+ splits in two connected
components C = {z : Re(z) < 0} and C+ = {z : Re(z) > 0}. As
F+ and F may dier only by a constant on each component, we are
allowed to choose them in such a way that F = F+ on C . However,
they will then dier by a constant c on C+ ! So if we choose a closed
loop around 0, say the circle of radius 1, then in a neighbourhood of 1
we may take exp F+ as a local solution. Moving along the circle away
from 1 we may still use the same local solution for some time around
each point of the circle, but at some point in C we must switch from
F+ to F because F+ is not dened along the whole circle. So when
we arrive back to 1, we get exp F as the continuation of the local
solution, which is ec (exp F+ ). This constant ec is the monodromy of
the equation around 0.
The above procedure can be generalized to the general equation (1).
Choose closed loops i passing through x such that each i has only
the point xi in its interior but not the other xj . For xed i we can do
the same as in the special case discussed above: we start with a local
solution yx of (1) around x and we move along i continuing it to local
solutions around each point. When we return to x we obtain another
local solution zx around x. It is usually not a constant multiple any
more because the equation can be more complicated but it lies in the
same nite-dimensional vector space of local solutions around x.
We can proceed likewise for all i. An apparent drawback is that the
results a priori depend on the choice of the loops i . However, the
classical monodromy theorem of complex analysis says that this is not
the case: if we replace i by another loop i having the same properties,
then the resulting zx will be the same. From the modern point of view
this is because i and i are homotopic: they can be continuously
deformed into each other without touching any of the xi along the way.
This observation is the germ of the notion of the fundamental group:
one may consider all closed loops passing through x and not touching
the xi up to continuous deformation. It is possible to introduce a group
operation on this set by taking as the product of two loops and the
loop obtained by going through and then . (It goes through x twice
but one may deform it so that it is not the case any more. For instance,
one may represent the product i j by a loop around x not meeting
itself and having xi and xj in its interior.) The resulting group is
generated by the classes of the i . Observe now that the operation of
 GALOIS THEORY AFTER GALOIS 9

continuing local solutions of (1) along loops representing elements of

is C-linear. In other words, we obtain an action of on the C-vector
space of solutions of (1) around x. Fixing a basis y1 , . . . , yn of this
space we thus obtain a homomorphism : GL(n, C). It is called
the monodromy representation of (1) around x.
The question now arises: can one classify dierential equations by
their monodromy representations in the same way as permutation rep-
resentations of the Galois group classify nite eld extensions? The
perceptive reader will immediately notice that the question is not cor-
rectly formulated. Giving a eld extension is not the same as giv-
ing a polynomial equation, although one knows that every nite eld
extension does indeed come from the polynomial. The question for
dierential equations must be similarly split in two parts.
The rst part is usually called the RiemannHilbert correspondence.
An elementary form of it says:
There is a one-to one correspondence between isomorphism classes of
monodromy representations : GL(n, C) and local systems, i.e.
systems of holomorphic functions on open subsets of C avoiding the xi
which locally around each point form an n-dimensional C-vector space.
The second part is the RiemannHilbert problem which asks:
Does every local system (and hence every monodromy representation)
come from a linear dierential equation (1)?
The answer is not unambiguous, especially if one requires additional
properties of the equation. Let us cite the most famous classical result,
that of Plemelj [5]:
Every local system as above comes from a linear dierential equation
(1) whose coecients are meromorphic functions in C. Moreover, the
equation can be chosen to be Fuchsian, which means that at each sin-
gular point the the coecient ai has a pole of order at most i.
It is important to observe that though in the above statement the
coecients of the equation have at most poles at the xi , they may have
poles at nitely many other points as well. One may consider a mon-
odromy representation taking these extra points into account as well,
but with the additional condition that the action of group elements
coming from loops around the extra points is trivial. In other words,
continuing a local solution around the extra points does not change
the solution. Such points are classically called apparent singularities.
If one does not allow apparent singularities, the statement does not
hold. For a beautiful introduction to the RiemannHilbert problem
and its generalizations, see Beauvilles report [1].
We have thus indeed arrived at statements that resemble the modern
formulation of Galois theory. Thanks largely to insights of Grothendieck
it is now possible to develop several general theories encompassing both;
10 TAMAS SZAMUELY

see e.g. my book [7] where the concepts surveyed in this article are also
explained in more detail. But the ties between these branches of math-
ematics are stronger. In recent decades dierential equations, and even
more the topological considerations arising from their theory, have been
successfully applied to the construction of interesting Galois extensions
of elds like C(t) and even Q. On the other hand, methods imported
from Galois theory have proven to be fundamental for analyzing dif-
ferential equations. We can thus happily observe that two centuries
after Galois his ideas are not only more alive than ever but have also
invaded a large part of present-day mathematical research.
References
[1] A. Beauville, Monodromie des systemes dierentielles lineaires a poles simples
sur la sphere de Riemann [dapres A. Bolibruch], Seminaire Bourbaki, expose
765, Asterisque 216 (1993), 103119.
[2] R. Dedekind, Uber die Permutationen des Korpers aller algebraischen Zahlen,
Abhandlungen der Gesellschaft der Wissenschaften zu Gottingen, 1901.
[3] A. Grothendieck, Revetements etales et groupe fondamental (SGA 1), Lecture
Notes in Mathematics, vol. 224, Springer-Verlag, Berlin-New York, 1971. New
annotated edition: Societe Mathematique de France, Paris, 2003.
[4] W. Krull, Galoissche Theorie der unendlichen algebraischen Erweiterungen,
Math. Ann. 100 (1928), 687698.
[5] J. Plemelj, Riemannsche Funktionenscharen mit gegebener Monodromie-
gruppe, Monatshefte Math. Phys. 19 (1908), 211246.
[6] E. Steinitz, Algebraische Theorie der Korper, J. reine angew. Math. 137 (1908),
167309.
[7] T. Szamuely, Galois Groups and Fundamental Groups, Cambridge Studies in
Advanced Mathematics, vol. 117, Cambridge University Press, 2009.