The Threefold Way

JOURNAL OF MATHEMATICAL PHYSICS
VOLUME
3,
NUMBER
NOVEMBER-DECEMBER
1962
The Threefold Way.

Algebraic Structure of Symmetry Groups and Ensembles in Quantum Mechanics
FREEMAN
J.
DYSON
Imtitute for Advanced Study, Princeton, New Jersey

(Received June 22, 1962)
Using mathematical tools developed by Hermann Weyl, the Wigner cla.eeification of group-representations and co-representations is clarified and extended. The three types of representation, and
the three types of co-representation, are shown to be directly related to the three types of division
algebra with real coefficients, namely, the real numbers, complex numbers, and quaternionB. The
author's theory of matrix ensembles, in which again three poeeible types were found, is shown to be in
exact correspondence with the Wigner cla.eeification of co-representations. In particular, it is proved
that the moot general kind of matrix ensemble, defined with a symmetry group which may be completely arbitrary, reduces to a direct product of independent irreducible ensembles each of which
belongs to one of the three known types.
I. INTRODUCTION
HE purpose of this paper is to bring together

and' unify three trends of thought which have
grown up independently during the last thirty years.
These are (i) the classification by Wigner1 of representations of groups which include time-inversion,
(ii) Weyl's general theory of matric algebras and
their commutator algebras/ and (iii) the study of
ensembles of random matrices, begun by Wigner3
and continued by various other physicists.4 It will
be shown that these three theories are all variations
upon a single mathematical theme. It is not surprising that the three theories should turn out to
be closely related, since they all took their origin
from the work of the great algebraists Frobenius
and Schur at the beginning of the twentieth century.6
Our way is threefold in another and deeper sense.
1 E. P. Wigner, Nachr. Akad. Wiss. Gottingen, Math.
physik. ID., 546 (1932). See also, E. P. Wigner, Group Theory
and its Application to the Quantum Mechanics of Atomic
Spectra (Academic PreBS Inc., New York, 1959), English
edition, Chaps 24 and 26.
2 H. Weyl, The Classical Groups, Their Invariants and
Representations (Princeton University Press, Princeton, New
Jersey, 1939). Chapter 3 of this book contains the essential
theorems on which all of our arguments hang. For Weyl's
treatment of semilinear representations, see Duke Math. J. 3,
200 (1937).
3 E. P. Wigner, Ann. Math. 53, 36 (1951); 62, 548 (1955);
65, 203 (1957); 67, 325 (1958).
4 F. J. Dyson, J. Math. Phys. 3, 140, 157, and 166 (1962).
This series of three papers includes references to earlier work
by others in the same field. Paper IV in the series is being
written in collaboration with Dr. M. L. Mehta and will be
published later. The present paper should logically be considered to be number zero in the series, since it provides an
improved mathematical and logical foundation for the rest of
the series. Since Roman numerals contain no symbol for zero,
we preferred to publish the present paper under a separate
title.
Ii A sketch of the historical development is to be found in
the section headed "Remembrance of Things Pa.et" in Weyl's
book (reference 2), p. 27.
In each of the three theories which we aim to unify,

there appears a triple alternative, a choice between
three mutually exclusive possibilities. (i) The irreducible representations of a group by unitary
matrices fall into three classes, which are called
potentially real, complex, and pseudoreal.6 Another,
and quite independent, threefold choice exists for
representations of a group by unitary and antiunitary matrices. Wigner7 calls such representations
co-representations, and he classifies them into
types I, II, and III. (ii) The classical groups studied
by Weyl are of three types, namely orthogonal,
unitary, and symplectic. (iii) The present author4
found three distinct kinds of ensembles of random
matrices, to which he attached the same three names
as are given to the classical groups. In the previous
discussion of matrix ensembles,4 the question whether
all irreducible ensembles belong to one of these three
types was not raised. This question will here be
answered in the affirmative.
The recurrence of the threefold choice in all these
contexts gave the first hint that a unified mathematical treatment of group representations, commutator algebras, and ensembles should be possible.
It was Bargmann who pointed out to the authorS
that the root of the matter is to be found in the
classical theorem of Frobenius. 9
Frobeniua' Theorem. Over the real number field
6 Chapter 24 of Wigner's book (reference 1). This claSBmcation was discovered by A. Loewy, Trans. Am. Math. Soc. 4
171 (1903). See also G. Frobenius and I. Schur, Sitzber:
preuss. Akad. Wiss., Physik.-math. Kl. 186 (1906).
7 Chapter 26 of Wigner's book (reference 1).
8 V. Bargmann (private communication).
9 G. Frobenius, J. reine u. angew. Math. 84, 59 (1878)
L. E. Dickson, Linear Algebras (Cambridge University PreBS;
New York, 1914), p. 10.
1199
This article is copyrighted as indicated in the article. Reuse of AIP content is subject to the terms at: http://scitation.aip.org/termsconditions. Downloaded to
IP: 134.157.146.58 On: Sun, 21 Jun 2015 20:49:04
FREEMAN J. DYSON
1200
10
there exist precisely three associative division algebras, namely the real numbers, the complex numbers,
and the real quaternions.
Once this is understood, the further development

of the theory is extremely simple. All that is necessary is to apply the general theorems of Weyl' to
the special case in which the ground field of the
matric s.lgebras is the field of res.l numbers.
Probs.bly all these connections would have been
clarified long ago, if quantum physicists had not
been hampered by a prejudice in favor of complex
and s.gs.inst real numbers. It hs.s been generally
believed that only the complex numbers could
legitimately be used as the ground field in discussing
quantum-mechanical operators. Over the complex
field, Frobenius' theorem is of course not valid;
the only division algebra over the complex field is
formed by the complex numbers themselves. However, Frobenius' theorem is relevant precisely beCs.use the appropriate ground field for much of
quantum mechanics is real rather than complex.
Specifically, as soon as anti-unitary opers.tors such
as time inversion are included, it is simpler and
more natural to work with a real ground field than
to follow Weyl2 in studying semilinear operators
over the complex field.
Physicists have known for a long time that in
practice, when invariance under time-inversion is
in question, complex phases are no longer arbitrary
s.nd undetermined coefficients may be taken to be
real. Physicists are, in fs.ct, like M. Jourdain talking
prose, using the real numbers for their ground field
without knowing it. One purpose of this paper is
to make the use of the real ground field in quantum
mechanics official and undisguised. l l No change in
the physical content of the theory is thereby implied. Only it may be es.sier for students to understand what they are doing if the mathematical
10 The restriction to associative a4ebras is forced by the
fact that the rule of matrix multiplicatIOn is associative. In all
applications of group theory to quantum mechanics we
identify the operation of multiplication with ordinary matrix
multiplication. It is well-known that a fourth division algebra
over the real number field exists, namely the algebra of
octonions, if multiplication is allowed to be nonassociative.
It is interesting to speculate uponlOBBible physical interpre[see . Pais, Phys. Rev. Letters
tations of the octonion
7, 291, 1961]. We have tried, and failed, to find a natural way
to fit octonions into the mathematical framework developed
in this paper.
11 The general formalism of quantum mechanics over a
real ground field has been worked out by E. C. G. Stueckelberg, Helv. Phys. Acta 32, 254 (1959); 33, 727 (1960). Two
further papers by Stueckelberg and collaborators have been
circulated as preprints and will appear in Helv. Phys. Acta.
These papers have many points of contact with the present
work. For a brief summary of Stueckelberg's conclusions, see
also the paper of Finkelstein et al. (reference 12).
formalism is brought into closer correspondence with

physical pmctice.
A final by-product of the work described in this
paper is that it defines an area of quantum mechanics
within which quaternions play a natural and essential role. Several attempts have been made in the
past 12 to construct a radically new version of quantum mechanics in which complex numbers are from
the beginning replaced by quaternions. Our analysis
has nothing to do with these attempts. Proceeding
in a modest and conservative spirit, we merely
show that quaternions form the appropriate algebraic basis for a description of nature whenever
we have to deal either with pseudoreal group
representations or with co-representations of Wigner's type II. The context in which quaternions
arose historically, in a study of the three-dimensional rotation group, can now be seen to be an
extremely special case of this general principle.
Every group which admits pseudoreal representations equally admits a natural description in terms
of real quaternions.
U. GROUP ALGEBRA AND COMMUTATOR ALGEBRA
The starting point of our analysis is a group G

which is supposed to be a symmetry-group for some
quantum-mechanical system. For example, G could
be a rotation group, or an isotopic-spin group, or
a time-inversion group, or all of these in combination. The quantum-mechanical states belong to a
linear vector space H. of finite dimension n over
the field C of complex numbers. An element g of
G is represented in H. by an operator A(g) which
is either unitary or antiunitary. Physically, the
antiunitary A(g) will correspond to operations g
which involve time-inversion. We make the convention that the letter g may denote any element of
G, the letter u denotes an element for which A(u)
is unitary, and the letter a denotes an element for
which A(a) is antiunitary. The set of u forms a
subgroup G1 of G. We assume that G contains some
antiunitary elements a. Then G1 is an invariant
subgroup of G with index 2. The a form a set G,
which is the unique co-set of G1 in G.
The A(a) are not matrices over the field of complex numbers. The notion of group representation
can be enlarged, following Weyl2 and Wigner,7 so
as to include such semilinear operations. However,
we find it simpler and more fruitful to represent
12 G. BirkhofJ and J. von Neumann, Ann. Math. 37, 823
(1936). E. J. Schremp, Phys. Rev. 99, 1603 (1955); 113, 936
(1959). D. Finkelstein, J. M. Jauch, S. Schiminovich, and D.
Speiser, J. Math. Phys. 3, 207 (1962).
IP: 134.157.146.58 On: Sun, 21 Jun 2015 20:49:04
1201
THE THREEFOLD WAY
the A(a) by true matrices over the field R of real

numbers. We define the correspondence
(1)
A(g) +-+ M(g)
in the following way. M(g) is a [2n X 2n] matrix

with real elements. Each (2 X 2) block in M (u)
is derived from a single element of the [n X n]
complex matrix A(u) by the replacement
a
+-+
fa
lp
(2)
a matric algebra over R. In particular we define
= commutator algebra of A,
(11)
Y = commutator algebra of B,
(12)
X
Z
commutator algebra of D.
The inclusion relations

ACBCD
(14)
Y:) Z.
(15)
immediately imply
X:)
Each A(a) is of the form

A(a) = U(a)j,
(3)
where U(a) is unitary and j is the operation of

complex conjugation. The M(a) are defined by
making the substitution (2) in U(a) together with
the replacement
j = I ..
X[1 OJ.
(4)
-1
The space H R in which the matrices M(g) operate

is a real 2n-dimensional vector space. Each vector
in H B is composed of the real and imaginary parts
of the components of the corresponding vector in H c'
It is convenient to consider the symbol
(5)
also as a matrix operating in HR'

The M (g) now form a true 2n-dimensional representation of the group G over the field R. The distinction between unitary and antiunitary elements
of G is provided by the commutation rules
M(u)i = iM(u) J
(6)
(7)
M(a)i
-W(a).
All the M (g) are orthogonal matrices.

A matric algebra over R is defined as a set of
matrices which is closed under the three operations
of addition, matrix multiplication, and multiplication by scalar coefficients in R. Three such algebras
will now be introduced:
A generated by the M(u),
(8)
B generated by the M(u) and i,
(9)
D generated by the M(u), M(a), and i.
(10)
The commutator algebra of a given algebra K is

defined as the set of matrices which commute with
all matrices in K. The commutator algebra is itself
(13)
The algebra A is given the name "group algebra

of Glover R." In an obvious sense, B is identical
with the group algebra of Glover O. The algebra
D is not a group algebra over 0 in the ordinary
sense, but it may be considered to be the group
algebra of Gover O. However, it is important that
we have defined each of A, B, D as algebras with
coefficients in R.
We next introduce some convenient notations;
following Weyl. 2 If K is any algebra and m a positive
integer, we denote by mK the algebra of matrices
consisting of m identical blocks,
MOO
o ...
(16)
with MinK. Symbolically, we may write this as

an outer product,
mK
= I", X K.
(17)
We denote by [Kl ... the algebra of all matrices consisting of m 2 blocks,
(18)
..1
M ..,
M..
with each M" independently a matrix in K. In

particular, when K = R is the algebra of scalars,
R", is the algebra of all real matrices of degree m.
Two algebras A, A' are said to be equivalent
(A ,...., A') if there exists a fixed nonsinguIar matrix
N such that every matrix M of A is related to the
corresponding M' in A' by
M
= NM'N-1
(19)
If A is any algebra, the inverse algebra A is obtained
IP: 134.157.146.58 On: Sun, 21 Jun 2015 20:49:04
FREEMAN J. DYSON
1202
from A by inverting the order of factors in all

products, thus
C is commutative, C = C. The quaternion division

algebra Q has degree 4 and is generated by
o
o
(21)
Finally, a division algebra is defined to be an algebra

in which every nonzero element M has a reciprocal
M- 1
With these notations and definitions, we are in a
position to state the main theorem of Weyl.13
Weyl's Theorem. Let K be any group algebra over
R, and L its oommutator algebra. Then K and L
are simultaneously equivalent to the canonioal forms
o
o
(20)
if and only if
rv
L s;[E;]'iI
;
rv
L
;
t;[.E;]'f'
" r-:
The summations here represent direot sums over

diagonal blooks of matrioes. Each value of j corresponds to one inequivalent irreducible representation
of the group r whioh generates Kover R. For each j,
E; is a division algebra, and s;, t; are positive integers.
The matrix block corresponding to index j has degree
T2
Ta =
(23)
where e; is the degree of E;.
The following remarks may be made concerning
this theorem.
Remark 1. The relation between the algebras K
and L is symmetrical. Thus K is also the commutator
algebra of L.
Remark 2. When the sums (22) reduce to a single
term, the algebras K and L are called simple. In
this case the suffixes j may be dropped.
Remark 3. When K is generated by an irreducible
representation of r, K is simple and the integer-s
is equal to unity. In this case
rv
L '" tEo
-[0 -1J
e2 -
o
o
o
o
o
-1
o
o
o
o
-1
o
o
o
o
-1
(25)
13 This is theorem (3.5B) on p. 95 of Weyl's book (reference

2), combined with the theorem that every group ring is fully
reducible (p. 101 of the same book).
-1
o
o
o
(26)
0-1
o
o
o
0_
o
o
o
o
o
1
The inverse algebra Q is then generated by
Note that this e2 is not necessarily identical with the

original imaginary unit i defined by Eq. (5). Since
-1
(24)
"
Remark 4. By Frobenius' theorem (see Sec. I),
the possible division algebras over R are three in
number, and are denoted by R, C, and Q. R has
degree 1, and is generated by the scalar Ii = 1.
C has degree 2 and is generated by
o -1
(22)
o
o
o
o
o
o
-1
o
o
1
o
o
o
o
o
o
-1
o
o
1
o
o
1
o
o
o
o
1
o
o
o
1
o
o
-1
o
o
o
o
o
-1
o
o
o
o
o
(27)
This particular representation of Q and
Q by real
IP: 134.157.146.58 On: Sun, 21 Jun 2015 20:49:04
THE THREEFOLD WAY
matrices is called the regular representation. It has

the property that all matrices in Q commute with
all matrices in Q. Thus Q and Q are commutator
algebras of each other, as required by the theorem.
Remark 5. When K is a simple algebra, the division
algebra E is uniquely fixed and must be either R, C,
or Q. In these three cases we say that the representation of r by K is of type R, type C, or type Q.
respectively.
Remark 6. We shall apply Weyl's theorem to the
algebras A, B, and D defined at the beginning of
this section. In the case of A, the group r is identical
with G 1 In the case of B, the group r is the direct
product of G1 with the Abelian group r 4 generated
by (I, i). In the case of D, the group r is the product
of G with r 4, the commutation rules between G
and r 4 being given by Eqs. (6) and (7). Each of
A, B, D is thus a group algebra over R in the
ordinary sense, although only B is a group algebra
over C.
The following lemma14 is important in determining the structural relations between the algebras
B andD.
Lemma. Let M1(g), M2(g) be two inequivalent irredtwible oomponents of the algebra D. Then tke subalgebras M1(u) , M2(U) in B are inequivalent, and
no irreducible component of M 1 (u) oan be equivalent
to any irreducible component of M 2 (u).
To prove the lemma, we assume that M1(g) and
M 2(g) are inequivalent and that M1(u) and M2(U)
have two equivalent irreducible components. There
then exists a matrix P in the algebra Y, linking
the two inequivalent blocks Ml and M2 of the
algebra D, but commuting with the algebra B.
This P satisfies
Pi
= iP,
(28)
for any two antiunitary elements all a2 in G. Therefore,

[M2 (al)]-lPM 1(al) = [M2 (a2)r 1PM1(a2) = W, (29)
where W is a matrix independent of all a2. Hence
(30)
for all a in G. Since Eq. (30) also holds with a replaced by a-I, we have
(31)
14 This lemma could probably be deduced as a special case
from the general theorems of A. H. Clifford, Ann. Math. 38,
533 (1937), concerning the connections between representations of groups and subgroups. However, it seemed simpler to
give a direct and elementary proof of the lemma without
appeal to Clifford's work.
1203
Therefore,
(P
+ W)M1(a)
M 2(a)(P
+ W)
(32)
+ W).
(33)
for all a in G, and this implies

(P
+ W)M1(g)
M 2(g)(P
Since M1(g) and M2(g) are supposed irreducible

and inequivalent, Schur's lemma 15 now implies
P
+ W = O.
(34)
But then Eq. (30) becomes

PM1(a) = -M2(a)p.
(35)
Equations (28) and (35) together give

iPM1(g)
= M 2(g)iP
(36)
for all g in G, and therefore by Schur's lemma again

iP = O.
Thus the operator P cannot exist, and the lemma

is proved.
Remark 7. An equivalent statement of this lemma
is as follows. Let the algebras Y and Z be written
in the canonical form of Weyl's theorem as direct
sums of diagonal blocks,
Y
= Lk
Y.. ,
= L;
Z;,
(37)
where the Z; are inequivalent simple algebras and

likewise the Y k The lemma states that each Y ..
is confined to a single block containing precisely
one Z;. This means that the structural relation
between Y and Z is completely determined by considering the separate blocks Zj.
m.
WIGNER'S CLASSIFICATION OF IRREDUCIBLE

REPRESENTATIONS
In this section we shall establish the connection

between Weyl's theory of group algebras (Sec. II)
and the classification of group representations by
Wigner. 1
A few preliminary observations must first be made.
The equivalence relations in Eq. (22) refer to a
transformation to canonical forms by a similarity
relation (19) in which N may be an arbitrary real
nonsingular matrix. According to a standard argument,16 since the algebra. K is generated by an
orthogonal group representation, the transformation
matrix N may be chosen to be orthogonaL Next we
show that, when the algebra K is one of the trio A,
B, or D, the matrix N may be chosen so as to
commute with i. The operator i belongs to Band D,
16
16
See Wigner (reference 1), p. 75, Theorem 2.

See Wigner (reference 1), p. 78.
IP: 134.157.146.58 On: Sun, 21 Jun 2015 20:49:04
FREEMAN J. DYSON
1204
and to the commutator algebras X and Y of A and B.

So in each of the three cases, i belongs either to
K or to L. When the transition to canonical forms
is made, i is transformed into some matrix i' which
has nonzero elements only within the blocks where
the canonical form of K or L exists. The transformed i' still satisfies
[i,]2
-I.
(38)
It is therefore possible to transform i' back into the
standard form i by a real orthogonal transformation

working within each diagonal block separately. As
a result, we have an orthogonal matrix N which
transforms K and L into the canonical forms (22)
and transforms i into i. This N then commutes
with i.
When N is chosen to be orthogonal and to commute with i, N is identical with a unitary transformation of the original complex vector space H .
Thus the canonical forms (22) are obtained by a
change in the representation of state vectors, according to the usual terminology of quantum mechanics.
It is convenient for us to choose N to be a transformation of this special kind. When this is done,
the division algebras C and Q will not in general
appear in the particular representations (25) and
(26). For the quantum-mechanical applications it is
useful to have i in the standard form (5), whereas
there is no strong reason to prefer the representations
(25), (26) of C and Q to other equivalent representations.
Let now G1 be a group composed of unitary operators only. For the moment we are not concerned
with the antiunitary part of G, and so we consider
the algebras A, B, X, Y only. Suppose that the
operators A(u) form an irreducible representation
of Glover C. Since C is the only division algebra
over C, the forms of the algebras Band Yare completely determined by Weyl's theorem
B = (0)",
Y=nC.
(39)
Equation (39) is in fact merely a statement of

Schur'S lemma. Ie Also, it follows from the definitions that every matrix in X which commutes with
i belongs to Y.
The order (number of linearly independent elements) of the algebra B is 2n2 According to Eqs. (8)
and (9), the order of A is 2n2 if i belongs to A,
n" if i does not belong to A. Weyl's theorem then
gives precisely three possible canonical forms for
the algebras A, X, as follows:
A
= 2R",
(40)
C,,'
A = Q... ,
= nC,
= mQ,
(41)
(42)
where we have written m = In. In all three cases

the operator i belongs to X, and the order of X
is 2 or 4.
Wigner's classification of irreducible representations A(u) is the following. Let A*(u) be the representation formed by taking the complex conjugate
in each element of A(u). If
A*(u)
MA(u)M-\
all u,
(43)
with M unitary and symmetric, then A(u) is

"potentially real." If Eq. (43) holds with M unitary
and antisymmetric, then A(u) is "pseudorea1." If
A*(u) is not equivalent to A(u), then A(u) is "complex." We write M as usual as a (2n X 2n) real
matrix, and define
P = jM
(44)
with j given by Eq. (4). Then Eq. (43) holds if and

only if the matrix P belongs to the commutator
algebra X. Therefore an equivalent statement of
Wigner's classification is this. If X contains atl antiunitary operator P with
p2
I,
(45)
then A(g) is potentially real. If X contains an antiunitary operator P with

p2
-I,
(46)
then A(g) is pseudoreal. If X contains no antiunitary operator P, then A(g) is complex. An inspection of the canonical forms (40), (41), (42) then
yields the following theorem.
Equivalence ThelYrem I. Let A(u) be an irreducible
repre8entation over C of a unitary group G 1 Let the
algebra A be defined by Eq. (8) with real coejfinients.
Then
(i) If A i8 of type R, A(u) i8 potentially real;
(li) If A i8 of type C, A(u) is oomplex;
(iii) If A i8 of type Q, A(u) i8 p8eudlYreal.
In eaoh case the oonver8e i8 also true.
The following remarks are corollaries of Theorem I.

Remark 1. The matrices M(u) form a real representation of the group G1 This representation is
irreducible over R when A is of type C or Q. It
reduces to two equivalent irreducible components
when A is of type R.
Remark 2. It is well known8 that A(u) is potentially
IP: 134.157.146.58 On: Sun, 21 Jun 2015 20:49:04
THE THREEFOLD WAY
real if and only if it is equivalent to a representation

composed entirely of real matrices. We now can
make another statement of the same kind. The
irreducible representation A(u) is pseudoreal if and
only if it is equivalent to a representation composed
entirely of matrices whose elements are quaternions
with real coefficients.
Remark S. It is well-known (see Wigner's book,t
p. 289) that the irreducible representations of the
3-dimensional rotation group are potentially real
for integer spin, pseudoreal for half-integer spin.
From remark 2 it then follows that the integerspin representations may be taken to be real, and
the half-integer-spin representations may be written
in terms of real quaternion matrices.
We now turn our attention to the full group G
including antiunitary operators. We shall be concerned with the algebras B, D and their commutators
Y, Z. An irreducible co-representation of G is a set
of matrices M(u) , M(a) such that the algebra D
is irreducible over R. According to Weyl's theorem
there are then three possibilities for the canonical
forms of D and Z.
1205
co-representation over C of a group G including antiunitary operations. Let the algebra D be. defined by
Eg. (10) with real ooefficients. Then
(i) If D is of type R, A(g) is of Wigner type I,
(ii) If D is of type C, A(g) is of Wigner type III,
(iii) If Dis of type Q, A(g) is of Wigner type II.
In each case the converse is also true.

Remark 4. If follows from this theorem that an
irreducible co-representation is of type II if and
only if it can be expressed in terms of matrices
whose elements are real quaternions.
Remark o. According to Eqs. (47) to (52), the
algebra Y has always precisely double the order
of the algebra Z. Also, it is known that Y contains
the matrix i, which commutes with Z but does not
belong to Z. Therefore, in the case here considered
(D being irreducible and Z a simple algebra), Y
is precisely the direct product of Z with the algebra
generated by (I, i).
Remark 6. The statement that Y is the direct
product of Z with (I, i) has been established for
the case of Z simple. However, by virtue of the
lemma of Sec. II (see remark 7 following the lemma)
(47)
Z = (2n)R,
the same relation between Y and Z holds in the
(48) general case.
Z = nC,
Remark 7. The lemma of Sec. II can be stated
(49)
Z = mQ, m = in.
D = Q""
very concisely as a statement about co-representaThe algebra B may now be reducible, but its tions: inequivalent irreducible co-representations of
irreducible components must be of the form sC ,. Also, G contain inequivalent irreducible representations
by Eqs. (9) and (10), the order of D must be exactly of GI
twice that of B. Equations (47), (48), and (49) then
IV. FURTHER ANALYSIS OF THE WIGNER
imply that the order of B is 2n2, n 2, in' in the three
CLASSIFICATION
cases. The only possibilities are
The equivalence Theorems I and II are so alike
(50)
Y = nC)
B = C.. ,
in form that one might suppose them to be two
(51) statements of the same triple alternative. We shall
Y= mC+ mC,
B = Cm. + C".,
show that in fact the precise opposite is true. The
(52)
Y = mC2 ,
B = 2C... ,
two triple alternatives are entirely independent.
and these correspond precisely to the three alterna- Within the same irreducible co-representation of G,
anyone of the three types of algebra D may occur
tives (47) to (49).
Wigner's classification of irreducible co-representa- in combination with anyone of the three types
tions1 is the following. The co-representation is type I of algebra A.
To study the relation between the two theorems,
if its unitary part is irreducible. It is type II if its
unitary part reduces to two equivalent irreducible we fix a particular irreducible co-representation A(g)
components. It is type III if its unitary part reduces of G and investigate the possible structure of the
to two inequivalent irreducible components. Now six algebras A, B, D, X, Y, Z in combination.
when the co-representation generates the algebra D, Since A(g) is irreducible, the possible structures
the unitary part of it generates the algebra B. An for D, Z, B, Yare described by Eqs. (47) to (52).
inspection of Eqs. (50) to (52) shows that these The representation A(u) of G1 is however not
three alternatives correspond to the Wigner types I, necessarily irreducible. Theorem I, and the three
alternatives given by Eqs. (40) to (42), apply
III, II, respectively.
Equivalence Theorem II. Let A(g) be an irreducible directly only to the irreducible components of A(u).
IP: 134.157.146.58 On: Sun, 21 Jun 2015 20:49:04
FREEMAN J. DYSON
1206
When D is of type R, then A(u) is irreducible

and Eqs. (40) to (42) apply unchanged. When D
is of type Q, then, according to Eq. (52), A(u) splits
into two identical irreducible components, or symbolically A(u) = 2A'(u). In this case Eqs. (40) to
(42) apply to A'(u). When D is of type C, then
Eq. (51) holds, and so A(u) splits into two irreducible
components inequivalent over C,
(53)
The real representation M(u) of Gl splits correspondingly into two components

M(u) = M 1(u)
+M
(u).
(54)
Equations (40) to (42) apply to Al and A2 separately.

However, we shall prove that the algebra A is
necessarily of the same type (R, C, or Q) for the
representations Al and A 2 Thus one of Eqs. (40)
to (42) applies to both components of A(u).
Let a be anyone of the antiunitary operators
in G. The transformation
tions of G into nine possible cases, which we denote

by RR, RC, RQ, CR, ... , QQ. Case CR, for example,
means that algebra D is of type C while algebra A
is of type R, Le., we have a co-representation of
Wigner type III whose unitary part splits into two
irreducible inequivalent representations each of
which is potentially real.
Using Eqs. (40) to (42) we can write down the
possible forms of the algebras A and X in each of
the nine cases:
caseRR,
A = 2R.. ,
X = nR2'
(61)
easeRC,
A = C.. ,
X =nC,
(62)
case RQ,
A = Qm,
X= mQ,
(63)
case CR,
A == 2Rm
+ 2R
m,
+ mR
X = mR2
2,
(64)
case CCI,
= Cm+ Cm ,
+ mC,
X == mC
(65)
case CC2,
A = 2Cm ,
(55)
case CQ,
A = Qp
is an automorphism V of the unitary group Gl

The representations
case QR,
A = 4R.. ,
X = mR4'
(68)
case QC,
A = 2C""
X = mCz ,
(69)
case QQ,
A = 2Qp,
X = pQ2,
(70)
u --+ V(u) = a -100
Ay(u) = A(V(u,
M y(u)
= M(V(u
(56)
+ Qp,
X = mC2,
X = pQ
(66)
+ pQ, (67)
differ from A(u) and M(u) only by a relabeling of

the elements of G1 Thus M y(u) and M(u) generate
isomorphic group algebras. Moreover, Eq. (54) implies
For convenience we wrote here m = tn, p = In.

The forms of A and X are uniquely fixed in all
cases except CC. Case CC divides into two alternatives CC1 and CC2. Case CC1 holds when the
M y(u) = [M(a)r 1 M(u)M(a) == M(u) ,
(57) representations M 1 (u) and M2(U) are inequivalent
over Rj case CC2 holds when Ml and M2 are equivawhere the equivalence is over R and not over C. lent over R.
Suppose now that D is of type C and Eq. (54) holds.
The results (61) to (70) follow immediately from
Then Eq. (57) means either
Eqs. (40) to (42) when D is of type R or Q. However,
when D is of type C some further argument is
M1y(u) = [M(a)]-IMl(u)M(a),
needed. Suppose then that D is of type C, so that
M2Y(U) = lM(a)rlMiu)M(a),
(58) Eqs. (48) and (51) hold, and the representation
M(u) splits according to Eq. (54). When Ml and M2
or
are inequivalent over R, every matrix commuting
Ml y(u) = [M(a)rlM2(u)M(a) ,
with the M(u) must commute separately with M1(u)
1
M2Y(U) = [M(a)r M 1 (u)M(a).
(59) and M2(U). The algebra X is then the direct sum
of the commutator algebras of Ml and M 2 ThereBecause the algebra D generated by M(u), M(a), fore for Ml and M2 inequivalent, Eq. (64), (65),
and i is irreducible, Eq. (58) cannot hold. There- or (67) holds according as A is of type R, C, or Q.
fore Eq. (59) must hold and
It remains to consider the case in which D is
of
type C while Ml and M2 are eauivalent over R.
(60)
There is then a real matrix L which commutes with
The algebra A generated by MICU) is therefore all the M(u) but does not commute with Ml(U),
necessarily of the same type as that generated M (u) separately. This L satisfies
2
by M2(U).
(71)
We may thus classify irreducible co-representa- M1(u) = L- 1M 2 (u)L, Miu) = L- 1M 1 (u)L.
IP: 134.157.146.58 On: Sun, 21 Jun 2015 20:49:04
THE THREEFOLD WAY
Since A 1 (u) and A2 (u) are inequivalent over C, L

must anticommute with i. Now suppose if possible
that A were of type R or Q. Then there would exist
also a matrix L' in X, anticommuting with i and
commuting with each of M 1 (u), M 2 (u) separately.
The product U = LL' would be a matrix commuting with i and also satisfying Eq. (71). This
is impossible since Al and A2 are inequivalent over C.
We have thus proved that, if D is of type C and
M 1 and M 2 are equivalent, A is also necessarily
of type C. There exists then only the case CC2
with A and X given by Eq. (66).
We next discuss a special situation in which the
above enumeration of possibilities simplifies considerably. We say that the group a is "factorizable"
if the automorphism V given by Eq. (55) is an inner
automorphism of a1 Suppose that a is factorizable.
Then there exists an element w in a1 such that
V(u)
a-Iua = w-Iuw, all u in
(72)
1207
The operator [M(T)J2 commutes with all M(g)

and with i, and it is also equal to M(u) with u = T2.
Thus [M(T)J2 belongs to both the algebras D and Z.
By Eqs. (47) to (49), the common part of D and Z
is (2n)R when D is of type R or Q, and is nC when
D is of type C. Since [M(TW is a real orthogonal
matrix, it must be a scalar
[M(T)]2 =
= 1,
(75)
in any of the four cases RR, RQ, QR, QQ. However,

Eq. (75) need not hold in case CC2.
We determine lastly which cases go with the plus
sign and which with the minus sign in Eq. (75).
When
is factorizable and Eq. (75) holds, the
algebra D is a direct product of the commuting
algebras A and W, where W is the algebra of order 4
generated by [I, i, M(T), iM(T)]. The structure
of W = (D / A) is then determined as follows:
caseRR,
A = 2Rn,
D = R 2n , W '" R 2 ,
case RQ,
A = Qm,
D = R 2n , W",
(73)
case QR,
in a which commutes with all elements of a1

Conversely, if such T exists, then V (u) is an inner
automorphism for any choice of the antiunitary
operator a in Eq. (55). In many physical applications, when such an operator T exists it is convenient to give it the name "time-inversion operator." In any representation M(g) of a, the antiunitary matrix M(T) belongs to the algebra X.
We now classify the possible types of irreducible
co-representation of a factorizable group a. Many
cases can be immediately eliminated. First, the
matrix M(T) belongs to X but does not belong to
Y since it anticommutes with i. Therefore X Y
for a factorizable group. Hence, by comparing Eqs.
(50) to (52) with Eqs. (62), (65), and (69), the
cases RC, CC1, and QC are excluded. Next, suppose
that D is of type C. Then Eq. (72) gives
case QQ,
Then there exists an antiunitary operator
M v(u)
[M(w)r1M(u)M(w),
(74)
with M(w) unitary. Since A1(u) and A2 (u) are inequivalent over C, Eq. (74) implies
M (u) = [M(w)r 1M (u)M(w),
1V
M 2V (u) = [M(w)r 1M 2 (u)M(w).

This together with Eq. (60) shows that M1(u) and
M 2 (u) are equivalent. We proved earlier that cases
CR, CC1, and CQ are then impossible.
The surviving cases for a factorizable group
are RR, RQ, QR, QQ, and CC2.
(76)
Q,
(77)
A = 4R m, D = Qm,
W,,-, Q,
(78)
A = 2Q",
W "-' R 2.
(79)
D = Qm,
The sign of E in Eq. (75) is plus when W is of type R 2 ,

minus when W is of type Q. These results will now
be summarized in a theorem.
Theorem III. Let M(g) be an irreducible co-representation of a factorizable group a, in which M(T)
is anti-unitary and commutes with all the M(g). Then
the following three possibilities alone exist:
(i) case RR or QQ with [M(T)]2 = +1,
(li) case RQ or QR with [M(T)]2 = -1,
(iii) case CC2 with [M(T)]2 = cos a + e sin a,
where a may be any real angle, and e is an element
of the algebra A with e2 = -1.
Remark 1. It is noteworthy that the sign of
[M(T)]2 is determined neither by the Wigner type
of the co-representation M(g), nor by the reality
type of the unitary subrepresentation M(u) , but
only by these two types in combination. Thus
[M(TW = +1 corresponds to Wigner type I and
potentially real, or to Wigner type II and pseudoreal; [M(T)1 2 = -1 corresponds to Wigner type II
and potentially real, or to Wigner type I and
pseudoreal.
Remark 2. In the majority of applications of the
theorem, T will be identified with the physical
operation of time inversion. In these circumstances
[M(T)]2 = +1 for co-representations with integer
spin, and [M(T)]2 = -1 for co-representations with
IP: 134.157.146.58 On: Sun, 21 Jun 2015 20:49:04
FREEMAN J. DYSON
1208
half-integer spin.7 Therefore cases RR and QQ occur

only with integer spin, cases RQ and QR only with
half-integer spin. Case CC2 may occur with either
integer or half-integer spin.
V. EXAMPLES
The classification theory of Sees. III and IV

would be empty if one could not produce examples
to show that each of the enumerated possibilities
can actually occur. We list here one example of
each of the ten possibilities (61) to (70). The first
five examples are factorizable and illustrate Theorem
III. The last five are nonfactorizable.
To simplify the notations we write (2 X 2)
matrices in terms of the standard basis
12 =
el =
e2 =
[01 -lJ,
0
ea =
[10 -1OJ. (SO)
The quaternion units are defined by Eqs. (26) and

(27). The order of the co-representation is the order
of the real matrices M(g); this is twice the dimension of the complex vector space He.
Example 1. Case RR. Order 2.
GI contains identity I only. G = [I, T], T2 = I.
MW = e2,
M(T)
= ea,
Case QR. Order 4.

GI generated by [I, T2], G
with T' = I.
+ sin cp[ea
X e2],
M(t}
M(T)
[M(T)]2
=
=
cos !a[el X ea]

cos a[I2 X 12 ]
M(cp)
(cos cp)I2 + (sin
= e2,
M(T) = ea.
Example 7. Case QC. Order 4.

Same group as example 6.
M(cp)
(cos cp)I,
+ (sin cp)T2'
M(T)
Tl'
Example 8. Case CR. Order 4.

G1 is a 4-element group generated },y the
reflections Rz and R. in two perpendicular
planes. G = [GIr TGIJ, T2 = I, where T is a
combination of time-inversion with a reflection
in the plane x = y.
M(T)
TI,
T{,
[M(T)]
2=
= ea X 12 ,
M(n, cp) = exp [lqm. fJ X 12 ,

M(']') = T{ X e2,
X 12 ,
-ea X 12 ,
= el X ea.
Example 9. Case CQ. Order 8.

GI is a direct product lOa X Oa] of two 3-dimensional rotation groups. G = [Gt, TG 1 ],
where T interchanges the two groups.
M(m, cp; n, If)
= [t(I2 + ea)J X
+ [l(I2 -
-I,.
Example 4. Case QQ. Order 8.

Same group as example 3. Representation M(u)
has two spin-l components which are interchanged by the T operator.
12 X e2'
Example 6. Case RC. Order 2.

G is generated by the 2-dimensional rotation
group O2 with an operator T not commuting
with O2 , T is now time-inversion without space
reflection.
M(n,cp) = exp[!cpnfJ,
+ sin 1a[e2 X el],

+ sin area X e2] = M(a).
M(R,,)
Example 3. Case RQ. Order 4.

group 0., with the time-inversion operator T
commuting with Os. Representation M(u) is
with spin 1.
M(T)
M(cp) = cos cp[I2 X 12J
T2,
M(T)
Example 5. Case CC2. Order 4.

group O2 together with an operator T commuting with O2 , The operator T is a combination of time-inversion with space reflection. The
phase angle a is a fixed parameter.
exp [#m TJ
ea)J X exp [!1/tfi'T],
12 X
Example 10. Case CCI. Order 4.

G1 = [02 X O2 ], G = [G l , TG I ] where T interchanges the two O2 groups.
M(cp, If) = [l(I 2 + ea)] X [cos 1/>12
+ [!(I
ea)] X
+ sin cpe2]
[cos #2 + sin Ife
2 ],
IP: 134.157.146.58 On: Sun, 21 Jun 2015 20:49:04
THE THREEFOLD WAY
The most interesting of these examples are numbers 4, 5, 8. They have some features which are
nontrivial and appear to be novel. We leave to the
reader the exercise of verifying that in each case
the commutator algebras X, Y, Z have the structure
described in Eqs. (47) to (52), (61) to (70).
VI. ALGEBRAIC CHARACTERIZATION OF
REPRESENTATION TYPES
In this section we conclude the study of representation types by proving a generalized version
of a classical theorem of Frobenius and Schur. Let
M(g) be a representation of a group G, irreducible
over some ground field cP with characteristic zero.
We suppose that the group G is either finite or
compact, and that the matrices M(g) have a finite
order d. If f(g) is any function of the group element
g, the average of f(g) over G is defined by
avo f(g) = h-1
L..J
f(g) ,
(81)
1209
To prove the theorem, let g' be any element of G;

Then
P H kl = avo [M.;(g,-lg-I)M kZ (gg')]
=
L..... M ....(g,-I)P
m; .k..Mnl(g')
(85)
.
Thus Pij .u, considered as a matrix in the indices

(i, l), commutes with all M(g') and belongs to the
commutator algebra L. Similarly, PH .1:1 belongs to
L when considered as a matrix in (k, j). Let el',
p. = 1, ... , e, be a linearly independent basis for
the algebra E. Then Eq. (84) gives
(86)
where the 0,., are coefficients in CPo
Now consider the sum
(87)
On the one hand, by Eq. (86),

}.
or by
avo f(g) = V-I
Jf(g) dp.(g) ,
(L);I:
(82)
where h is the order of G when G is finite, and where

is the volume of G in the invariant group measure
dp.(g) , when G is compact. We consider the fourthrank !tensor
= avo [MH(g-l)Mkl(g)].
(83)
Let K be the group algebra generated by the M(g)

with coefficients in CPo
The structure of K is given by Eq. (24), since
Weyl's theorem holds in any field with characteristic zero. The commutator algebra L of K has
the structure
I, X E,
(It),,; X Ls}.,.c,.,(e'),,;,
where
s}.,.
11
P H kl
(84)
where E is an irreducible division algebra of order

e over CP, and d = teo
The type of the representation M(g) is specified
by the division algebra E. For example, when cP = R
is the field of real numbers, there are three types
of representation corresponding to E = R, C, or Q.
The tensor PH .kl is useful in classifying representations by virtue of the following theorem.
Theorem IV. The tensor PH.J:I depends only on
the integer t and on the algebra E, and is otherwise
independent of the group G and of the representation
M(g).
Thus PH .kl is characteristic of the type of the
representation M(g).
,.,
= spur [eV].
(88)
(89)
On the other hand, by Eq. (83), since e}. commutes

with all the M(g),
}.
(L);I:
= avo [M(g)(I, X i)M(g-I)]ki

= (It X ih;.
(90)
Comparison of Eqs. (88) and (90) shows that

(91)
so that the matrix 0,., is the invel'S(;l of the matrix s}.,..

The coefficients 0,., are thus uniquely determined
by E, and Eq. (86) establishes the truth of Theorem IV.
We shall be interested in applying Theorem IV
to cases in which the matrices M(g) are orthogonal.
So we assume
(92)
The algebra K then contains the transposed of

every matrix in K, and L has the same property.
We can therefore choose the basis elements e>' of
the algebra E to be either symmetric or antisymmetric. Suppose that the number of symmetric i
is q, and the number of antisymmetric i is q'.
The invariant
(93)
IP: 134.157.146.58 On: Sun, 21 Jun 2015 20:49:04
1210
FREEMAN J. DYSON
provides a simple criterion for the type of the

representation M(g).
Theorem V. When M (g) is an orthogonal irreducible
representation of G over a field 4.> of characteristic zero,
= avo spur [M(l)] =
q - q'.
(94)
The first part of Eq. (94) follows at once from the

definition of P and Eq. (92). To prove the second
part, we suppose the e>' chosen so that
(95)
with each '1/>. equal to 1. Then Eqs. (86), (89),
and (91) give
Remark 3. We apply remark 1 to the situation

discussed in Theorem II of Sec. III. Let A(g) be
an irreducible co-representation over C of a group G.
According to remark 6 of Sec. II, the group algebra
D is generated over R not by the group G itself
but by an extended group r. The representation
of r which generates D consists of the matrices
M(u) ,
p.llii
(96)
= q - q'.
Remark 1. Suppose that 4.> is the field of real

numbers. Then Theorem V gives the following
characterization of the type of the representation
M(g):
M(a) ,
iM(a) ,
(105)
which are all real and orthogonal. When Theorem V

is applied to the group r, the contributions from
M(u), iM(u) to P cancel each other, while the
contributions from M(a) , iM(a) are equal. Thus
P = L cl'p(el');;'I/.(eP);i
L'I/.
iM(u) ,
= t avo spur
[M(a 2 )],
(106)
averaged over the antiunitary part only of G. If

A(u) is the unitary part of the co-representation,
A(u) is irreducible when D is of type R, while A(u)
has two irreducible components when D is of type
Cor Q. In any case we let A'(u) be one of the (one
or two) irreducible components of A(u) , and we write
II' = avo [spur A'(a
)].
(107)
By Eq. (100), this II' is equal to P when D is of

type R, and is equal to tP when D is of type Cor Q.
The criterion of Eqs. (97)-(99) then becomes
+1 for M(g) of type R,
P =
0 for M(g) of type C,
(97)
(98)
P = -2 for M(g) of type Q.
(99)
II' =
Remark 2. We apply remark 1 to the situation

discussed in Theorem I of Sec. III. Let A(u) be an
II' =
+1 for A(g) of Wigner type I,

0 for A(g) of Wigner type III,
(108)
(109)
II' = -1 for A(g) of Wigner type II.

(110)
irreducible representation over C of a unitary group
This elegant analog to the Frobenius-Schur criterion
GI Then the corresponding real representation M(u)
was discovered by Bargmann. ls
splits into two equivalent irreducible representations
M'(u) when M(u) is of type R, while M(u) is irVII. THEORY OF MATRIX ENSEMBLES
reducible when it is of type C or Q. The correIn this section we deal with the problem for which
spondence between A(u) and M(u) gives
the theory of Sec. II was specifically introduced,
(100) namely the classification of ensembles of matrices
spur M(u) = 2 Re spur A(u),
with given symmetry properties. An ensemble is
and therefore the quantity
a set of objects with an assigned probability dis2
II = avu [spur A(u )]
(101) tribution. We shall define the probability distributions later; it is necessary first of all to study the
is equal to tP. This P is given by Eqs. (98), (99)
classification of sets of matrices invariant under
when M(u) is of type C, Q, but is equal to (+2)
some symmetry-group G.
when M(u) is of type R since Eq. (97) then refers
As in Sec. II, we suppose that the matrices S
to the irreducible component M'(u). So we derive
which we are studying operate in a complex vector
the classical criterion of Frobenius and Schur17 for
space Hoof finite dimension n. We are given a
the type of an irreducible unitary representation:
representation of the group G in H 0' consisting of
II = +1 for A(u) potentially real,
(102) unitary operators A(u) and antiunitary operators
A(a). The matrices S are supposed to be invariant
II =
0 for A(u) complex,
(103)
under G, but this notion of invariance already introII = -1 for A(u) pseudoreal.
(104) duces an ambiguity. There is a choice between two
---G. Frobenius and I. Schur, reference 6.

17
18
V. Bargmann (private communication).
IP: 134.157.146.58 On: Sun, 21 Jun 2015 20:49:04
THE THREEFOLD WAY
definitions of invariance. We say that 8 is "formally

invariant under G" if
8A(g) = A(g) 8,
all
in G.
(111)
Formal invariance means that 8 is unchanged by

any of the transformations
8 - A(g)S[A(g)r1,
(112)
whether g be unitary or antiunitary. We say that S

is "physically invariant under G" if for every pair
of vectors (cp, 1/;) in He
(cp, S1/;) = (A(u)cp, 8A(u) 1/;) = (A(a) 1/;, SA(a)cp). (113)
Note that the initial and final state vectors are

interchanged in Eq. (113) in the case of antiunitary
elements of G. The effect of Eq. (113) is that we
have instead of Eq. (111)
8A(u)
A(u) 8,
8A(a)
A(a) 8+ ,
(114)
where S+ means the Hermitian conjugate of S.

The two types of invariance are relevant in different circumstances. If S is, for example, a unitary
operator describing a change in the representation
of states, then formal invariance under G is a
meanmgful requirement, signifying that this change
in representation does not disturb the symmetry
relations of the states under the operations of G.
If 8 is an operator characterizing a physical system,
for example a scattering matrix, then the antiunitary operations of G are associated with a reversal of the physical roles of initial and final states;
in this case physical invariance of S is the physically
meaningful requirement, signifying that the system
to which 8 belongs is invariant under the operations
of G in the usual dynamical sense. The two definitions
of invariance under G become equivalent only when
the matrix S is Hermitian, for example when S is
the Hamiltonian of a system.
It is convenient to transcribe the matrix S into
a real (2n X 2n) matrix operating in the real vector
space H R according to Eq. (2). The real form of S
then satisfies
(115)
Si = is,
with i defined by Eq. (5). For 8 to be invariant

under the unitary subgroup G1 (in either sense) it
is necessary and sufficient that
8M(u) = M(u)S,
(116)
where the matrices M(g) are the representation of

G defined in Sec. II. The condition for 8 to be
formally invariant under G is
SM(a)
M(a)S,
a in G,
(117)
1211
in addition to Eq. (116). The condition for S to be

physically invariant under G is Eq. (116) and
SM(a) = M(a)ST,
in G,
(118)
where ST means the transpose of 8.

From Eqs. (115), (116) we see that the set Y
of matrices in H R invariant under G1 is identical
with the commutator algebra Y defined by Eqs. (9)
and (12). From Eqs. (115)-(117), the set Z of
matrices formally invariant under G is identical
with the commutator algebra Z defined by Eqs.
(10) and (13). We define the set W to consist of
those matrices which are both formally and physically invariant under G. Then W is the set of all
symmetric matrices in Z. Lastly, we define V to be
the set of matrices physically invariant under G.
Then we shall prove
Theorem VI. For S to be in V, it is neoessary and
sufficient that
(119)
where S1 and S2 are matrices in W.
The sufficiency follows immediately from the

relations
iM(a) = -M(a)i.
(120)
To prove the necessity, we observe that all the

matrices M(g) are orthogonal, and thus
(121)
Hence ST belongs to V whenever S does, and we

may then write
S = Sf
+ S",
(122)
where S' is symmetric and S" antisymmetric, and

both Sf, S" belong to V. The matrices S1 = Sf
and 8 2 = -is'' now satisfy both Eq. (117) and
Eq. (118), and therefore belong to W.
The results of Sec. II, and in particular Weyl's
theorem, provide us with a complete structural
analysis of the sets V, W, Y, Z. We use Frobenius'
theorem (Sec. I) in order to replace the division
algebras E; of Weyl's theorem by the standard
trio R, C, and Q. The integers tj of Weyl's theorem
are now irrelevant since they contribute to the
structure of the group algebra D but not to the
commutator algebra Z. We thus state the main
result of the theory of matrix ensembles as follows.
Theorem VII. The set Z of matrioes in H R formally
invariant under G is a direot product of irreducible
oomponents, one component Zj corresponding to eaoh
inequivalent irreducible ao-representation of G oontained in the given oo-representation A(g). Each oom-
IP: 134.157.146.58 On: Sun, 21 Jun 2015 20:49:04
1212
FREEMAN J. DYSON
ponent Z I may be written as the set of all square

matrices of order s/ with elements 'in an algebra
Eooh
is either R, the algebra of real numbers, or 0,
the algebra of oomplex numbers, or Q, the algebra of
reat quaternions.
Remark 1. The structure of the set Z i depends

on the Wigner type of the corresponding co-representation of G in the manner specified by equivalence
Theorem II. The reality type of the unitary part
of the representation, specified by equivalence
Theorem I, is here entirely irrelevant, except insofar as the Wigner type and the reality type may be
correlated for factorizable groups G according to
Theorem III.
According to remark 7 at the end of Sec. II, the
sets V, W, Y are direct products of independent
components, one corresponding to each component
Zj of Z. To avoid unnecessary repetition, we describe
the structure of V, W, Y corresponding to a single
component of Z. Thus in the following theorems
we assume that Z is irreducible, which means that
all irreducible co-representations contained in A(g)
are equivalent. From this special case the general
case is easily derived by writing Z j for Z and taking
a direct product over j.
We have seen, in Remarks 5 and 6 of Sec. III,
that the algebra Y is generated by i and Z. The
matrix i commutes with Z, and therefore commutes
with the algebras
Hence we may form a new
algebra cfli by adding the independent unit i to cfl j
Theorem VIII. When Z is irreducible, the set Y of
matrices in HB invariant under G1 may be written
as the set of all square matrices of order s with elements in an algebra cfl', derived from
by allowing
each element of cfl to have complex instead of real
coeffidents.
Remark S. When cfl = R, cfl' is the algebra of ordi= 0, cfl" is the
nary complex numbers. When
algebra of complex-complex numbers with two commuting imaginary units; in this case cfl" is reducible
and has the stl'llcture
+ O.
(123)
When
= Q, is the algebra of complex quaternions, which is equivalent to an algebra of complex
(2 X 2) matrices,
(124)
follows. We define the conjugate of a number in

to be the number obtained by reversing the signs
== 0) or of
of the coefficient of e2 (in the case
the coefficients of 'TlJ 1'2, 'Ta (in the case
= Q).
We define the dual of a matrix to be the transposed matrix with each element conjugated. Since
the units e2, 'T1, 'TI, 'Ta when written in expanded form
are antisymmetric, a matrix which is symmetric
in expanded form becomes self-dual when written
with elements in
Theorem IX. When Z is
the set W of
matrices in H B invariant under G in both physical
and formal senses may be written as the set of all
square self-dual matrices of order s with elements
in = R, 0, or Q.
The conjugate of an element of

is obtained
by changing the signs of the coefficients of the
units, leaving the unit i unchanged. So from Theorems VI and IX follows immediately the result:
Theorem X. When Z is irreducible, the set V of
matrices in H 11 invariant under G in the physical
sense may be written as the set of all square self-dual
matrices of order s with elements in
Remark 3. We now finally make contact with the

theory of matrix ensembles developed earlier by
the author.' Let V u be the subset of unitary matrices
in V. Then Theorem X states that, for the most
general symmetry group G and the most general
quantum-mechanical representation of G, the set
V u is a direct product of independent components,
each of which is identical with one of the three
ensemble-spaces T I , T 2 , and T, defined in reference 4.
The cases T I , T" T. correspond, respectively, to
= R, 0, Q. The spaces T I , T 2 , and T. were
originally obtained by considering special groups G
of a very simple kind. It is satisfactory to find
that the same three spaces, and no others, occur
in all possible circumstances.
The reason for choosing V u as the space in which
to construct an ensemble is that no natural definition of uniform probability appears to exist in V.
For the same reason we study the subset Z u of
unitary matrices in Z. The following theoremfollows
from Theorem VII together with well-known properties of the classical groups.'
Theorem XI. The set Zu of unitary matrices in HJf
formally invariant under G is a direct product of irreducible components, each of which is a simple
The algebra W consists of matrices which are classical group. When

= R, 0, or Q. the corresymmetric when written in expanded form in H B' sponding component of Z u is an orthogonal, unitary,
When S is written, as in Theorems VII and VIII, or symplectic group of dimension Si'
as a smaller matrix with elements in
condition
In the same way we define the unitary subset
of symmetry becomes a condition of
duality, as Y u of Y. The components of Y u are
IP: 134.157.146.58 On: Sun, 21 Jun 2015 20:49:04
THE THREEFOLD WAY
Yu
U(s) ,
U(2s) ,
U(S) X U(s) ,
(125)
Zu = O(s),
U(s) ,
Sp(2s).
(126)
The unitary space V u is not a group. But it can be

represented conveniently in terms of the groups
Y u , Zu in the following way. A matrix S belongs
to V u if and only if it can be expressed as a
symmetric product
S = UUD ,
in
Yu,
(127)
where D denotes cP dual. All matrices U' of the form
U' = UU l ,
of V u is of the form
S = exp [iH],
corresponding to
(128)
1213
in W,
(131)
while a matrix S of Zu is of lhe form

S = exp [A],
in ZA,
(132)
where ZA is the subset of Z containing anti-Hermitian

matrices. Thus Wand ZA are the spaces of infinitesimal generators for V u and Zu, respectively.
We define the Hermitian Gaussian ensemble EE
as the space W of matrices H with the probability
distribution
dp.(H) = C exp [ - (spur H2) /4a2 ]
II dH
(133)
where c, a are constants and the product extends

correspond to the same S by Eq. (127), and every over all the independent real coefficients of the
U' corresponding to S is of the form (128). Thus elements of H in the algebra CPo The anti-Hermitian
each matrix S in V u corresponds to a unique co-set Gaussian ensemble EE is defined as the space Z A
of the subgroup Zu in the group Yu. We have thus with probability distribution
proved
2
(134)
Theorem XII. The set V u of unitary matrices in V dp.(A) = C exp [+(spur A2)/4a ] II
is abstractly equivalent to the homogeneous space These ensembles have an algebraic structure pre(Y u/Zu) , the quotient of the group Y u by its sub- cisely analogous to that of E P and E P , respectively.
group Zu.
They divide into irreducible components each of
Having defined the spaces Z u and V u, we are which is of one of the three types R, C, or Q. In
now in a position to define the corresponding in- particular, EE is the natural ensemble to use in
variant matrix ensembles. The ensemble E P of describing the statistical properties of the Hamilunitary matrices formally invariant under G is tonian H of a system known to be physically indefined as the space Z u with probability distribution variant under the group G.
given by the invariant group measure in Zu. Since
The physical motivation for considering ensembles
Z u is a direct product of simple classical groups, of matrices with probability distributions defined
the group measure in Z u is merely the product of in these various ways has been discussed by Wigner l9
the invariant measures in the irreducible com- and by the author.4 In the case of the ensembles
ponents of Z u. The ensemble E P of unitary matrices E P and E P , consisting of unitary matrices, the
physically invariant under G is defined as the space existence of a natural uniform measure provides
V u with measure given according to Theorem XII by an intuitively plausible definition of "equal a priori
A
E
(129) probability." In the case of the ensemblesE andE ,
consisting of Hermitian and anti-Hermitian matrices,
Here dp.(Y u) and dp.(Zu) are the invariant group the choice of a Gaussian probability distribution is
measures in Y u and Zu, and the quotient measure mainly a matter of mathematical convenience.
is defined in the obvious way. Alternatively, the
has argued that one should use in
quotient measure may be uniquely defined as the preference to Eq. (133) a "microcanonical ensemble"
measure in V u which is invariant under all auto- with the exponential replaced by a delta function
morphisms
t5[spur (H2) - 0'2].
USUD ,
(130)
U in
The algebraic structure of EE and EA would of
P
of V u into itself. The ensemble E is a direct product course not be affected by such a change.
In any physical situation to which the ensembles
of irreducible components, each of which is identical
P
E
or EE are relevant, we have a system specified
with one of the three types E l , E 2 , E. which were
by a unitary operator S or by a Hermitian H.
studied in reference 4.
Two other types of ensemble naturally suggest
lt E. P. Wigner, Proceeding8 of the 4th Canadian MatheCongres8 (University of Toronto Press, Toronto,
themselves for study, composed of Hermitian and matic8
Canada, 1959), p. 174.
anti-Hermitian matrices, respectively. A matrix S
10 N. Rosenzweig, Bull. Am. Phys. Soc. 7, 91 (1962).
IP: 134.157.146.58 On: Sun, 21 Jun 2015 20:49:04
1214
FREEMAN J. DYSON
Since the system is invariant under G, every p(el, ... , 8,.) = e II (1 - cos e;)
;
stationary state is associated with a particular irreducible co-representation of G. Each irreducible
II
[cos ei - cos
(137)
i<j
co-representation fixes the values of a certain set
of quantum numbers (spin, parity, isotopic spin, (iv) s = 2n + 1, ..:l = -1, eigenvalues -1,
etc.) which are attached to the energy levels belong- exp ( iej ),
ing to that co-representation. The fact that the
(1 + cos ej )
ensemble E P or EH is a direct product of irreducible peel' '" , en) = e II
i
components means that the energy levels belonging
II
[cos ei - cos B;]2.
(l38)
to different sets of quantum numbers are stai.<i
tistically uncorrelated. Thus the statistical proper((3) cP = c, Zu = U(s), eigenvalues exp (iB j ).
ties of energy levels are entirely determined by the
behavior of the individual level-series, each asso- PCBI' '" , e,) = c
lexp Cie i ) - exp (i8 j ) 12. (139)
i<i
ciated with one set of quantum numbers. A single
level-series is described by an irreducible ensemble.
C'Y) cP = Q, ZU = Sp(2s), eigenvalues exp ( iB;).
The final result of our analysis may then be stated
PCBI , , B,) = c II (1 - cos2 B;)
as follows: When we consider a single series of
I
energy levels of a complex system, having definite
II
[cos ei - COSB;]2.
(140)
values for all quantum numbers of the symmetry;<i
group G, the statistical behavior of these levels
13. EA. Gaussian Ensemble of Anti-Hermitian
follows one of three possible laws, corresponding to
P
Matrices Formally Invariant under G
the three types of irreducible ensemble E or EH
Ca) cP = R, matrices real and antisymmetric.
(i) s = 2n, eigenvalues iE;,
II
vm.
EIGENVALUE DISTRIBUTIONS
peEl, ... ,E,.) = e[n (E; - E;)2]

i<i
In this section we list without proof the joint
probability distributions of the eigenvalues of
X exp [matrices belonging to the irreducible ensembles
F
P
E , EA, E , EH. In each case the integer s is the (ii) 8 = 2n + 1, eigenvalues 0, iE;,
dimension of the algebra Z over the field P which
may be R, C, or Q. The constant c will not be the peEl, ... ,E,.) = e[Il
i
i<i
same each time it appears.
(141)
(142)
1. EF. Ensemble of Unitary Matrices
((3) cP
Formally Invariant under G
peel, .. , ,en)
II
[cos e
i<i
i -
cos e;]2.
II (1
i<i
+ 1, ..:l =
1, eigenvalues
peEl,
(143)
= Q. Eigenvalues iE
.. , ,E.) = e[Il EmIl
('Y) cP
1<1
X exp [ -
L:;
(144)
3. E P Ensemble of Unitary Matrices

Physically Invariant under G
Eigenvalues exp (ie;), each doubly degenerate

in the case P = Q.
- cos2 8 j )
cos
- EYJ
i<i
(135)
x II [cos eo -
e[II (Ei
X exp [-
1, eigenvalues exp (iej ),
, en-I) = e
(iii) s = 2n
(ii) s = 2n, ..:l = -1, eigenvalues 1, exp ( ie;),
pel, '"
C. Eigenvalues iE j
peEl, ... ,E.)
(a) P = R, Zu = O(s). In this case Zu (the

orthogonal group) splits into two disconnected parts,
consisting of matrices with determinant ..:l equal
to + 1 and -1, respectively. There are thus four
distinct eigenvalue distributions to be listed.
(i) s = 2n, ..:l
elY'
(136)
+1, exp ( ie;),
p(el , " ' , e.) = c
II
/exp (ie;)
i<i
with f3 = 1, 2, 4 for cP
- exp
(145)
R, C, Q, respectively.
IP: 134.157.146.58 On: Sun, 21 Jun 2015 20:49:04
THE THREEFOLD WAY
4. EN. Gaussian Ensemble of Hermitian

Matrices Invariant (in either sense) under G
Eigenvalues E j , each doubly degenerate in the
case
= Q.
PCEI, ... ,E.)
= c[II
i<i
E;/4a
] ,
(146)
with {3 = 1, 2, 4 for
= R, C, Q.
Proofs of Eqs. (135) to (140) are to be found in
Chapter 7 of Weyl's book. 2 Equations (141) to (144)
can be deduced as limiting cases of Eqs. (135) to
(140) when all angles 6; are small. Similarly Eq. (146)
can be deduced from Eq. (145). The proof of Eq.
(145) has been given by the author.4
The statistical properties of the eigenvalues resulting from each of these ensembles can be studied by
following the method used by the author4 for the
case of Eq. (145). The eigenvalue distribution in
each ensemble has an exact mathematical analog
in the form of a classical Coulomb gas.
We briefly describe the Coulomb gas analogs to
EF and EA when
= R or Q. In EF the numbers
Xj
cos 6j
(147)
are considered to be positions of unit charges, constrained to move on the segment [-1
x
+1],
which may be imagined to be a straight conducting
wire of length 2. Every two charges repel each
other with the potential
W(X; -
Xj)
= -
In
Ix; - xd.
(148)
In addition there are fixed charges of q+ units at

x = +1 and of q_ units at x = -1. When
= R,
the angles 6j are rotation angles of a random rotation
in the orthogonal group O(s). The values of q+,
q_ are
(i)
= 2n,.1
(ii)
= 2n, .1 =
(iii)
s = 2n
(iv)
s =
l;q+ = q- =
+ 1, .1
2n + 1,.1
-1.
+1.
= 1; q+ = +1, q- = -1.
= -1; q+ = -1, q- = +1.
-1; q+
gas is the same for

= R or Q, namely, T = !.
The Gaussian antisymmetric ensembles EA. for
= R or Q have a Coulomb analog composed of
unit charges with positions
Xj
IE; - Ed ll]
X exp [ -
1215
q- =
When
= Q, the angles 6j are rotation angles of
a random matrix in the symplectic group Sp(2s).
In this case q+ = q_ = +1. The temperature of the
E;,
(149)
constrained to move on the semi-infinite straight

wire 0
x < ro. The repulsion between charges
is again given by Eq. (148), and T = ! as before.
There is a fixed charge of q units at x = 0, where
q
= -1
when
= R,
2n,
q = +1 when
= R, s = 2n + 1, or when cI> = Q.
In addition to the Coulomb forces, each charge Xj
is subject to a constant downward force produced
by a "gravitational potential"
(150)
When cI> = C, the ensembles EF and EA become

identical with E P and E H , for which the Coulomb
analogs have been described previously.4
The whole of the previous analysis4 of level distributions, based on the ensembles E P , can be repeated with minor modifications for the other
ensembles E F , EA, EH. However, there is one basic
difference between the physical ensembles E P , EH
on the one hand and the formal ensembles E F , EA
on the other.
Theorem XIII. Consider an irreducible ensemble
of matrices over the field cI>, with order s -+ ro. In
EF or EA, the local statistical behavior of eigenvalues
is described by an infinite Coulomb gas with temperature T = ! independent of cI>. In E P or EN the local
behavior of eigenvalues is described by an infinite
Coulomb gas with temperature T = 1, !, 1 corresponding to cI> = R, C, Q.
The most striking qualitative feature of the
physical ensembles E P , EN is that the strength of
the repulsion between neighboring energy levels
depends on the Wigner type of the co-representation
to which these levels belong. This feature is absent
in the formal ensembles E F , EA.
ACKNOWLEDGMENTS
The author is grateful above all to Dr. V. Bargmann for suggesting to him this line of inquiry, and
for help on specific points as noted in the text. He
also profited greatly from conversations with Dr. E.
Wigner and Dr. A. Grossmann.
IP: 134.157.146.58 On: Sun, 21 Jun 2015 20:49:04

The Threefold Way - Dyson (1961)

Uploaded by

Copyright:

Available Formats

The Threefold Way - Dyson (1961)

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

The Threefold Way - Dyson (1961)

Uploaded by

Copyright:

Available Formats

JOURNAL OF MATHEMATICAL PHYSICS

The Threefold Way.

Imtitute for Advanced Study, Princeton, New Jersey

HE purpose of this paper is to bring together

In each of the three theories which we aim to unify,

Once this is understood, the further development

formalism is brought into closer correspondence with

The starting point of our analysis is a group G

the A(a) by true matrices over the field R of real

A(g) +-+ M(g)

in the following way. M(g) is a [2n X 2n] matrix

a matric algebra over R. In particular we define

The inclusion relations

Each A(a) is of the form

where U(a) is unitary and j is the operation of

The space H R in which the matrices M(g) operate

also as a matrix operating in HR'

All the M (g) are orthogonal matrices.

B generated by the M(u) and i,

D generated by the M(u), M(a), and i.

The commutator algebra of a given algebra K is

The algebra A is given the name "group algebra

with MinK. Symbolically, we may write this as

We denote by [Kl ... the algebra of all matrices consisting of m 2 blocks,

with each M" independently a matrix in K. In

If A is any algebra, the inverse algebra A is obtained

from A by inverting the order of factors in all

C is commutative, C = C. The quaternion division

Finally, a division algebra is defined to be an algebra

The summations here represent direot sums over

13 This is theorem (3.5B) on p. 95 of Weyl's book (reference

The inverse algebra Q is then generated by

Note that this e2 is not necessarily identical with the

This particular representation of Q and

THE THREEFOLD WAY

matrices is called the regular representation. It has

for any two antiunitary elements all a2 in G. Therefore,

for all a in G, and this implies

Since M1(g) and M2(g) are supposed irreducible

But then Eq. (30) becomes

Equations (28) and (35) together give

for all g in G, and therefore by Schur's lemma again

Thus the operator P cannot exist, and the lemma

where the Z; are inequivalent simple algebras and

WIGNER'S CLASSIFICATION OF IRREDUCIBLE

In this section we shall establish the connection

See Wigner (reference 1), p. 75, Theorem 2.

and to the commutator algebras X and Y of A and B.

It is therefore possible to transform i' back into the

standard form i by a real orthogonal transformation

Equation (39) is in fact merely a statement of

where we have written m = In. In all three cases

with M unitary and symmetric, then A(u) is

with j given by Eq. (4). Then Eq. (43) holds if and

then A(g) is potentially real. If X contains an antiunitary operator P with

In eaoh case the oonver8e i8 also true.

The following remarks are corollaries of Theorem I.

THE THREEFOLD WAY

real if and only if it is equivalent to a representation