Genetic algebras∗

I. M. H. Etherington
Mathematical Institute, University of Edinburgh
Received May 8, 1939

Abstract
A population can be classified genetically at various levels, according to
the frequencies of the gametic types which it produces, of the zygotic types
of the individuals which it contains, of types of mating pairs in the preceding
generation, and so on. It is represented accordingly by means of hypercomplex
numbers in one or other of a series of linear algebras (gametic, zygotic, copular,
. . . ), each algebra being isomorphic with the quadratic forms of the preceding
algebra. Such a series of genetic algebras exists for any mode of genetic inheri-
tance which is symmetrical in the sexes. (Genetic algebras for unsymmetrical
inheritance also exist, but are not considered here.) Many calculations which
occur in theoretical genetics can be expressed as manipulations within these
algebras.
The algebras which arise in this way are all commutative nonassociative
linear algebras of a special kind. Firstly, they are baric algebras, i.e. they possess
a scalar representation; secondly, they are train algebras, i.e. the rank equation
of a suitably normalized hypercomplex number has constant coefficients. Some
theorems concerning such algebras are enunciated.

1 Introduction
Two classes of linear algebras, generally nonassociative, are defined in §3 (baric al-
gebras) and §4 (train algebras), and the process of duplication of a linear algebra
in §5. These concepts, which will be discussed more fully elsewhere, arise naturally
∗
Proceedings of the Royal Society of Edinburgh 59 (1939) 242–258.

1
in the symbolism of genetics, as shown in §§6–15. Many of their properties express
facts well-known in genetics; and the processes of calculation which are fundamental
in many problems of population genetics can be expressed as manipulations in the
genetic algebras. In cases where inheritance is of a simple type (e.g. §§10–13, 15) this
constitutes a new point of view, but perhaps amounts to little more than a change of
notation as compared with existing methods. §14, however, indicates the possibility
of generalisations which would seem to be impossible by ordinary methods.
The occurrence of the genetic algebras may be described in general terms as fol-
lows. The mechanism of chromosome inheritance, in so far as it determines the
probability distributions of genetic types in families and filial generations, and ex-
presses itself through their frequency distributions, may be represented conveniently
by algebraic symbols. Such a symbolism is described, for instance, by Jennings [8]
(chap. 9); many applications are given by Geppert and Koller [4]. It is shown in
the present paper that the symbolism is equivalent to the use of a system of related
linear algebras, in which multiplication (equivalent to the procedure of “chessboard
diagrams”) is commutative (P Q = QP ) but nonassociative (P Q · R 6= P · QR). A
population (i.e. a distribution of genetic types) is represented by a normalized hyper-
complex number in one or other algebra, according to the point of view from which
it is specified. If P, Q are populations, the filial generation P × Q (i.e. the statistical
population of offspring resulting from the random mating of individuals of P with
individuals of Q) is obtained by multiplying two corresponding representations of P
and Q; and from this requirement of the symbolism it will be obvious why multipli-
cation must be nonassociative. It must be understood that a population may mean
a single individual, or rather the information we have concerning him in the form of
a probability distribution.
Inheritance will be called symmetrical if the sex of a parent does not affect the
distribution of genetic types produced. Paying attention only to the inheritance of
gene differences (not of phenotypes), every regular mode of symmetrical inheritance
in theoretical genetics has its fundamental gametic algebra, from which other algebras
(zygotic, etc.) are deduced by duplication. From the nature of the symbolism these
are of necessity baric algebras; but it appears on closer examination that they belong
in all cases to the narrower category of train algebras.
(The fundamental algebras can be modified to take account of various kinds of
selection. They are then no longer train algebras, although the baric property and
the relation of duplication sometimes persist.)
Symmetry of inheritance may be disturbed by unequal crossing over in male and
female, by sex linkage, or by gametic selection. These cases are not discussed at all
in the present paper; but it may be stated briefly that in the absence of selection the

2
corresponding genetic algebras (of order n, say) possess train subalgebras (of order
n − 1).
The occurrence of a nonassociative linear algebra in the simplest case of Mendelian
inheritance was pointed out by Glivenko [5].

2 Notation
By principal powers in a nonassociative algebra, I mean powers in which the factors
are absorbed one at a time always on the right or always on the left (see (3.6)).
Otherwise, for the notation and nomenclature for nonassociative products and powers,
see my paper ”On nonassociative combinations” [3]. The word pedigree which occurs
there can now be interpreted almost in its ordinary biological sense.
Elements of a linear algebra (i.e. hypercomplex numbers) will be called elements
and denoted by Latin letter, generally small (a, b, . . .); but normalized elements, i.e. el-
ements of unit weight (§3), will be denoted by Latin capitals (A.B. . . .). The letters
m, n, r, however, denote positive integers.
Elements of the field F over which a linear algebra is defined will be called numbers
and denoted by small Greek letters (α, β, . . .). Thus, an element is determined by
its coefficients, which are numbers. In the genetical applications, F may be taken as
the field of real numbers. The enumerating indices (subscripts and superscripts) take
positive integer values, either 1 to m, 1 to n, or 1 to r, according to the context.
Block capitalsP(A, B, . . .) denote algebras.
The symbol denotes summation with respect to repeated indices, e.g. with
respect to σ in (3.3), with respect to σ and τ in (5.3).
The symbol 1µ stands for a set of 1’s. Thus the formula (6.3) means the same as
n
X
γσµν = 1.
σ=1

The advantage of this notation is that such formulae retain their form under linear
transformations of the basis of a genetic algebra, 1µ being replaced by the vector ξ µ
(cf. (6.12)).

3 Baric algebras
It is well known that a linear associative algebra possesses a matrix representation.
Nonassociative algebras in general do not, but may. The simplest such representation

3
would be a scalar representation on the field F over which the algebra is defined. A
linear algebra X, associative or not, which possesses a non-trivial representation of
this kind, will be called baric.
The definition means that to any element x of X there corresponds a number ξ(x)
of F, not identically zero, such that for all x, y ∈ X, α ∈ F

ξ(x + y) = ξ(x) + ξ(y), ξ(αx) = αξ(x), ξ(xy) = ξ(x)ξ(y). (3.1)

ξ(x) will be called the weight of x, or the weight function of X. If ξ(x) 6= 0, x can be
normalized – that is, replaced by the element

X = x/ξ(x) (3.2)

of unit weight. Elements of zero weight will be called nil elements. The set U of all
nil elements is evidently an invariant subalgebra of X; i.e. XU ⊆ U: it will be called
the nil subalgebra.
Let the multiplication table of a linear algebra X be
X
aµ aν = γσµν aσ , µ, ν, σ = 1, . . . , n, (3.3)

and let the general element be denoted

X
x= αµ a µ . (3.4)

For X to be a baric algebra, it is necessary and sufficient that the equations (3.3),
regarded as ordinary simultaneous equations in F for the unknowns a µ , should possess
a non-null solution aµ = ξ µ . For this is obviously necessary, the ξ µ being the weights
of the basic elements aµ . Conversely, if the condition is satisfied and we take
X
ξ(x) = αµ ξ µ , (3.5)

then (3.1) are at once deducible. The basic weights ξ µ form the weight vector of X.
In the genetical applications, ξ µ = 1µ .
Let the right rank equation ([2], §19), or equation of lowest degree connecting the
right principal powers,
x, x2 , x3 , . . . , xm = xm−1 x, (3.6)
be
f (x) = xr + θ1 xr−1 + θ2 xr−2 + · · · + θr−1 x = 0, (3.7)

4
where each coefficient θm is a homogeneous polynomial of degree m in the coordinates
αµ of x. Then f (x), being zero, is of zero weight. Hence the equation is satisfied when
we substitute ξ(x) for x; consequently, x − ξ(x) must be a factor of f (x). The same
is true for the left rank equation. Thus

ξ(x) is a root of the right and left rank equations. (3.8)

The weight function of an algebra is not necessarilyPunique. In fact, a commutative

associative linear algebra for which the determinant | γσµν γτστ | does not vanish has n
independent weight functions; and its rank equation is hence completely determined
by (3.8) ([2], §55, and the references given there).

4 Train algebras
A baric algebra with the weight function ξ(x) and right rank equation (3.7) will
be called a right train algebra if the coefficients θm , in so far as they depend on the
element x, depend only on ξ(x). A left train algebra is defined similarly. For simplicity,
suppose multiplication commutative, so that we may drop “left” and ”right”.
Since θm is homogeneous of degree m in the coordinates of x, it must in a train
algebra be a numerical multiple of ξ(x)m . Hence (if the field F be sufficiently extended,
e.g. to include complex numbers) the rank equation can be factored:

f (x) ≡ x(x − ξ)(x − λ1 ξ)(x − λ2 ξ) · · · = 0. (4.1)

(It is implied that when the left side is expanded, powers of x are interpreted as
principal powers.) The numbers 1, λ1 , λ2 , . . . are the principal train roots of the
algebra.
For a normalized element (3.7) becomes

f (X) ≡ X r + θ1 X r−1 + θ2 X r−2 + · · · + θr−1 X = 0, (4.2)

where now the θ’s are constant (i.e. independent of X); and (4.1) becomes

f (X) ≡ X(X − 1)(X − λ1 )(X − λ2 ) · · · = 0. (4.3)

Since (4.2) can be multiplied by X any number of times, it can be regarded as a linear
recurrence equation with constant coefficients connecting the principal powers of the
general normalized element X. Solving the recurrence relation for X m (m ≥ r) in
the usual way, we obtain 1, λ1 , λ2 , . . . as the roots of the auxiliary equation; hence a

5
formula for X m can be written down in terms of X, X 2 , . . . , X r−1 . Hence also for
the general non-nil element x = ξX, the value of xm = ξ m X m is known; while for a
nil element u, um = 0 (m ≥ r).
The properties of train algebras will be studied elsewhere, and the following the-
orem is proved:
If (1) X is a baric algebra; (2) its nil subalgebra U is nilpotent [9] (p. 111); (3)
for m = 1, 2, 3, . . ., the subalgebra U(m) , consisting of all products of altitude m [3]
(p. 156) formed from nil elements is an invariant subalgebra of X (as it necessarily
is of U); then X is a train algebra.
For train algebras of rank r = 2 or 3, provided that the principal train roots
do not include 12 , the conditions are necessary as well as sufficient; but I cannot
say whether this converse holds more generally or not. I will call X a special train
algebra if it satisfies the conditions (1), (2), (3). In such algebras it can be shown that
there are many other sequences which have properties like those of the sequence of
principal powers; i.e. sequences of elements derived from the general element, which
satisfy linear recurrence equations whose coefficients, being functions of the weight
only, become constants on normalization. Such sequences will be called trains. For
example, the sequence of plenary powers

x, x2 , (x2 )2 , ((x2 )2 )2 , . . . , (4.4)

and the sequence of primary products

x, Y x, Y (Y x), Y (Y (Y x)), . . . , (4.5)

form trains in a special train algebra.

It is convenient to denote the m-th element of a train by x[m] , and to regard it as a
symbolic m-th power of x. Let the normalized recurrence equation, or train equation,
be
g[X] ≡ X [s] + φ1 X [s−1] + φ2 X [s−2] + · · · + φs−1 X = 0, (4.6)
where the φ’s are numerical constants. It is implied that the equation may be symbol-
ically “multiplied through” by X any number of times. It may also be symbolically
factorized:
g[X] ≡ X[X − 1][X − µ1 ][X − µ2 ] · · · = 0. (4.7)
The square brackets indicate that after expansion powers of X are to be interpreted
as symbolic powers. The expansion being performed as in ordinary algebra, multipli-
cation of the symbolic factors is commutative and associative. Extra factors may be
introduced without destroying the validity of the train equation; but assuming that

6
all superfluous factors have been removed, s is the rank of the train, and the numbers
1, µ1 , µ2 , . . . are the train roots, by means of which a formula for X [m] (m ≥ s) can
be written down.
In the applications to genetics, it will be found that all the fundamental sym-
metrical genetic algebras are special train algebras. Various trains have genetical
significance; the X [m] represent successive discrete generations of an evolving popula-
tion or breeding experiment, and the train equation is the recurrence equation which
connects them.
Thus, for example, plenary powers (4.4) refer to a population with random mat-
ing; principal powers (3.6) to a mating system in which each generation is mated back
to one original ancestor or ancestral population; and the primary products (4.5) to
the descendants of a single individual or subpopulation X mating at random within a
population Y . Other mating systems are described by other sequences, and in various
well-known cases these have the train property – that is, the determination of the
m-th generation depends ultimately on a linear recurrence equation with constant co-
efficients. It usually happens that the train roots are real, distinct, and not exceeding
unity. Hence it may be shown that X [m] tends to equilibrium with increasing m; the
rate of approach to equilibrium is ultimately that of a geometrical progression with
common ratio equal to the largest train root excluding unity; but it may be some
generations (depending on the number of train roots) before this rate of approach is
manifest.
Train roots may be described as the eigenvalues of the operation of symbolic mul-
tiplication by X, or in genetic language, the operation of passing from one generation
to the next.
Train algebras of (principal) rank 3, which occur in several contexts in genetics,
have certain special properties. For example, if the train equation for principal powers
is X(X − 1)(X − λ) = 0, then the train equation for plenary powers is X(X − 1)(X −
2λ) = 0; and vice versa. Examples may be seen below in (10.12), (12.4,5), (15.3),
where respectively λ = 0, 21 (1 − ω), 61 .

5 Duplication
Let X
aµ aν = γσµν aσ (5.1)
be the multiplication table of a linear algebra X with basis aµ (µ = 1, . . . , n). Then
X X
aµ aν · aθ aφ = γσµν aσ · γτθφ aτ .

7
Writing
aµ aν = aµν , (5.2)
this becomes X
aµν aθφ = γσµν γτθφ aστ , (5.3)
which may be regarded as the multiplication table of another linear algebra, isomor-
phic with the totality of quadratic forms in the original algebra. It will be called
the duplicate of X, and denoted X0 . It is commutative and of order 21 n(n + 1) if
X is commutative; noncommutative and of order n2 if X is noncommutative. It is
generally nonassociative, even if X is associative. It is not to be confused with what
may be called the direct square of X, or direct product of two algebras isomorphic
with X; this would be an algebra of order n2 , having the multiplication table
X
aµν aθφ = γσµθ γτνφ aστ , (5.4)
differing from (5.3) in the arrangement of indices.
Some theorems on duplication will be proved elsewhere. It will be shown that
the duplicates (i) of a linear transform of an algebra, (ii) of the direct product of
two algebras, (iii) of a baric algebra with weight vector ξ µ , (iv) of a train algebra
with principal train roots 1, λ, µ, . . . , are respectively (i) a linear transform of the
duplicate algebra, (ii) the direct product of the duplicates, (iii) a baric algebra with
weight vector ξ µ ξ ν , (iv) a train algebra with principal train roots 1, 0, λ, µ, . . . .
These theorems are relevant as follows: (iii) in view of §§7, 8; (ii) in view of §9; (i) in
connection with the method used in §14; (iv) in deriving equations such as (10.10),
(12.6).
Duplication of an algebra may be compared with the process of forming the second
induced matrix of a given matrix ([1], cf. also [10]).

6 Gametic algebras
Consider the inheritance of characters depending on any number of gene differences
at any number of loci on any number of chromosomes in a diploid or generally au-
topolyploid species. Assume that inheritance is symmetrical in the sexes: the sex
chromosomes are thus excluded, and crossing over if present must be equal in male
and female.
Let G1 , G2 , . . . , Gn denote the set of gametic types determined by these gene
differences. Then there will be
1
m = n(n + 1) (6.1)
2
8
zygotic types Gµ Gν (= Gν Gµ ). The formula giving the series of gametic types pro-
duced by each type of individual, and their relative frequencies, may be written
X
Gµ Gν = γσµν Gσ , (6.2)

with the normalizing conditions

X
γσµν 1σ = 1; (6.3)

γσµν is then the probability that an arbitrary gamete produced by an individual of

zygotic type Gµ Gν is of type Gσ .
(I speak of zygotic types – individuals distinguished by the gametes from which
they were formed – rather than genotypes – individuals distinguished by the gametes
which they produce – because the Gµ Gν are not all distinct genotypes if more than
one chromosome is involved: the zygotic algebra, §7, will have the same train equation
if genotypes are used, but will then not be a duplicate algebra.)
A population P which produces gametes Gµ in proportions αµ may be represented
by writing X
P = αµ G µ . (6.4)
Imposing the normalizing condition
X
αµ 1µ = 1, (6.5)

αµ denotes the probability that an arbitrary gamete produced by an arbitrary indi-

vidual of P is of type Gµ .
A population may also be described by the proportions of the zygotic types Gµ Gν
which it contains; thus we may write
X
P = αµν Gµ Gν , (6.6)

with the normalizing condition

X
αµν 1µ 1ν = 1, (6.7)

and a similar probability interpretation. We may suppose without loss of generality

that αµν = ανµ , so that in (6.6) the coefficient of Gµ Gν is 2αµν if µ 6= ν. The two
representations are connected by the gametic series formulae (6.2): that is to say,
from the zygotic representation (6.6) follows the gametic representation
X
P = αµν γσµν Gσ . (6.8)

9
 If two populations P , Q intermate at random, representations of the first filial
generation are obtained by multiplying the gametic representations of P and Q; i.e. if
X X
P = αµ G µ , Q= βµ G µ , (6.9)

the population of offspring is

X X
PQ = α µ βν G µ G ν = αµ βν γσµν Gσ . (6.10)

In particular, the population of offspring of random mating of P within itself is given

by P 2 .
We may now view the situation abstractly. The gametic series (6.2) form the
multiplication table of a commutative nonassociative linear algebra with basis Gµ
(µ = 1, . . . , n). It will be called the gametic algebra for the type of inheritance
considered, and denoted G. The equations (6.3) show that G is a baric algebra with
weight vector
ξ µ = 1µ . (6.11)
With regard to its gametic type frequencies, a population is represented by a normal-
ized element (6.4) of G. Multiplication in G has the significance described in §1, and
it follows from the multiplicative property of the weight in a baric algebra that P Q
will be automatically normalized if P and Q are.

7 Zygotic algebras
When individuals of types Gµ Gν , Gθ Gφ mate, the probability distribution of zygotic
types in their offspring can be obtained by multiplying the gametic representations
(given in (6.2)) together, and leaving the product in quadratic form (as in (6.10)).
We obtain X
Gµ Gν · Gθ Gφ = γσµν γτθφ Gσ Gτ ; (7.1)
or, writing
Z µν = Gµ Gν (7.2)
to emphasize the union of paired gametes into single individuals,
X
Z µν Z θφ = γσµν γτθφ Z στ . (7.3)

These 12 m(m + 1) equations, then, are the formulae giving the series of zygotic types
produced by the mating type of couple Z µν × Z θφ , the probability of Z στ being the
corresponding coefficient γσµν γτθφ + γτµν γσθφ (if σ 6= τ ) or γσµν γτθφ (if σ = τ ).

10
 The linear algebra with basis Z µν and multiplication table (7.2) will be called the
zygotic algebra for the type of inheritance considered. It is a baric algebra with weight
vector 1µ 1ν , the duplicate of the gametic algebra G, and will be denoted
Z = G0 . (7.4)
A population, regarded as a distribution of zygotic types, is represented by a normal-
ized element X X
P = αµν Z µν where αµν 1µ 1ν = 1;
and multiplication in Z, as in G, has the significance described in §1. A product left
in quadratic form in the Z’s gives now the probability distribution of couples Z µν Z θφ
among the parents; or, as I shall call it, the copular representation of the population
of offspring.

8 Further duplicate genetic algebras

The process of duplication can be applied repeatedly. Thus the 12 m(m + 1) types of
paired zygotes, or couples,
K µν·θφ = Z µν Z θφ , (8.1)
can be taken as the basis of a new linear algebra
K = Z0 = G00 . (8.2)
Call it the copular algebra. A normalized element with positive coefficients
X X
P = αµν·θφ K µν·θφ where αµν·θφ 1µ 1ν 1θ 1φ ,

is the copular representation of a population – the probability distribution of couples

in the parents of the individuals comprised in the population.
Similarly, in the next duplicate algebra K0 , the basic symbols would classify tetrads
of grandparents.
In all these algebras, multiplication has the significance described in §1.

9 Combination of genetic algebras

Consider two distinct genetic classifications referring to the same population P , firstly
into a set of m genetic types
A1 , A2 , . . . , Am ;

11
secondly into a set of n genetic types
B1, B2, . . . , Bn
of the same kind (gametic, zygotic, etc.). Let the corresponding genetic algebras be
A, B with multiplication tables
X X
Aµ Aν = γσµν Aσ , B θ B φ = δτθφ B τ .
By taking account of both classifications at once, we obtain a third classification
which may be called their product, into mn genetic types:
C µθ = Aµ B θ .
The type C µθ comprises all individuals (gametes, zygotes, etc.) who are of type Aµ
in the first classification, B θ in the second.
If the characters of the two classifications are inherited independently, i.e. if they
involve two quite distinct sets of chromosomes, then the probabilities γσµν , δτθφ refer to
independent events. Hence the genetic algebra with basis C µθ is the direct product
C = AB;
i.e. its multiplication table is
X
C µθ C νφ = γσµν δτθφ C στ .
It follows that a genetic algebra which depends on several autosomal linkage groups
must be a direct product ABC · · · of genetic algebras, one factor algebra for each
linkage group.
If, however, the A and B classifications are independent but genetically linked,
i.e. if they involve two quite distinct sets of gene loci but not distinct sets of chromo-
somes, then the probabilities γσµν , δτθφ are not independent. Regarded as a linear set,
C is still the product of the linear sets A and B; but the algebra C will not be the
direct product of the algebras A and B (except in the very exceptional case when
all crossing over values between A and B are precisely 50 per cent). It is, however,
still the case that C contains subalgebras isomorphic with A and B. For example,
if these algebras are gametic, and if we keep the first index of C µθ constant, we are
virtually disregarding all the A-loci, so we obtain a subalgebra isomorphic with B;
and this can be done in m ways.
Hence a genetic algebra based on the allelomorphs of several autosomal loci pos-
sesses numerous automorphisms.
It will be shown in §14 that even when linkage is involved the gametic algebra can
be symbolically factorized, and regarded as a symbolic direct product of noncommu-
tative factor algebras, one for each locus (see (14.12)).

12
10–15 Examples of symmetrical genetic algebras
A more detailed description of practical applications will be given elsewhere. My
object here is simply to show that the genetic algebras are train algebras. I give in
each case the principal and plenary train equations, i.e. the identities of lowest degree
connecting respectively the sequences of principal and plenary powers of a normalized
element. As explained in §4, these are really recurrence equations, and have a special
significance in genetics.

10 Simple Mendelian inheritance

For a single autosomal gene difference (D, R), the gametic multiplication table is

DD = D, DR = 12 D + 12 R, RR = R. (10.1)

Writing
A = DD, B = DR, C = RR, (10.2)
we find, e.g.,
2
B2 = 1
2
D + 12 R = 41 A + 21 B + 14 .
Hence and similarly the zygotic multiplication table is

A2 = A, B 2 = 41 A + 12 B + 14 , C 2 = C, (10.3)
AB = 21 A + 21 B, AC = B, BC = 21 B + 12 C.

Call these two algebras G2 , Z2 (Z2 = G02 ), and denote their general elements

G2 : x = δD + ρR, (10.4)
Z2 : x = αA + 2βB + γC. (10.5)

The principal rank equations are

G2 : x2 − (δ + ρ)x = 0, (10.6)
Z2 : x3 − (α + 2β + γ)x2 = 0; (10.7)

and the plenary rank equations (or identities of lowest degree connecting plenary
powers of the general elements) are (10.6) and

Z2 : (x2 )2 − (α + 2β + γ)2 x2 = 0. (10.8)

13
 A population P is represented by an element of unit weight in either algebra,
i.e. (10.4) or (10.5) with
δ + ρ = 1, α + 2β + γ = 1, (10.9)
the ratios δ : ρ, α : 2β : γ giving the relative frequencies of the gametic types which
it produces or genotypes which it contains. In this case (10.6), (10.7), (10.8) become
the train equations
G2 : P 2 = P, (10.10)
Z2 : P 3 = P 2, 2 2
(P ) = P, (10.11)
expressing facts well-known in genetics. It is convenient to write these equations in
the form (cf. 4.7)
G2 : P (P − 1) = 0, (10.12)
Z2 : P 2 (P − 1) = 0, 2
P [P − 1] = 0. (10.13)

11 Multiple allelomorphs
For n allelomorphs Gµ (µ = 1, . . . , n), the gametic and zygotic multiplication tables
are
Gµ Gν = 21 Gµ + 12 Gν , (11.1)
µν θφ 1 µθ 1 µφ 1 νθ 1 νφ
Z Z = 4
Z + 4
Z + 4
Z + 4
Z , (11.2)
where Z µν = Gµ Gν . The algebras Gn , Zn so determined reduce to G2 , Z2 when
n = 2; and they have in general the same train equations (10.11), (10.12).

12 Linked allelomorphs
For two linked series of multiple allelomorphs, respectively m and n in number, with
crossing over probability ω, the gametic multiplication table is
Gµα Gνβ = 12 (1 − ω)(Gµα + Gνβ ) + 21 ω(Gµβ + Gνα ), (12.1)
where Gµα (µ = 1, . . . , m; α = 1, · · · , n) are the mn gametic types. Denote this
gametic algebra Gmn (ω). The principal and plenary rank equations are
x3 − 12 (3 − ω)ξx2 + 21 (1 − ω)ξ 2 x = 0, (12.2)
(x2 )2 − (2 − ω)ξ 2 x2 + (1 − ω)ξ 3 x = 0, (12.3)

14
giving for a normalized element P the train equations
 
3 1 2 1 1−ω
P − 2 (3 − ω)P + 2 (1 − ω)P ≡ P (P − 1) P − = 0, (12.4)
2
(P 2 )2 − (2 − ω)P 2 + (1 − ω)P ≡ P [P − 1][P − (1 − ω)] = 0. (12.5)

In the duplicate algebra Zmn (ω) = Gmn (ω)0 the corresponding equations are
 
2 1−ω
P (P − 1) P − = 0, P 2 [P − 1][P − (1 − ω)] = 0. (12.6)
2

13 Independent allelomorphs
Consider two series of multiple allelomorphs in separate autosomal linkage
groups. This being indistinguishable from the case of §12 with µ = 12 , the gametic
algebra is Gmn ( 21 ). As in §9, it may also be expressed as the direct product Gm Gn .

14 Linkage group
I will first rewrite equations (12.1) with a change of notation. I will then write down
the analogous equations for the case of three linked loci, and examine the structure
of the corresponding algebra. This will be a sufficient indication of the procedure
which can be followed out quite generally for a complete linkage group comprising
any number of loci on one autosome, with any number of allelomorphs at each locus.
The method may be extended to include any number of linkage groups.
Equations (12.1) may be written

AB · A0 B 0 = 21 (1 − ω)(AB + A0 B 0 ) + 21 ω(AB 0 + A0 B). (14.1)

Here A and B refer to the two gene loci. Aµ B α would mean the same as Gµα – a
gamete with the µ-th allelomorph at the A-locus and the α-th at B; but dropping the
indices AB and A0 B 0 stand for any particular gametic types, the same or different.
(14.1) may again be rewritten

AB · A0 B 0 = 21 ω(A + χA0 )(B + χB 0 ), (14.2)

where ω = 1 − ω and χ is an operator which interchanges ω and ω, so that χ2 = 1

and ωχ = ω.

15
 Now consider the case of three loci A, B, C, having respectively m, n, r al-
lelomorphs, and crossing over probabilities ωAB , ωBC , ωAC. The gametic algebra
may be symbolized conveniently as Gmnr (ω), where ω is the symmetrical matrix of
the crossing over values, with diagonal zeros. Its multiplication table, comprising
1
2
mnr(mnr + 1) formulae, is

ABC · A0 B 0 C 0 = 21 λ(ABC + A0 B 0 C 0 ) + 21 µ(A0 BC + AB 0 C 0 ) (14.3)

+ 12 ν(AB 0 C + A0 BC 0 ) + 21 ρ(ABC 0 + A0 B 0 C),

where

λ + µ + ν + ρ = 1, (14.4)
µ + ν = ωAB , ν + ρ = ωBC , µ + ρ = ωAC . (14.5)

The ω’s are not independent, but are connected only by an equality [6]:

ωAC = ωAB + ωBC − κωAB ωBC , where 0 ≤ κ ≤ 2, (14.6)

from which may be deduced

µρ ≥ νλ. (14.7)
Now introduce the following operators:

χ1 interchanges λ with µ, and ν with ρ,

χ2 interchanges λ with ν, and ρ with µ, (14.8)
χ3 interchanges λ with ρ, and µ with ν.

Together with 1, they form an Abelian group, having the relations

χ2 χ3 = χ 1 , χ3 χ1 = χ 2 , χ1 χ2 = χ 3 , (14.9)
χ21 = χ22 = χ23 = χ1 χ2 χ3 = 1.

(14.3) may then be rewritten:

ABC · A0 B 0 C 0 = 12 λ(A + χ1 A0 )(B + χ2 B 0 )(C + χ3 C 0 ). (14.10)

This symbolism can be manipulated with considerable freedom. For example, an

expression such as (αABC + βA0 BC) can be written (αA + βA0 )BC; and when two
such expressions are multiplied, the distributive law works. The interchange symbols
cooperate in the same way.

16
 (14.10) may again be rewritten

ABC · A0 B 0 C 0 = (χ0 A + χ1 A0 )(χ0 B + χ2 B 0 )(χ0 C + χ3 C 0 ), (14.11)

where χ0 = 1, and the operand 12 λ is implied. Finally, (14.11) may be analyzed into

AA0 = χ0 A + χ1 A0 , BB 0 = χ0 B + χ2 B 0 , CC 0 = χ0 C + χ3 C 0 . (14.12)

This separation of the symbols, or factorization of the algebra (cf. end of §9), will
evidently yield valid results, provided that after recombination and application of
(14.9), χ0 is interpreted as 21 λ, χ1 as 12 µ, χ2 as 12 ν, χ3 as 12 ρ. It must be noted that
the symbols when separated in this way are noncommutative; e.g. AA0 6= A0 A, since
ABC · A0 B 0 C 0 6= A0 BC · AB 0 C 0 .
Select a particular gametic type ABC, and write

A − A = u, B − B = v, C − C = w, (14.13)

where A 6= A, B 6= B, C 6= C. Thus the symbols u, v, w are nil elements having

respectively m − 1, n − 1, r − 1 possible values. We have from (14.12):

A2 = (χ0 + χ1 )A,
Au = A2 − AA = (χ0 + χ1 )A − (χ0 A + χ1 A) = χ1 u,
uA = A2 − AA = (χ0 + χ1 )A − (χ0 A + χ1 A) = χ0 u,
u2 = A2 − AA − AA + A2
= (χ0 + χ1 )A − (χ0 A + χ1 A) − (χ0 A + χ1 A) + (χ0 + χ1 )A = 0,

and eight similar equations.

Now write

ABC = I, uBC = u, AvC = v, ABw = w, (14.14)

Avw = vw, uBw = wu, uvC = uv, uvw = uvw.

The symbols I, u, v, w, vw, wu, uv, uvw thus introduced are linear and linearly
independent in the gametic type symbols; and their number is

1 + (m−1) + (n−1) + (r−1) + (n−1)(r−1) + (r−1)(m−1) + (m−1)(n−1)

+ (m − 1)(n − 1)(r − 1) = mnr,

17
which is equal to the number of gametic type symbols. They may thus be taken as
a new basis for the gametic algebra. The transformed multiplication table is then
easily deduced. We find, for example,

I 2 = I,
Iu = Au · B2 · C2 = χ1 (χ0 + χ2 )(χ0 + χ3 )u = (χ0 + χ1 + χ2 + χ3 )u = 21 u,

since χ0 + χ1 + χ2 + χ3 is to be interpreted as 12 λ + 12 µ + 21 ν + 12 ρ = 21 . Similarly:

Ivw = 21 (λ + µ)vw, Iuvw = 12 λuvw, ,

2
u v = 21 (ν + µ)uv, u vw = 21 µuvw, u = u uv = u uvw = 0.

These results are typical, all other products in the transformed multiplication table
being obtainable from them by cyclic permutation of u, v, w and µ, nu, ρ and 1, 2,
3.
It is now readily verifiable that the algebra has the structure of a special train
algebra as defined in §4, with

U = (u, v, w, vw, wu, uv, uvw), U(1) = (vw, wu, uv, uvw),
U(2) = (uvw), U(1) = 0.

Many of its properties can be most easily deduced from this transformed form. It can
be shown that its principal and plenary train roots, other than unity, are the results
of
χ0 , χ0 + χ 1 , χ0 + χ 2 , χ0 + χ 3 ,
operating respectively on 21 λ and λ. Further details are postponed until the properties
of special train algebras are studied elsewhere.

15 Polyploidy
A single example – the simplest possible – will illustrate the occurrence of special
train algebras in this connection. The gametic algebra with multiplication table

A2 = A, B 2 = 61 A + 23 B + 16 C, C 2 = C, (15.1)
AB = 21 A + 12 B, AC = 61 A + 23 B + 16 C, BC = 12 B + 12 C,

refers to the inheritance of a single autosomal gene difference in autotetraploids.

(Cf. [7], the case m = 2, with A, B, C written for A2 , Aa, a2 .)

18
 This is a special train algebra, as may be seen by performing the transformation

A = A, A − B = u, A − 2B + c = p. (15.2)

It has the principal and plenary train equations

P (P − 1) P − 61 = 0, P [P − 1] P − 13 = 0.

 
(15.3)

References
[1] A. C. Aitken, The normal form of compound and induced matrices, Proc. Lond.
Math. Soc. (2) 38 (1935) 354–376.

[2] L. E. Dickson, Linear algebras, Cambridge Tract, No. 16, 1914 (reprinted 1930).

[3] I. M. H. Etherington, On nonassociative combinations, Proc. Roy. Soc. Edin. 59

(1939) 153–162.

[4] H. Geppert and S. Koller, Erbmathematik, Leipzig, 1938.

[5] V. Glivenko, Algèbre mendelienne, Moscow Acad. Sci. C. R. 4 (1936) 385–386.

[6] J. B. S. Haldane, The combination of linkage values, Journ. Gen. 8 (1918) 299–
309.

[7] J. B. S. Haldane, Theoretical genetics of autopolyploids, Journ. Gen. 22 (1930)

359–372.

[8] H. S. Jennings, Genetics, London, 1935.

[9] J. H. M. Wedderburn, On hypercomplex numbers, Proc. Lond. Math. Soc. (2) 6

(1908) 77–118.

[10] J. H. M. Wedderburn, On certain theorems in determinants, Proc. Edin. Math.

Soc. (1) 27 (1908) 67–69.

19
Mathematical Reviews
MR0000597 (1,99e): Many significant features of genetic theory can be expressed
mathematically in terms of linear algebras or hypercomplex number-systems, gener-
ally nonassociative, over the rational or real fields; for example, if the units are taken
to represent different genetic types, their multiplication table represents a tabulation
of the probable distribution among these types of the issues of the possible crosses
between the several types. The simple instances previously noted by Glivenko (C. R.
(Doklady) Acad. Sci. USSR 4 (1936) 385–386) are multiplied in the present paper so
as to embrace a wide range of genetic situations. The author points out a number of
algebraic properties and processes common to the several genetic algebras described,
and proposes to discuss them elsewhere. In particular, he observes that the genetic
algebras are all “special train algebras”, that is, possess non-trivial homomorphic
mappings on their coefficient fields and exhibit special features in respect to their
nilpotent elements. (M. H. Stone)

Zentralblatt
Zbl 0027.29402: Eine barische Algebra (für diesen Begriff siehe dies. Zbl. 27, 155)
heißt Train-Algebra, wenn die Koeffizienten der Ranggleichung eines allgemeinen El-
ementes x der Algebra nur von dem Gewicht ξ(x) abhängen. Wird der Koeffizien-
tenkörper passend erweitert, so hat die Ranggleichung einer Train-Algebra die Gestalt

x(x − ξ)(x − λ1 ξ)(x − λ2 ξ) · · · = 0,

wobei das Produkt so zu bilden ist: Das Produkt der ersten zwei Faktoren wird mit
dem dritten Faktor multipliziert, das entstehende Produkt mit dem vierten usw. Eine
barische Algebra X heißt eine spezielle Train-Algebra, wenn die Unteralgebren U (m) ,
bestehend aus allen Produkten von der Höhe m aus Nilelementen von X, sämtlich
nilpotent sind. An anderer Stelle (vgl. vorst. Referat) hat der Autor beweisen, daß
jede spezielle Train-Algebra eine Train-Algebra ist. Das Duplikat X 0 einer Algebra
X mit der Multiplikationstafel
X
aµ aν = γσµν aσ

wird durch die Multiplikationstafel

X
aµν aθφ = γσµν γτθφ aστ

20
definiert. — Alle diese Begriffe finder in der Erbmathematik Anwendung. Es seien G1 ,
G2 , . . . , Gn die Gametentypen, die durch die verschiedenen Gene in irgendwelchen
Chromosomen, Geschlechtchromosome ausgenommen, möglich sind. Die Gameten
n


paaren sich dann zu 2 Zygotentypen G G = Gν Gµ . Die Gameten-Nachkommenschaft
µ ν

eines Individuums vom Typus Gµ Gν sei durch

X
Gµ Gν = γσµν Gσ (15.4)

gegeben, wobei γσµν die Warscheinlichkeit ist, daß eine Zygote vom Typus Gµ Gν eine
Gamete vom Typus Gσ hervorbringt. Dann ist (15.4) die Multiplikationstafel einer
kommutativen barischen Algebra, der Gametenalgebra G. Wenn Individuen von den
Typen Gµ Gν und Gθ Gφ gepaart werden, so wird ihre Nachkommenschaft durch
X
Gµ Gν · Gθ Gφ = γσµν γτθφ Gσ Gτ

gegeben.P Die Zygotenalgebra G0 ist also das Duplikat der Gametenalgebra. Ele-
mente aµν Gµ Gν dieser Algebra stellen, wenn ihr Gewicht zu 1 normiert wird,
Bevölkerungen dar, in denen die Individuen Gµ Gν mit Wahrscheinlichkeiten aµν
vertreten sind. An mehreren Beispielen (einfache Mendelsche Vererbung, multi-
ple Allelomorphie, Kopplung, Kopplungsgruppen, Tetraploide) wird gezeigt, daß die
wichtigsten in der Erbmathematik vorkommenden Algebren spezielle Train-Algebren
sind. (B. L. van der Waerden)

English translation: A baric algebra (for this concept see Zbl. 27, 155) is called a
train algebra if the coefficients of the rank equation of an arbitrary element x of the
algebra depend only on the weight ξ(x). After a suitable extension of the coefficient
field, the rank equation of a train algebra has the form

x(x − ξ)(x − λ1 ξ)(x − λ2 ξ) · · · = 0,

in which the product is built up as follows: the product of the first two factors is
multiplied with the third factor, the resulting product with the fourth, etc. A baric
algebra X is called a special train algebra if the subalgebras U (m) , consisting of all
products of height m of nil elements of X, are all nilpotent. In another place (see
the previous review) the author has shown that every special train algebra is a train
algebra. Given an algebra X with multiplication table
X
aµ aν = γσµν aσ ,

21
the duplicate X 0 is defined by the multiplication table
X
aµν aθφ = γσµν γτθφ aστ .

All these concepts have application to the mathematics of heredity. Let G1 , G2 , . . . ,

Gn be the gametic types which are possible for the different genes in any

chromosome
(sex chromosomes are excluded). The gametes pair together to form n2 zygotic types
Gµ Gν = Gν Gµ . The gametic progeny of an individual of type Gµ Gν are given by
X
Gµ Gν = γσµν Gσ (15.5)

where γσµν is the probability that a zygote of type Gµ Gν produces a gamete of type Gσ .
Then (15.5) is the multiplication table of a commutative baric algebra, the gametic
algebra G. If individuals of types Gµ Gν and Gθ Gφ are paired, then their progeny are
given by X
Gµ Gν · Gθ Gφ = γσµν γτθφ Gσ Gτ .
The
P zygotic algebra G0 is thus the duplicate of the gametic algebra. Elements
aµν Gµ Gν of this algebra represent, if the weight is normalized to 1, populations in
which the individuals Gµ Gν have probability aµν . Several examples (simple Mendelian
inheritance, multiple allelomorphs, linking, linkage groups, tetraploidy) show that the
most important algebras arising in the mathematics of heredity are special train al-
gebras. (B. L. van der Waerden)

22