Representation Theory Rise and Role in Number Theory 0

Download as pdf or txt
Download as pdf or txt
You are on page 1of 27

REPRESENTATION THEORY

ITS RISE AND ITS ROLE IN NUMBER THEORY

ROBERT P. LANGLANDS

Introduction
By representation theory we understand the representation of a group by linear transfor-
mations of a vector space. Initially, the group is finite, as in the researches of Dedekind and
Frobenius, two of the founders of the subject, or a compact Lie group, as in the theory of
invariants and the researches of Hurwitz and Schur, and the vector space finite-dimensional,
so that the group is being represented as a group of matrices. Under the combined influences
of relativity theory and quantum mechanics, in the context of relativistic field theories in the
nineteen-thirties, the Lie group ceased to be compact and the space to be finite-dimensional,
and the study of infinite-dimensional representations of Lie groups began. It never became,
so far as I can judge, of any importance to physics, but it was continued by mathematicians,
with strictly mathematical, even somewhat narrow, goals, and led, largely in the hands of
Harish-Chandra, to a profound and unexpectedly elegant theory that, we now appreciate,
suggests solutions to classical problems in domains, especially the theory of numbers, far
removed from the concerns of Dirac and Wigner, in whose papers the notion first appears.
Physicists continue to offer mathematicians new notions, even within representation theory,
the Virasoro algebra and its representations being one of the most recent, that may be as
fecund. Predictions are out of place, but it is well to remind ourselves that the representation
theory of noncompact Lie groups revealed its force and its true lines only after an enormous
effort, over two decades and by one of the very best mathematical minds of our time, to
establish rigorously and in general the elements of what appeared to be a somewhat peripheral
subject. It is not that mathematicians, like cobblers, should stick to their lasts; but that
humble spot may nevertheless be where the challenges and the rewards lie.
J. Willard Gibbs, too, offered much to the mathematician. An observation about the
convergence of Fourier series is classical and now known as the Gibbs phenomenon. Of
far greater importance, in the last quarter-century the ideas of statistical mechanics have
been put in a form that is readily accessible to mathematicians, who are, as a community,
very slowly becoming aware of the wealth of difficult problems it poses. His connection
with representation theory is more tenuous. In the period between his great researches on
thermodynamics and statistical mechanics, he occupied himself with linear algebra, even in
a somewhat polemical fashion. In particular, he introduced notation that was more than
familiar to mathematics students of my generation, but that perhaps survives today only in
electromagnetic theory and hydrodynamics.
In these two subjects, the dot product α · β and the cross product α × β introduced
by Gibbs are ubiquitous and extremely convenient. The usual orthogonal group comes to
us provided with a natural representation because it is a group of 3 × 3 matrices, and all
objects occurring in elementary physical theories must transform simply under it. The
Appeared in Proceedings of the Gibbs Symposium, AMS (1990).
1
2 ROBERT P. LANGLANDS

two products of Gibbs appear because the tensor product of this natural representation
with itself is reducible and contains as subrepresentations both the trivial representation,
which assigns 1 to every group element, and, if attention is confined to proper rotations, the
natural representation itself. Tensor products play an even more important role, implicitly or
explicitly, in spectroscopy, and it is their explicit use for the addition of angular momenta,
especially by Wigner, that drew representation theory so prominently to the attention of
physicists.
The use of compact groups and their Lie algebras became more and more common, especially
in the classification of elementary particles, often with great success, as a terminology
peppered with arcane terms from occidental literature and oriental philosophy attests, but the
investigation of the representations of groups like the Lorentz group, in which time introduces
a noncompact element, was left by and large to mathematicians. To understand where it led
them, we first review some problems from number theory.
Number Theory
Two of the most immediate and most elementary aspects of number theory that are yet
at the same time the most charged with possibilities are diophantine equations and prime
numbers. Diophantine equations are equations with integral coefficients to which integral
solutions are sought. A simple example is the equation appearing in the pythagorean theorem,
(a) x2 + y 2 = z 2 ,
with the classic solutions,
32 + 42 = 52 , 52 + 122 = 132 ,
that were in my youth still a common tool of carpenters2,3 and surveyors. It is also possible
to study the solutions of equations in fractions and to allow the coefficients of the equations
to be fractions, and that often amounts to the same thing. If a = x/z and b = y/z then the
equation (a) becomes
a2 + b2 = 1.
Prime numbers are, of course, those like 2, 3, 5 that appear in the factorization of other
numbers, 6 = 2 · 3 or 20 = 2 · 2 · 5, but that do not themselves factor. They do not appear
in our lives in the homely way that simple solutions of diophantine equations once did, but
their use in cryptography, in for example public-key cryptosystems, has perhaps made them
more a subject of common parlance than they once were.
The two topics may be combined in the subject of congruences. Consider the equation
x2 + 1 = 0. If p is any prime number we may introduce the congruence,
x2 + 1 ≡ 0 (mod p).
REPRESENTATION THEORY ITS RISE AND ITS ROLE IN NUMBER THEORY 3

A solution of this is an integer x such that x2 + 1 is divisible by p. Thus,


x = 1, 12 + 1 = 2 ≡ 0 (mod 2),
2
x = 2, 2 +1=5≡0 (mod 5),
2
x = 3, 3 + 1 = 10 ≡ 0 (mod 2, mod 5),
x = 4, 42 + 1 = 17 ≡ 0 (mod 17),
x = 5, 52 + 1 = 26 ≡ 0 (mod 2, mod 13),
x = 6, 62 + 1 = 37 ≡ 0 (mod 37).
The list can be continued and the regularity that is already incipient continues with it. The
primes for which the congruence can be solved are 2 and 5, 13, 17, . . . all of which leave the
remainder 1 upon division by 4, whereas primes like 7, 19, 23, . . . that leave the remainder 3
upon division by 4 and that have not yet appeared never do so; the congruence is not solvable
for them.
This regularity, at first blush so simple, is the germ of one of the major branches of the
higher number theory and was the central theme of a development that began with Euler and
Legendre in the eighteenth century, and continued down to our own time, with contributions
by Gauss, Kummer, Hilbert, Takagi, and Artin. Current efforts to extend it will be the theme
of this essay.
We shall be concerned with more complex congruences, involving, for example, two
unknowns and our concern will be strictly mathematical, but I mention in passing that
these more abstruse topics may also impinge on our daily life, even disagreeably, since the
sophisticated theory of congruences modulo a prime is exploited in coding theory and thus in
the transmission of information (and of misinformation, not to speak of unsolicited sounds
and images that are simply unpleasant). It is sometimes also necessary, and even important,
to consider congruences modulo integers that are not prime, not only for strictly theoretical
purposes but also for ends that may be regarded as more practical. The difference between
congruences modulo primes and modulo composite numbers is great enough that it may be
used to very good effect in the testing of numbers for primality, and so can be exploited in
cryptography.
Zeta-functions
The most familiar of the zeta-functions, and the one that has given its name to the others,
is that to which the name of Riemann is attached,

X 1
ζ(s) =
n=1
ns
 1 −1  1 −1  1 −1
= 1− s 1− s 1− s ···
2 3 5
Y 1
=
p
1 − p1s
The equality, due to Euler, is obtained by applying the expansion
1
= 1 + x + x2 + x3 + · · ·
1−x
4 ROBERT P. LANGLANDS

to
1
1 − p1s
to obtain
1 1
1+ s
+ 2s + · · · ,
p p
and then recalling that every positive integer can be written in one and only one way as a
product of prime powers.
The principal use of the Riemann zeta-function is for the study of the distribution of the
primes amongst the integers. Observing that the sum
1 1
1 + s + s + ···
2 3
behaves like the integral Z ∞
1
ds,
1 xs
we conclude correctly that the series defining the zeta-function converges for s > 1. To use it
to any effect in the study of the distribution of primes it is, however, necessary to define and
calculate it for other values of s.
There are manifold ways to do this. The Γ-function, defined by the integral
Z ∞
Γ(s) = e−t ts−1 dt
0
when s > 0, is readily shown upon an integration by parts to satisfy the relation Γ(s) =
(s − 1)Γ(s − 1). It can therefore be defined and its value calculated for any s.
To deal with the zeta-function, introduce two further functions defined by infinite series:

2
X
φ(t) = e−πn t ;
n=1

2
X
θ(t) = 1 + 2φ(t) = e−πn t .
−∞

Calculating the integral,


Z ∞
s
(b) φ(t)t 2 −1 dt,
0
term by term, we obtain s
s
π− 2 Γ ζ(s).
2
As a consequence, to calculate the zeta-function we need only calculate the integral (b) for all
values of s. The integrand behaves badly at t = 0 if s 6 2, so that as it stands the integral
still does not serve our purpose.
We next exploit the elements of Fourier analysis, using a device to which Poisson’s name is
attached. We observe that θ(t) is the value at x = 0 of the periodic function

2
X
e−π(n+x) t .
n=−∞
REPRESENTATION THEORY ITS RISE AND ITS ROLE IN NUMBER THEORY 5

It is easy enough to calculate the Fourier expansion of this function. Doing so, and using the
expansion to calculate its value at x = 0, we discover that
 
1 1
(c) θ(t) = √ θ .
t t
With this relation in hand we return to the integral (b), replace it by two integrals, one from
0 to 1, and one from 1 to ∞. The second is defined for any value of s. In the first, we replace
φ(t) by  
1  −1 
− 12 1
t −1 +t φ
2 .
2 t
The first two terms yield integrals that can be integrated explicitly, the result being
1 1
− ,
s−1 s
which gives the anticipated pole at s = 1. In the last term, we substitute 1/t for t obtaining
an integral from 1 to ∞ that is readily seen to converge for any s.
Since the function φ(t) decreases so rapidly as t → ∞, the two integrals that we have not
calculated explicitly are easily evaluated to any degree of accuracy. This is the main lesson
to be drawn from this technical digression. We have not yet linked diophantine equations
to zeta-functions, but when we do, we shall see that as a result of a circle of conjectures
and theorems the use of zeta-functions offers a way of deciding whether a given diophantine
equation has a solution that can be thousands of times more effective than even sophisticated
searches for it. For this, however, it is necessary to be able to calculate their values at certain
points with precision, although it need not be very great. If our experience in this century is
any guide, this can be done only by showing that the zeta-functions attached to diophantine
equations are equal to others, defined ultimately in terms of representations of noncompact
groups, that are amenable to the same analysis as the Riemann zeta-function.
A classic example, in which the use of infinite-dimensional representations is unnecessary,
is provided by the equation x2 + 1 = 0. We attach to it the function,
1 1 1
L(s) = 1 · 1 · ··· ,
1 + 3s 1 − 5s 1 + 71s
the general factor being
1
,
1 ∓ p1s
according as the congruence
x2 + 1 ≡ 0 (mod p)
does or does not have a solution. This is a series defined by a diophantine equation. The
notation L(s) is due to Dirichlet and one often speaks of L-function rather than zeta-function,
following conventions that are unimportant here.
On the other hand, by the regularity whose importance we have already been at pains to
stress, the general factor could also be defined by the alternative that p leave the remainder
+1 or the remainder −1 upon division by 4. A Dirichlet character χ is a multiplicative
function on the integers, thus a function satisfying χ(ab) = χ(a)χ(b), and such that χ(a)
depends only on the remainder after division of a by some positive integer n. For simplicity,
6 ROBERT P. LANGLANDS

one customarily takes χ(a) = 0 if a and n have a divisor in common. For example a Dirichlet
character modulo 4 is given by:
χ(1) = 1; χ(2) = 0; χ(3) = −1; χ(4) = 0.
The general factor of the product defining L(s) is then
 −1
χ(p)
1− s .
p
Expanding the product defining L(s) just as we expanded the product for the zeta-function
we see that
1 1 1 1
L(s) = 1 − s + s − s + s − + · · ·
3 5 7 9

X χ(n)
= s
.
n=0
n
Since the numerator of this series is a periodic function of n we can once again use elementary
Fourier analysis to define and calculate L(s) for any s.
Because it is so important for the subsequent discussion, I repeat that the function L(s) is
originally defined by a product of factors determined according to a diophantine alternative,
simple though it be. At first glance there is no reason to think that the function has a
meaning outside the region, s > 1, where the product obviously converges, but thanks to the
regularity that we have observed, the product can be converted to a series that is defined by
a periodic function and that can be put in a form that has a meaning for any s.
We can not expect the regularity always to be so simple; for then it would not have been
so elusive. Two basic classes of examples serve as an introduction to the general problem;
both illustrate its difficulty. The first, equations with icosahedral Galois groups, has a greater
historical appeal since they are the simplest equations completely inaccessible to the classical
theory of equations with abelian Galois groups; the solution that appears to be correct
for them was suggested directly by ideas originating in representation theory. The second,
equations in two variables defining elliptic curves, illustrates, however, far more cogently to a
nonspecialist the value of L-functions for diophantine equations and is one of the earliest and
most striking uses of the computer as an aide to pure mathematics; here the solution that
appears to be correct was suggested earlier and on different grounds. Both solutions turn out
to be part of the same larger pattern, and although the evidence for them is extensive and
overwhelming, neither has been established in any generality. We begin with equations in a
single variable.
Galois groups
In our discussion of the congruence x2 + 1 ≡ 0 (mod p), we emphasized the search for
solutions, but this is tantamount to the search for factorizations,
x2 + 1 ≡ x2 + 2x + 1 = (x + 1)2 (mod 2),
2 2
x + 1 = x − 4 + 5 ≡ (x − 2)(x + 2) (mod 5).
The polynomial x2 + 1 cannot be factored modulo 7 or 11 or any prime that leaves the
remainder 3 upon division by 4, but factors into two linear factors modulo any prime leaving
the remainder 1.
REPRESENTATION THEORY ITS RISE AND ITS ROLE IN NUMBER THEORY 7

Another example, the reasons for whose choice will be explained later, is
(d) x5 + 10x3 − 10x2 + 35x − 18.
It is irreducible modulo p for p = 7, 13, 19, 29, 43, 47, 59, . . . and factors into linear factors
modulo p for p = 2063, 2213, 2953, 3631, . . . . These lists can be continued indefinitely, but it
is doubtful that even the most perspicacious and experienced mathematician would detect
any regularity. It is none the less there.
To explain why we have chosen this equation and not another as an example, we first
consider the general equation in one variable,
xn + an−1 xn−1 + an−2 xn−2 + · · · + a0 = 0,
in which all coefficients are rational numbers. This equation will have roots θ1 , . . . , θn and
there will be various relations between these roots with coefficients that are also rational,
F (θ1 , . . . , θn ) = 0.
√ √
For example, the roots of x3 − 1 = 0 are θ1 = 1, θ2 = (−1 + −3)/2, θ3 = (−1 − −3)/2
and two of the many valid relations are:
θ1 = 1; θ2 θ3 = θ1 .
We can associate to the equation the group of all permutations of its roots that preserve all
valid relations. In the example the sole possibility in addition to the trivial permutation is
the permutation that fixes θ1 and interchanges θ2 and θ3 . This group is extremely important,
and is known as the Galois group of the equation. We denote it by G.
For simplicity (it is easy to achieve) suppose that the equation has no multiple roots and
that the coefficients are integers, and introduce the discriminant,
Y
∆= (θi − θj ).
i6=j

It is an integer and it is fundamental to the theory of diophantine equations that, following


Dedekind and Frobenius, we can attach to any prime p that does not divide ∆ an element Fp
in G that determines among other things how the equation factors modulo p. More precisely,
it is the conjugacy class of Fp within the group G that is determined, and that suggests, as is
indeed the case, that to define Fp a little theory is necessary.
It also suggests the use not of Fp itself but of the trace of the matrix ρ(Fp ), where ρ is
some (finite-dimensional) representation of the group G, for the numbers trace(ρ(Fp )) taken
for all ρ determine the class of Fp . We can also, fixing ρ, consider the matrices ρ(Fp ) or,
better, their conjugacy classes.
Rather than asking how the factorization of the original equation f (x) = 0 varies with p
and whether it manifests any regularity, we can ask whether, for a given ρ, the conjugacy
classes ρ(Fp ) do. The simplest possibility is that ρ is one-dimensional, and this is the classical
theory; ρ(Fp ) is then a number that is determined by the remainder left upon division of p
by a certain integer n that depends on the equation and on ρ.
The next possibility is that ρ is a representation by 2 × 2 matrices that we may suppose
unitary. Recalling the close relation between the group of unitary matrices in two variables
and the group of proper rotations in three variables, the second being a homomorphic image
of the first, we classify finite subgroups of the unitary group by their image in the group of
proper rotations. Taking the finite subgroup to be ρ(G) and excluding the possibility that
ρ is reducible, we obtain dihedral, tetrahedral, octahedral, and icosahedral representations.
8 ROBERT P. LANGLANDS

Dihedral representations can be treated by the classical theory; tetrahedral and octahedral
representations require modern ideas but have been completely dealt with [L2]. Icosahedral
equations, however, remain intractable in general, and even for specific examples verifying
numerically the validity of the regularity suggested is a task that requires great skill and
ingenuity and is by no means assured of success.
The first step is to find examples of equations with icosahedral Galois groups, and that
requires a computer search of equations of degree 5. Two possibilities that present themselves
are:
x5 + 10x3 − 10x2 + 35x − 18 = 0 ;
x5 + 6x3 − 12x2 + 5x − 4 = 0 .
It is the first that has been studied closely and not the second, although it has smaller
coefficients, so that one might believe calculations with it would be more efficient. A peculiar
aspect, however, of the theory of diophantine equations is that a great deal of theory is
required in order to recognize which are simpler. Neither the size of the coefficients nor the
form of the equation is a guide.
For equations in one variable there are two criteria: the Galois group and the conductor. We
have already chosen the Galois group as simple as possible if the equation is to offer difficulties.
The conductor is a positive integer related to the discriminant but more complicated. To
calculate it requires a painstaking examination of the properties of the equation modulo
powers, often quite high, of the primes that divide the discriminant, but with enough time
and effort that can always be done. The conductor of the first of our two equations is 800
and that of the second 4256. This difference in size entails a great reduction in the number
of calculations, and not being an expert I am not even sure whether the second equation is
accessible to numerical investigations.
A good-sized monograph [Bu] was found necessary to explain the methods used to establish
the proposed regularity for the first. This regularity, which is expressed in terms of modular
forms, has yet to be described; we have first to establish a conviction that it is necessary,
for it is of a transcendental nature, and simply asserts the equality of two quite differently
defined sequences. The criterion that the statement of regularity has to fulfill in order to be
accepted as significant is analytic. For equations in a single variable, its importance is obscure
except in a theoretical context like that of Emil Artin’s papers of the nineteen-twenties, in
which the representation theory of finite groups was fused with the notion of an L-series. For
elliptic curves the importance will be clearer .
It is curious that, although one of the two currents that merged in Artin’s papers, repre-
sentation theory, owed its very existence to Richard Dedekind and F. G. Frobenius [Ha], who
also contributed essential ideas to the study of L-series and the theory of equations, it is not
clear to what extent the rise of representation theory was a response to problems posed by
L-series. The correspondence between Dedekind and Frobenius is extant, and the letters of
Dedekind have been published [D], but it would be imprudent for a reader inexperienced in
the ways of historical research to read too much into them.
If G is the icosahedral group attached to the equation (d) and ρ its two-dimensional
representation, then for all primes p that do not divide the conductor we can form the 2 × 2
REPRESENTATION THEORY ITS RISE AND ITS ROLE IN NUMBER THEORY 9

matrix ρ(Fp ), which has two eigenvalues, αp and βp . The L-function of Artin is then
Y 1 1
L(s) = αp · βp
,
p
1 − ps 1 − sp

the primes dividing the conductor being omitted from the product. This product converges
for s > 1 and one of the simplest cases of the problem posed by Artin in 1923 is to show that
as an analytic function of a complex variable it can be defined for all s. For equation (d),
this was first achieved in the monograph [Bu] in 1970. It was a critical test of the general
ideas formulated in [L1].
Elliptic Curves
As examples of elliptic curves we take the equation in two variables,
(e) y 2 = x3 + Dx,
borrowing from the account in a lecture of G. Harder before the Rheinische-Westfälische
Akademie der Wissenschaften [H]. Take D = (6577)2 . The equation has the solution x = 0,
y = 0. Does it have further solutions with x and y both rational? One way to attempt to
answer this question is to conduct a computer search.
An intelligent search entails an appeal to the theory of the equation, in the hope of replacing
x and y by numbers with possibly smaller numerators and denominators. The first step is to
set
y2 y1 (x21 + 4D)
x = 12 ; y= .
4x1 8x21
Then we put
x1 = lt2 ; y1 = lst,
and finally
U
t= ; s = lW (2V )2 ; l = 6577.
2V
If U, V and W are integers satisfying the equation
U 4 − 64V 4 = 6577W 2 ,
then x and y will be rational solutions of the original equation. Further transformations of a
similar nature but with more complex equations are made, and finally, at some point, a more
or less blind search for integers U, V, W satisfying the new equation begins. That takes time,
apparently some twenty minutes on an IBM 370 if no false starts are made, and ultimately a
triple is found.
U = 10 500 084 257 375 984 596 799
V = 1 980 407 963 453 953 023 564
W = 1 303 262 616 226 128 053 329 966 805 106 514 807 822 601
Of course, there is no need except as a mathematical amusement to find solutions to
the equation (e), but the problem of finding some method more efficient than promiscuous
searching to decide whether it has a solution has an immediate appeal that needs no further
justification. That such a method may exist is, nevertheless, an astonishing discovery due to
Bryan Birch and Peter Swinnerton-Dyer [BSD], one of the earliest uses of the computer in
pure mathematics and perhaps still that with the deepest implications.
10 ROBERT P. LANGLANDS

Their investigations were partly inspired by the study of congruence zeta-functions, which
also originated with Dedekind and Artin but whose evolution was strongly determined by
the geometrical and topological ideas that infused the mathematics of the middle decades of
this century. More than anything else, it is these functions that have shaped current number
theory.
As a simple example we take D = −1 in equation (e) and then for each prime number p,
except 2 and 3, we count the number of solutions modulo p. For example, if p = 5 we simply
list the possibilities for x and y and substitute them in the equation to see if it is satisfied
modulo p.
y\x 0 1 2 3 4
0 + + − − +
1 − − + − −
2 − − − + −
3 − − − + −
4 − − + − −
The plus sign indicates that the congruence is satisfied and the minus sign that it is not.
There are seven plus signs so that the number of solutions of the congruence is N5 = 7. For
any prime p we can count Np in the same way and define αp and βp by the conditions:
Np = p + αp + βp ; βp = ᾱp ; αp βp = p.

These conditions will define αp and βp only if |Np − p| 6 2 p. This is so, and can be proved
by considerations that began with Gauss, but the relation is also a special case of a powerful
general theorem whose statement and proof are one of the great triumphs of the application
to number theory of the topological ideas appearing during this century.
Given the two sequences αp and βp , for any D but now in particular for D = (6577)2 , we
can introduce, in what has become almost a reflex action, the function
Y 1 1
(f) L(s) = αp · βp
,
p
1 − ps 1 − sp

omitting p = 2 and p = 3 from the product. Because the equation (e) has a symmetry
x → −x, y → iy, it satisfies an analogue of the first criterion for simplicity, which was that
the Galois group be commutative. I forego attempting an exact formulation; the upshot is
that there is a regularity that allows this function, like the zeta-function of Riemann and the
L-functions defined by Dirichlet characters, to be calculated in a fraction of a second for any
s and not just in the region s > 32 where the product (f) converges.
Its value at s = 1 is of special interest. According to a special case of the conjecture of
Birch and Swinnerton-Dyer, but not one that has been established, the equation (e) must
have a solution if
L(1) = 0.
This is readily decidable since L(1) should be an integer times
1 ∞
Z
dx
√ .
4 0 3
x + Dx
Judgements as to the value of a given mathematical theorem are diverse and various criteria
may be applied, but one that mathematicians of all casts accept as decisive is that in concrete
REPRESENTATION THEORY ITS RISE AND ITS ROLE IN NUMBER THEORY 11

classical problems it replaces a laborious calculation with uncertain results with one that is
rapid, at least with modern tools, and certain. The conjecture of Birch and Swinnerton-Dyer
achieves this, for it predicts that the existence of nontrivial solutions to certain equations is
decided by the values of zeta-functions, and Yutaka Taniyama had already proposed [Sh]
a method for calculating these values. Taniyama’s proposal was not known, however, to
Birch and Swinnerton-Dyer who tested their conjecture on elliptic equations that like (e) are
sufficiently simple, in the appropriate sense, to be susceptible to a more classical treatment.
Not long after the appearance of the conjectures of Birch and Swinnerton-Dyer the proposal
of Taniyama was taken up by Andre Weil, who made it more precise, and it has since been
subject to a thorough numerical testing, which curiously enough is much more readily
performed for elliptic curves than for icosahedral equations. Once again the apparently
simpler equations are the more difficult of access. The regularity in the numbers Np implied
by Taniyama’s proposal is ultimately of the same nature as that for the traces αp + βp of
the elements Fp in the Galois group of an icosahedral equation, but it has a more immediate
geometric meaning.
A typical form for the equation of an elliptic curve is:
(g) y 2 = x3 + ax2 + bx + c.
Then dx
R
y
is an elliptic integral and that is of course the source of the terminology.
For these equations too, there is a second criterion for simplicity, the conductor that can
be calculated by examining the associated congruence modulo high powers of primes. The
conductor of the equation,
1
y 2 = x3 − x2 + ,
4
is for example 11 and this is the smallest conductor that can be achieved. It is surprisingly
small. Replacing the variable y by y + 21 , we may also write the equation as
(h) y 2 + y = x3 − x2 .
The equation is also surprisingly simple. We change it in two steps. We first set, following
Vélu [V],
1 2 1
X =x+ 2 + + ;
x x − 1 (x − 1)2
 
1 1 1
Y = y − (2y + 1) + +
x3 (x − 1)3 (x − 1)2
If x and y satisfy (g) then the new quantities satisfy the equation,
(i) Y 2 + Y = X 3 − X 2 − 10X − 20,
which, although slightly more complicated, is still elliptic. There is clearly a close relation
between solutions of the two equations. If we then set,
2Y + 1 X −5
σ= , τ =− ,
11 11
we obtain two quantities satisfying the equation,
(j) σ 2 = 1 − 20τ + 56τ 2 − 44τ 3 ,
which ceases altogether to recommend itself by its simplicity. Its advantage is that it appeared
in an altogether different context long ago, in Klein’s lectures on modular forms ([KF], p. 440).
12 ROBERT P. LANGLANDS

As a consequence, although this was understood only much later after the researches of M.
Eichler and G. Shimura in the nineteen-fifties, the L-series of the curve (h), (i), or (j)—they
are all equal—is readily calculated, and provides the simplest example of the method proposed
by Taniyama.
Modular Curves
The set of all complex solutions to the equation (j) forms a surface, because for each
complex value of τ there are in general two possible values of σ. This surface can be obtained
in another way from the upper half-plane of complex numbers z = x + iy, y > 0. If
!
a b
γ=
c d
is a real matrix of determinant 1 and z lies in the upper half-plane, then γz = (az + b)/(cz + d)
again lies in the upper half-plane. Consider the group Γ of integral matrices γ that have
determinant 1 and for which c is divisible by 11.
In the lectures of Klein, as well as in the later treatise of Fricke ([F], p. 406, [Li], §4) two
meromorphic functions (thus complex-analytic except for poles), σ(z), τ (z), on the upper
half-plane are constructed such that σ = σ(z), τ = τ (z) is a point on the curve (j) for any z,
provided that we admit ∞, ∞ as such a point. Furthermore, all points of the curve except
τ = 0, σ = ±1 are obtained in this way, and two values z and z 0 yield the same point if and
only if there is a γ in the group Γ such that z 0 = γz.
If we replace z by the variable q = exp(2πiz), then σ1 dτdq
has an expansion,

1 X
An q n ,
q n=1
and the function,

X An
L(s) = ,
n=1
ns
is, according to Shimura, equal to the L-function attached to the elliptic curve (j) and to (h)
and (i), so that its value is easily calculated for any value of s
Our notions of simplicity are once again overturned. If one avoids equations that are in
any sense abelian then some equations in two variables, those of elliptic curves, appear to
be simpler than those in one variable. They exist with much smaller conductor, and the
connection with modular functions, like σ and τ , is much more immediate.
This is not peculiar to the example (j). If
1 dτ
f (z) = ,
σ dz
then, for any γ in the group Γ
(k) f (γz) = (cz + d)2 f (z).
Such a function is called a modular form of weight 2 with respect to the group Γ. If the
exponent 2 is replaced by 1 the equation (k) defines modular forms of weight 1. Forms of
weight 2 are simply derivatives of invariant functions and therefore much easier to treat than
forms of weight 1. Icosahedral equations are so much more difficult because their regularity
is at best expressed by forms of weight 1.
REPRESENTATION THEORY ITS RISE AND ITS ROLE IN NUMBER THEORY 13

Taniyama’s suggestion, as we now understand it, is that for any elliptic curve (g) with
rational coefficients and therefore with a conductor that will be a positive integer N , we can
find two meromorphic functions x(z), y(z) that parametrize the curve and that are invariant
under the group ΓN of matrices γ that are integral of determinant 1 and satisfy the conditions,
c≡0 (mod N ); a≡1 (mod N ).
It cannot, however, be required in general that two points z and z 0 yield the same point on the
curve only if z 0 = γz with some γ in ΓN . This suggestion can not yet be proved but it can be
easily tested, because for a given integer N there are efficient methods for finding all elliptic
curves that admit parametrizations by functions invariant under ΓN [M]. Of course, even if
Taniyama’s suggestion can be established, the conjecture of Birch and Swinnerton-Dyer will
remain.
We have introduced L-functions as adjuncts to the study of diophantine equations, but
they also appear in the study of automorphic forms, of which the modular forms of weight 1
and 2 in this section are a particular manifestation. It is in this second guise that the values
of the functions can be calculated and the regularity on which we have insisted asserts that
an L-function attached to a diophantine problem is always, in spite of the utterly different
way it is defined, equal to one attached to an automorphic form. It is possible to pass to
the general theory of automorphic forms from the theory of modular functions, but for the
purposes of this essay, in which the emphasis is on the notion of automorphic L-function, it is
best, in order to present it quickly and persuasively, to take our stance within representation
theory.
Representations of groups over the field of real numbers
We begin with a review of the pertinent facts from the representation theory of Lie groups,
taking as our first example the group GL(n, R), a group whose underlying manifold is neither
compact (closed) nor connected, so that we cannot expect its irreducible representations, and
it is these that are of primary interest, usually to be finite-dimensional.
There are at least two ways to analyze or to construct representations of a Lie group: by
passing to the Lie algebra and searching with purely algebraic means for realizations of it;
or to start from a natural action of the group on a manifold, then to pass to the associated
action on functions and to decompose it irreducibly. The standard example is the orthogonal
group in three variables, which acts on the sphere of radius 1 and thus on the functions on
this sphere. This action commutes with the angular term in the Laplacian,
 
1 1
Df = fφφ + (fθ sin θ)θ ,
sin θ sin θ
the subscripts denoting partial differentiation. Then the action of the orthogonal group on
the space of spherical harmonics of order n, defined by
Df + n(n + 1)f = 0,
yields, for n = 0, 1, 2, . . . , the irreducible representation of degree 2n + 1 of the orthogonal
group.
One manifold that is always available is the group manifold itself, on which the group acts
by left or right multiplication, and as appears most clearly in the paper of Peter-Weyl [PW],
if the group is compact all representations can be obtained in the space of functions on the
14 ROBERT P. LANGLANDS

group itself. For example, the representation π being given on the space V , we may choose a
fixed nonzero linear form u on V and sending the vector v to the function
(l) hπ(g)v, ui
realize V together with π on a space of functions. Since the group is compact these functions
are in fact square-integrable, so that, as in [PW], standard elementary techniques from
analysis can be applied. The functions (l) are known as matrix coefficients.
For groups such as GL(n, R) that are not compact similar techniques can be used, but
they cease to be elementary, the system being treated being best compared to a Schrödinger
equation for which both asymptotically independent and bound states appear.
Since every element of GL(n, R) is a product k1 ak2 where k1 and k2 are orthogonal matrices
and a is a diagonal matrix with positive eigenvalues,
 
α1
...
,
 

αn
the simplest representations should have n freely assignable parameters, and are analogous
to a system of n interacting but asymptotically independent particles to which arbitrary
momenta can be assigned. In addition the presence of the factors k1 and k2 may entail the
appearance of discrete quantum numbers. Moreover as a result of the symmetries implied by
the possibility of permuting the αi at the cost of modifying k1 and k2 we are dealing with
questions of reflection more than of scattering, and as a consequence permutations of the n
parameters do not change the representation.
These simple representations have simple constructions. In the group G = GL(n, R) let B
be the group of superdiagonal matrices,
 
α1 ∗ ∗ · · · ∗
α2 ∗ ∗
 

 
(m) b= α 3 ∗ .
.. .. 
 
. . 


αn
Given n real parameters a1 , . . . , an and n numbers mk = ±1 (the supplementary discrete
quantum numbers) we introduce the characters
χk : α → sgn(α)mk |α|iak
of R× , as well as the character χ of B that sends the matrix b to
Yn
χk (αk )|αk |δ−k .
k=1

The number δ = (n + 1)/2 and the supplementary exponent δ − k appear because the group
manifold of G = GL(n, R) is curved. We associate to χ the representation of G on the space
of functions f on G that satisfy
(n) f (bg) = χ(b)f (g)
REPRESENTATION THEORY ITS RISE AND ITS ROLE IN NUMBER THEORY 15

for all b in B and all g in G. It is irreducible. Moreover the representations associated to two
sequences of parameters χk are equivalent if and only if one sequence is a permutation of the
other.
If we replace the exponents iak by arbitrary complex numbers, we may still study the
representation on the space of functions satisfying (n). It will in general be neither unitary
nor irreducible. None the less, there is a well determined process for choosing a specific
irreducible factor of the representation that is then taken as the representation associated
to the sequence of characters, χ1 , . . . , χn . In other words, we allow complex momenta; the
theory of numbers demands that we do so. Once again, the representations attached to two
sequences are equivalent only if the sequences are permutations of each other.
In his basic paper [Ba], Bargmann first constructs representations algebraically by examining
commutation relations, but his most important discovery was, at least if we tailor its
formulation to our present purposes, that there are irreducible representations of GL(2, R)
whose matrix coefficients are square-integrable over the submanifold SL(2, R). In other
words analogues of two-particle bound states appear; they are known as discrete-series
representations of GL(2, R). They are associated to one continuous parameter (or, if the
analogy is pursued, one momentum) and one discrete parameter (or quantum number).
As long as we remain with the general linear group over the real field there are no analogues
of n-particle bound states for n > 2. The reason is that the real numbers have a single algebraic
extension, the field of complex numbers, and it is of degree two. Observe that in the context
of the analogy, a one-particle state is simply a (continuous) homomorphism of GL(1, R), the
multiplicative group of real numbers, into C× , thus, an irreducible representation of GL(1, R),
which is necessarily, the group being commutative, of dimension 1. To have a convenient
terminology, it is usual to call a (continuous) homomorphism of a group into C× a character,
even if its absolute value is not 1.
For the group GL(n, C) of complex matrices, the representation theory is simpler. An
irreducible representation of GL(1, C) is simply a homomorphism from GL(1, C) = C× to C× ,
thus a character of GL(1, C). There are no discrete series for n > 1 and therefore no n-particle
bound states, because, as the fundamental theorem of algebra implies, the field of complex
numbers has no algebraic extension.
For either field, the general irreducible representation of GL(n) is the analogue of r
interacting but asymptotically independent bound states with n1 , . . . , nr particles, the sum
of the nk being n, and the nk being subject to the constraints appropriate to the field. To
construct the representation, we introduce the group P of matrices of the form (m) except
that n is to be replaced by r, each αk is a matrix with nk rows and columns, and the asterisks
represent block matrices of the appropriate size. The space on which the representations act
is once again given by functions satisfying (n), where, however, χ is no longer a character, but
the tensor product of representations χk of GL(nk ), so that the function f takes values in an
infinite-dimensional space. If the momenta are all real, then this representation is irreducible;
otherwise it is necessary to pass again to a specific factor. As before, the order of the χk is
irrelevant.
We conclude that, proofs that involve an elaborate analysis of the pertinent scattering
problems aside, the classification of irreducible representations of the general linear group has
an extreme formal simplicity. For the complex field, to specify a representation of GL(n) we
simply specify n homomorphisms from GL(1, C) = C× to C× . Hence the conclusion, simple
to state but fraught with consequence, that the irreducible representations of GL(n, C) are
16 ROBERT P. LANGLANDS

parametrized by the (completely reducible) n-dimensional representations of the multiplicative


group C× of nonzero complex numbers. Observe that we are comparing collections of two very
different objects: on one hand, irreducible and in general infinite-dimensional representations
of the group GL(n, C); on the other, finite-dimensional indeed n-dimensional but in general
reducible representations of C× .
A similar statement for the real field requires a group that is not commutative but whose
irreducible representations are of degree at most two, in order to accommodate the existence
of two-particle bound states. The appropriate group is labeled WR or WC/R and is obtained
by adjoining to the group C× an element w such that w2 = −1 and wz = z̄w, for z in C× .
Thus it is a kind of dihedral group.
Bound states correspond to irreducible representations and their momentum to the deter-
minant of the representation. Since the homomorphism
z → z z̄, w → −1
yields R× as the maximal abelian quotient of WC/R , the determinant is indeed a character of
R× . Typical representations that yield two-particle bound states are:
! !
zm 0 0 1
z→ ; w→ .
0 z̄ m (−1)m 0
The integer m must, however, be different from zero for the representation to be irreducible.
The group WR is referred to as the Weil group of the real field R. The Weil group of
the complex numbers is just the multiplicative group C× . For both fields, the irreducible
representations of GL(n) are parametrized by n-dimensional representations of the Weil
group. The Weil group can be attached to many of the fields appearing in number theory.
Its simplicity for the real and the complex field is misleading, for it incorporates many of the
deepest facts about the theory of equations known at present. In the following section we
shall comment on its definition for p-adic fields, but we first add some observations on its use
for the field of real numbers.
As a device for inspiring some confidence that L-functions are important mathematical
objects, we have focused on the problem of calculating their values. For this purpose, the
groups GL(n) suffice. On the other hand, some central principles in the subject were suggested
by Harish-Chandra’s practice of treating all reductive groups uniformly, and, despite the
extra baggage they entail, the theory gains both in flexibility and power when reductive
groups are admitted.
In the present context a few remarks suffice, simply to suggest the possibilities. Take the
field to be R. A typical reductive group is the special orthogonal group G in 2n + 1 variables,
of which there are several forms depending on the index of the quadratic form, which can,
for example, be of Euclidean or Minkowski type, so that, even for a fixed n, we are dealing
with groups whose representation theory is at first glance quite distinct.
In order to parametrize the representations of G with the help of the Weil group, we attach
to G its L-group L G, whose general definition is given in terms of the theory of equations
and of the root systems of Killing and Cartan, and which, in this specific case turns out to
be the symplectic group in 2n variables over the complex number field, thus, the group of
complex matrices leaving invariant an alternating form,
x1 y2 − x2 y1 + · · · + x2n−1 y2n − x2n y2n−1 .
REPRESENTATION THEORY ITS RISE AND ITS ROLE IN NUMBER THEORY 17

The principle that the irreducible representations of the real Lie group G are classified by the
(continuous) homomorphisms of WR into L G is valid, but there are some cautionary remarks.
Observe first of all that a reducible n-dimensional representation of WR is a homomorphism of
WR into GL(n, C) such that the image of WR commutes with a subgroup of GL(n, C) that is
isomorphic to C× and not in the center of GL(n, C). Hence the notion of irreducibility has an
obvious analogue for homomorphisms of WR into L G, and homomorphisms that are irreducible
in this sense correspond to representations whose matrix coefficients are square-integrable
over the group manifold.
Of course, if the quadratic form chosen is Euclidean then G is compact and all its
representations have this property, so that any homomorphism of WR into L G that is not
irreducible in the sense indicated will have to be excluded from the list of parameters. A
similar but less restrictive condition must also be imposed for the groups associated to forms
with other indices, but with standard notions from the theory of reductive groups it can be
simply and elegantly formulated, and functions in general.
It also turns out that the classification provided by WR and L G is coarse; some homo-
morphisms correspond to several irreducible representations of G; but this has not marred
the theory. On the contrary, the analysis of the families of finitely many representations
associated to given homomorphisms has revealed unexpected facets of harmonic analysis and
representation theory.
The most fecund principle suggested by the introduction of the L-group is that of functori-
ality. A homomorphism from a group H to a group G does not provide any way to transfer
irreducible representations from one of these groups to the other, unless of course G is abelian,
nor are homomorphisms between H and G usually related to homomorphisms between L H
and L G. On the other hand a homomorphism from L H to L G does, by composition,
WR → L H → L G,
yield a map from the set of parameters for the irreducible representations of H to those
for the irreducible representations of G, and thus implicitly a way to pass from irreducible
representations of H to irreducible representations of G. It has not yet been possible to define
this passage directly, and it promises to be very difficult to do so. The possibility of this
passage is known as the principle of functoriality in the L-group. The analogous principle for
p-adic fields and for the field of rational numbers would be extremely powerful, strong enough,
for example, to imply all forms of the Ramanujan conjecture and to treat the L-functions
of icosahedral representations, and is certainly expected to be valid; but at present it is far
beyond our reach.
Representations of groups over p-adic fields
We have already insisted on the importance of the conductor and therefore on the importance
of studying equations not only modulo a prime number but modulo any integer. Because of
the Chinese Remainder Theorem, it is sufficient to treat congruences modulo a prime power.
We have seen, for example, that
x2 − 5 = (x − 1)2 + 2x − 6 ≡ (x − 1)2 (mod 2).
For the modulus 4 we have,
x2 − 5 = (x − 1)(x + 1) − 4 ≡ (x − 1)(x + 1) (mod 4).
18 ROBERT P. LANGLANDS

On the other hand, there is no integer m such that 8 divides m2 − 5 and hence no possibility
of factoring x2 − 5 modulo 8.
The polynomial x2 − 17 behaves in quite a different fashion.
x2 − 17 = (x − 1)(x + 1) − 16 ≡ (x − 1)(x + 1) (mod 16).
We can continue,
x2 − 17 = (x − 9)(x + 9) + 64 ≡ (x − 9)(x + 9) (mod 64),
x2 − 17 = (x − 41)(x + 41) + 1664 ≡ (x − 41)(x + 41) (mod 128),
because 1664 = 13 · 128, and then
x2 − 17 = (x − 105)(x + 105) + 11008 ≡ (x − 105)(x + 105) (mod 256),
because 11008 = 43 · 256. In general, if
x2 − 17 = x2 − m2 + a2k ,
with two integers m and a, and k > 3, then m is odd and
x2 − 17 = x2 − (m − a2k−1 )2 + b2k+1 ,
with some other integer b. Thus we can proceed indefinitely, constructing a sequence
m1 , m2 , . . . such that
x2 − 17 ≡ (x − mk )(x + mk ) (mod 2k ).
The sequence mk does not converge in the usual sense, because the difference between
mk and mk+1 is equal to an integer times 2k−1 . On the other hand, the difference between
mk and any ml , l > k, is also divisible by 2k−1 , so that if our concern is with divisibility
properties rather than with metric properties then mk and ml are close when k and l are
large. To simplify theoretical discussions, we treat such sequences as real objects, calling
them 2-adic numbers.
More generally, for any prime number p a sequence ak = mk /nk of rational numbers such
that the numerator of the differences,
ak − al = mk /nk − ml /nl , l > k,
is divisible by higher and higher powers of p as k approaches infinity is called a p-adic number,
the p-adic number that this sequence of rational numbers approaches. Of course, constant
sequences a, a, a, . . . are admitted, so that every rational number is a p-adic number. Indeed
in almost every respect p-adic numbers can be treated like real numbers. In particular, they
can be added,subtracted, multiplied, and divided. The only difference is metric.
For the purpose of studying congruences modulo powers of a given prime p, we agree that
a rational number is small if its denominator is prime to p and its numerator is divisible by a
high power of p. Thus, if p = 5 then, in this sense, 1, 000, 000 is small and 1, 000, 000, 000 even
smaller. More precisely, we take the size of a number a whose numerator and denominator
are both prime to p to be 1 and of apk , where k is any integer, positive, negative, or zero, to
be p−k .
|apk |p = p−k .
For example,
3 25 1
= 1, = .
2 5 2 5 25
REPRESENTATION THEORY ITS RISE AND ITS ROLE IN NUMBER THEORY 19

The p-adic distance between two rational numbers, a, b is then just |a − b|p . This distance
extends by passing to limits to a distance defined between any two p-adic numbers, as does the
p-adic absolute value |a|p , so that even the metric properties of p-adic numbers are analogous
to those of real numbers, although in practice quite dissimilar.
We have stressed in our discussion of zeta functions and diophantine equations that the
study of rational solutions of equations usually requires a preliminary examination, often
lengthy and painstaking, of the congruential properties of the equations. The introduction of
the field of p-adic numbers allows us to interpret this as the study of the equations in these
fields. Thus the study of rational solutions of equations is generally preceded by a study of
their p-adic solutions for every prime p.
This requires, among other things, some understanding over p-adic fields of equations of
the form,
xn + an−1 xn−1 + · · · + a0 = 0.
This is the theory of equations over p-adic fields, and is considerably more difficult than
the theory over the real field, because for any p-adic field there are irreducible equations of
arbitrary degree, and not, as for the real field, of degree at most two. Thus the field of p-adic
numbers has algebraic extensions of arbitrarily high degree and the Galois groups of these
extensions are quite complicated.
As our previous discussion suggests, the extensions with abelian Galois groups are much
more accessible than the others, and can be classified. Even so, this requires an elaborate
theory that we cannot rehearse here. The basic results can be encapsulated in the existence
of the Weil group, a way of expressing them that is concise and suggestive, especially for our
present purposes, but hardly reveals their import to the uninitiated.
If we denote the p-adic field by F , then a Galois extension K of F is of course a field
obtained by adjoining to F all the roots of some equation. Rather than using the Weil group
WF , we use the groups WK/F , from which WF is ultimately constructed. The group WK/F
has as subgroup the multiplicative group K × of the field K. It is such that if we divide it by
this subgroup to obtain a quotient group then this quotient is isomorphic to the Galois group
of the extension K. All the art is in piecing these two constituents together. Two properties
worth mentioning are: (i) there is always a homomorphism from WK/F onto F × ; (ii) if K
contains L then there is a homomorphism from WK/F to WL/F .
The Weil group is the correct tool for the classification of the representations of the group
GL(n, F ) of n × n matrices with entries from F . This is a topological group, although not a
Lie group. Nevertheless we can study its (continuous) irreducible representations. There are
now discrete-series representations for any value of n, thus n-particle bound states, because
there are irreducible representations of the Weil group of any degree. Over a p-adic field
the classification of the (continuous) irreducible, and of course generally infinite-dimensional
representations of GL(n, F ) by (continuous) finite-dimensional and, in general, reducible
representations of the Weil group is not yet complete, although much is known [He].
To have a statement exactly like that for the real field, one uses not the Weil group itself
but a thickened form, its product with the group SL(2, C) or, recalling the unitary trick of
Hermann Weyl, the group SU(2). In contrast to the Weil group over the real field, in which
the subgroup C× can transmit geometric information, the Weil group over a p-adic field
contains only information from the theory of equations, and could be defined entirely in terms
of Galois groups. The classification by the thickened group implies that representation theory,
like the theory of automorphic forms, to which we are using it as a steppingstone, is adequate
20 ROBERT P. LANGLANDS

to the description of geometric phenomena that first appear for the curves and surfaces defined
by equations in more than one variable. It is, however, rather the simplest representations of
the group GL(n, F ) that we need to understand before passing to automorphic forms.
The construction used over the real and the complex field may be used here to construct
the representation associated to r discrete-series representations πk of GL(nk , F ) with
n1 + · · · + nr = n.
This is then the counterpart of r interacting but asymptotically independent nk -particle
bound states.
The representations that are of particular importance, the counterparts of n asymptotically
independent particles all of whose supplementary quantum numbers are 0 and which are
therefore characterized by their momenta alone, are usually referred to as unramified repre-
sentations. More precisely, a 1-particle state corresponds to an irreducible representation
of GL(1, F ) = F × and is thus simply a homomorphism from F × to C× . The simplest such
homomorphisms are those of the form,
x → |x|ia
p .
They are characterized by the complex number a alone, which can be regarded as their
momentum, and putting n of them together, we construct an unramified representation on
the space of functions satisfying the relation (n).
Thus an unramified representation π is defined by n numbers a1 , . . . , an or, rather, by the
numbers pia1 , . . . , pian , because |x|p is always a power of p. Since it is only the set of these
numbers that matters and not their order, we can prescribe them simply by giving a matrix
with them as eigenvalues, for example,
 
pia1
pia2
 
 
A(π) =  
..
.
 . 

ian
p
Indeed, it suffices to give the conjugacy class of A(π).
Observing that we can attach to the matrix A(π) the function
L(s, π) = Det(I − A(π)/ps )−1 ,
we are ready to pass to automorphic forms.
Automorphic forms
If we are given for every prime p except those lying in some finite set S an unramified
representation πp of the general linear group GL(n, Qp ) of matrices with coefficients from the
p-adic field Qp , then we can form the infinite product
Y
(o) L(s) = L(s, πp ).
p6∈S

It converges if s is a sufficiently large real number, say s > a + 1 provided that the eigenvalues
of each A(πp ) are less than pa in absolute value. There is, however, no reason to believe that
it will even define a function having a meaning for any other values of s, and much less that
the function could be calculated.
REPRESENTATION THEORY ITS RISE AND ITS ROLE IN NUMBER THEORY 21

In order that this procedure lead to a function defined outside the region of convergence
of the infinite product, there must be some coherence between the matrices A(πp ), and the
appropriate way to assure this is to demand that the collection of representations πp be
associated to an automorphic form.
An efficient, but abstract, way to approach the subject of automorphic forms is by the
introduction of adeles, rather ungainly objects that nevertheless, once familiar, spare much
unnecessary thought and many useless calculations. We shall use them first as a conceptual
aid, adding later some remarks to show how one works with them in practice. The fields of
p-adic numbers were introduced as a compact way to study congruences modulo ever higher
powers of prime numbers. If we want to consider congruences modulo arbitrary integers m
and at the same time monitor the size, in the usual sense, of the numbers involved, then
we use the adeles. An adele is a sequence {a∞ , a2 , a3 , a5 , . . . , ap , . . . } in which a∞ is a real
number, and, for each p, ap is a p-adic number. The only constraint is that for all but a
finite number of p the number ap be integral, in the sense that a = bpk , with k > 0 and
|b|p = 1. Thus if b were rational then both its numerator and its denominator would be prime
to p. Observe, for it is very important to us, that if we take any rational number a then the
sequence {a, a, a, a, a, . . . } is an adele because the denominator of a is divisible by only a finite
number of primes. It is possible to add, subtract, and multiply adeles, but not necessarily
to divide one by another. In particular, in order that the adele 1/a, {a = a∞ , a2 , . . . }, be
defined it is necessary that |ap |p = 1 for all but finitely many p
We may consider matrices whose entries are adeles and form their determinants. If the
determinant d of an adelic matrix has an inverse 1/d then, using minors in the customary
way, we can form the inverse of the matrix itself. Thus the collection of all adelic matrices
with invertible determinant forms a group GL(n, A), or more briefly G. Taking the entries
one coordinate at a time we can construct from an element g of G a sequence of elements
g∞ , g2 , g3 , . . . with g∞ in the group GL(n, R) and with gp in the group GL(n, Qp ). As a
consequence, from an irreducible representation π of the group G we can construct irreducible
representations of the group GL(n, R) and of the groups GL(n, Qp ), and this is the first clue
to the construction of coherent sequences of πp .
Since every rational number is an adele, every matrix γ in Γ = GL(n, Q) is also a matrix
in G so that Γ is a subgroup of G. An irreducible representation is said to be automorphic if
it can be realized on a space of functions f on the group G satisfying
f (γg) = f (g),
for all g ∈ G and all γ ∈ Γ. The operator π(h), h ∈ G, sends f to the function f 0 (g) = f (gh).
The automorphic forms themselves are these functions f . An automorphic representation
then yields a sequence πp that is unramified for almost all p and sufficiently coherent that
the function
(p) L(s) = L(s, π)
defined by (o) is meaningful for all except a finite number of values of s. Moreover, from f
we can calculate its values as precisely as we like.
The functions (p) are called automorphic L-functions and appear to be exactly what we
need for diophantine purposes, although it is extremely difficult to show that the L-function
associated to a problem in diophantine equations is equal to an automorphic L-function, and
our conjectures have far outstripped our ability to prove them. None the less, although the
theorems necessary to deal with elliptic curves are not yet available in any real generality,
22 ROBERT P. LANGLANDS

there is a great deal of very solid evidence available for other problems [K1], so that there
are no serious doubts that we are on the right track.
Convenient though they are, the use of adeles does at first leave most operational questions
unanswered, so that, for example, it is by no means obvious how we are going to use elementary
Fourier analysis to calculate the values of our L-functions, or that the functions (p) are in
any way related to the functions (f).
To see that they are, take n = 1. The group G is then just the group of invertible ideles,
usually denoted A× . Since it is abelian, all irreducible representations are of dimension one,
simply characters of the group. What kind of coherence is entailed by the demand that the
representation be automorphic? In the present context, where the group is abelian, this is
the condition that the character equal 1 on the subgroup Q× of A× .
Introduce the subgroup U of elements a∞ , a2 , a3 , . . . such that |ap |p = 1 for all prime
numbers p and a∞ is positive. Every element of the group A× can be written in a unique
way as the product of an element in Q× and an element of U . Indeed given any adele with
an inverse, so that all its coordinates are nonzero, we first multiply it by ±1 to obtain an
adele with positive coordinate a∞ . If the other coordinates are ap and if
|ap |p = p−mp ,
then mp is zero for all but finitely many p, so that we can form the positive rational number
Y
b= pm p .
Since |ap /b|p = 1 for all p, the original adele is equal to ±b times an element of U . Upon a
little reflection one sees that this factorization is unique.
Thus an automorphic representation becomes even simpler, just a character χ of U . The
condition that it be continuous, on which we have not previously insisted, turns it into a
completely elementary object. It implies that χ depends on only finitely many coordinates
u∞ , up1 , . . . , upr of an element u and that further in its dependence on each upi , it depends
only on the residue of upi modulo some power pni of pi , the same for all u, but, of course, not
for all χ.
The set S of primes to exclude from the product (p) is {p1 , . . . , pr }. The representations
πp associated to the automorphic representation π defined by χ are themselves characters,
but of the group Q× p = GL(1, Qp ), and are easily constructed. The representation π sends an
element a = bu, where b ∈ Q× and u ∈ U , to π(a) = χ(u). To obtain πp we take an element
ap in Qp , extend it to an adele a with the coordinate ap at p but all other coordinates equal
to 1, and then set πp (ap ) = π(a).
The function L(s, π) is a product of the factors
L(s, πp ) = (1 − A(πp )/ps )−1 ,
the matrix A(πp ) becoming for n = 1 a number. It is equal to πp (p−1 ) because πp (p−1 ) = pa if,
in general, πp (ap ) = |ap |ap . To calculate πp (p−1 ), an expression in which p is a p-adic number,
we have to write the adele 1, 1, . . . , 1, p−1 , 1, . . . as the product of a rational number and an
element u in U . The rational number is p−1 itself, but then u must be
p, p, . . . , p, 1, p, . . . .
Thus it has the coordinate 1 only in Qp . Choosing a character χ that does not depend on
the coordinate u∞ of u in R, we conclude finally, on recalling the properties of χ, that A(πp )
depends only on the residue of p modulo the numbers pn1 1 , . . . , pnr r . Thus Dirichlet characters
REPRESENTATION THEORY ITS RISE AND ITS ROLE IN NUMBER THEORY 23

are automorphic representations of GL(1, A). Indeed, there is almost no distinction between
the two notions.
Hence, for n = 1, our constructions lead us back to objects that we had already introduced
in a much simpler way. This is not very persuasive evidence either for the necessity or even
for the utility of the notion of adele. For n = 2, however, the constructions lead to quite
different objects, in particular, to modular forms in the usual sense. This similarity between
Dirichlet characters and modular forms that the use of adeles renders evident went unnoticed
for decades, even apparently by Erich Hecke who did so much with both; so that, although
the arithmetic of equations with abelian Galois group, begun in the eighteenth century, was
well understood at the end of the first quarter of this century, and even in many respects
earlier, another forty years elapsed before it was realized that the vehicle for expressing the
regularities that, it was hoped, would appear for all equations was already at hand, and had
been so for a very long time. For n > 2, there is seldom an advantage in deciphering the
adelic language.
Take n = 2 and, for simplicity, consider an automorphic representation that is unramified
at all primes p. We now introduce another group U , a group of 2 × 2 matrices u with adelic
entries, and with further special properties to be prescribed. Suppose that
!
a b
u=
c d
Each of the four entries has coordinates that we denote by the appropriate subscripts. We
suppose first that ! !
a∞ b ∞ 1 0
= .
c∞ d ∞ 0 1
For each p we suppose that
|ap |p 6 1, |bp |p 6 1, |cp |p 6 1, |dp |p 6 1,
and that
|ap dp − bp cp |p = 1.
Thus up and its inverse are both integral matrices in the p-adic sense.
If an automorphic representation is unramified then it can be realized on a space containing
functions f satisfying
(q) f (γgu) = f (g),
for all g ∈ G, γ ∈ Γ, and u ∈ U . On the other hand, elementary considerations a little more
elaborate than those involved for n = 1 show that every element of the group G = GL(2, A)
is a product g = γhu, where h is a matrix in GL(2, R). Since f (g) = f (h), we may think of
f as a function on GL(2, R). It is by no means an arbitrary function. First of all, if δ is an
integral matrix with integral inverse then f (δh) = f (h); and secondly, the representation π∞
on the space of functions obtained in this way is irreducible.
To make the connection with modular forms, we suppose that π∞ belongs to the discrete
series and that the associated momentum is 0, or, in straightforward terms, that π∞ (z) = 1
if z is a scalar matrix. We can find an even positive integer l > 2 such that the function f of
(q) satisfies in addition the relation,
(r) f (hk) = e−ilθ f (h),
24 ROBERT P. LANGLANDS

if !
cos θ − sin θ
k=
sin θ cos θ
is an orthogonal matrix. If l is chosen minimal then, in addition, f satisfies the relation,
d d
(s) f (g exp(tX1 )) − i f (g exp(tX2 )) = 0,
dt dt
where ! !
1 0 0 1
X1 = ; X2 = .
0 −1 1 0
If we then define the function h(z) on the upper half-plane by
!
l ai + b a b
h(gi) = (ci + d) f (g), gi = , g= ,
ci + d c d
then equation (s) becomes the Cauchy-Riemann equation guaranteeing that h is an analytic
function of z and (q) and (r) imply that it is a modular form of weight l. Because the
automorphic representation was taken to be unramified it is even a modular form with respect
to the group Γ1 . In general, a conductor N appears.
As well as giving a little more solid or intuitive content to otherwise tenuous notions,
these somewhat technical discussions for n = 1 and n = 2 make it clear that there is very
little difference between functions on GL(n, A) invariant under GL(n, Q) and functions on
GL(n, R) invariant under GL(n, Z), or, more precisely, a subgroup of it. Since GL(n, R)
almost completely fills the vector space V of n × n real matrices, it is not astonishing that we
can use Fourier analysis on this space to define and calculate the value of the L-functions (p)
for all values of s, although the arguments and calculations are not completely patent. They
are more transparent if one remains with GL(n, A).
Final reflections
The introduction of infinite-dimensional representations entailed an abrupt transition
in the level of the discourse, from explicit examples to notions that were only described
metaphorically, but at both levels we are dealing with a tissue of conjectures that cannot be
attacked frontally.
The aesthetic tension between the immediate appeal of concrete facts and problems on
the one hand, and, on the other, their function as the vehicle to express and reveal not so
much universal laws as an entity of a different kind, of which these laws are the very mode of
being, is perhaps more widely acknowledged in physics, where it has long been accepted that
the notions needed to understand perceived reality may bear little resemblance to it, than in
mathematics, where oddly enough, especially among number theorists, conceptual novelty
has frequently been deprecated as a reluctance to face the concrete and a flight from it.
Developments of the last half-century have matured us, as an examination of Gerd Falting’s
proof of the Mordell conjecture makes clear [Fa], but there is a further stage to reach.
It may be that we are hampered by the absence of a central unresolved difficulty and by
the extremely large number of currently inaccessible conjectures, at whose extent we have
hardly hinted. Some are thoroughly tested; others are in doubt, but they form a coherent
whole. What we do in the face of them, whether we search for specific or general theorems,
REPRESENTATION THEORY ITS RISE AND ITS ROLE IN NUMBER THEORY 25

will be determined by our temperament or mood. For those who thrive on the interplay of
the abstract and the concrete, the principle of functoriality for the field of rational numbers
and other fields of numbers has been particularly successful in suggesting problems that are
difficult, that deepen our understanding, and yet before which we are not completely helpless.
In particular, there are methods, the Arthur-Selberg trace formula, and classes of diophantine
problems, those defining Shimura varieties, that are intimately connected to representation
theory and whose mastery probably has to precede any attack on the general form of the
principle of functoriality. The papers [AC] and [K2] surmount longstanding obstacles to the
application of the trace formula and to the analysis of the zeta-functions of Shimura varieties,
and are rich sources of technique for anyone wishing to penetrate the area.
Despite its importance as a guiding principle, I have given no prominence to functoriality,
alluding to it only briefly, preferring, for the sake of cogency, to place the sequence, diophantine
equation, values of L-functions, automorphic L-functions, in stark relief. This entails an
emphasis on a fixed GL(n), to the prejudice of other groups, whereas the primary force of the
principle of functoriality is its insistence on the intimate connection between representations,
above all automorphic representations, of different groups. I have also restricted myself to
the simplest possible equations that could serve as illustrations. The L-functions attached
to algebraic curves are closely related to algebraic integrals on these curves, so that their
theory cannot be simpler than that of these integrals, whose concrete geometric implications
for curves of higher genus are much less clear than its theoretical consequences. The same
disadvantages are attached to the use of L-functions.
References
[AC] J. Arthur and L. Clozel, Simple algebras, base change, and the advanced theory of the trace formula,
Annals of Math. Studies, no. 120, Princeton University Press (1989).
[Ba] V. Bargmann, Irreducible unitary representations of the Lorentz group, Ann. of Math. 48, 568–640
(1947)
[BSD] B. Birch and P. Swinnerton-Dyer, Notes on elliptic curves, I and II, Jour. f. d. reine u. ang. Math.,
Bd. 212, 7–25 (1963), and Bd. 218, 79–108 (1965).
[Bu] J. P. Buhler, Icosahedral Galois representations, SLN in Math., v. 654 (1978).
[D] R. Dedekind, Gesammelte mathematische Werke, reprinted by Chelsea Publishing Co. (1969).
[Fa] G. Faltings Endlichkeitssätze für abelsche Varietäten über Zahlkörper, Inv. Math. v. 73, 561–584
(1983).
[F] R. Fricke, Die elliptische Funktionen und ihre Anwendungen, T. II, B. G. Teubner, 1922.
[H] G. Harder, Experimente in der Mathematik, Rheinische-Westfälische Akad. der Wiss., Vorträge, N. 321
(1983)
[Ha] T. Hawkins, The creation of the theory of group characters, Rice University studies, v. 64 (1978)
[He] G. Henniart, Les conjectures de Langlands locales pour GL(n), in Journées arithmétiques de Metz,
Astérisque, no. 94, 67–85 (1982).
[KF] F. Klein and R. Fricke, Vorlesungen über die Theorie der elliptischen Modulfunktionen, Bd. II, B. G.
Teubner, Leipzig (1892)
[K1] R. Kottwitz, On the λ-adic representations associated to some simple Shimura varieties, to appear.
[K2] , Shimura varieties and λ-adic representations, Automorphic Forms, Shimura Varieties,
and L-functions, vol. I. (L. Cluzel and J. S. Milne, eds.), Academic Press, 161–209, (1990).
[L1] R. Langlands, Problems in the theory of automorphic forms, in SLN in Math., v. 170 (1970).
[L2] , Base change for GL(2), Annals of Math. Studies, no. 96 (1980).
[Li] G. Lizogat, Courbes modulaires de genre 1, Bull. de la Soc. Math. de France, Mém. 43, v. 103 (1975)
[M] J.-F. Mestre, Courbes de Weil et courbes supersingulières, Seminar on Number Theory 1984–1985,
Exp. No. 23, Univ. Bordeaux I, Talence, 1985.
26 ROBERT P. LANGLANDS

[PW] F. Peter and H. Weyl, Die Vollständigkeit der primitiven Darstellungen einer geschlossenen kon-
tinuierlichen Gruppe, Math. Ann. Bd. 97, 737–755 (1927).
[Sh] G. Shimura, Yutaka Taniyama and his time, very personal recollections, Bull. London Math. Soc., v.
21, 186–196 (1989).
[V] J. Vélu, Isogénies entre courbes elliptiques, C. R. Acad. Sc. Paris, A, t. 273, 238–241 (1971).
Compiled on July 3, 2018 4:55pm -04:00

You might also like