SPECIAL TRAIN ALGEBRAS

By I. M. H. ETHERINGTON (Edinburgh)
[Received 2 July 1940]
IN a previous paper (4) I have defined three kinds of non-associative
linear algebras, namely, baric algebras, train algebras, and special
train algebras, which appear in the symbolism of genetics. Train
algebras and special train algebras are by definition baric algebras;
and I have asserted that special train algebras are train algebras.
This paper supplies the proof.
The commutative law of multiplication is assumed in § 3 (not in
§§1, 2); and the theorem is proved on this assumption in §4, the
extension to non-commutative algebras being indicated briefly.
Methods of obtaining the principal and plenary train roots of a
special train algebra are shown by an example in § 5.

1. Powers of a non-associative linear algebra

I have denned elsewhere (3) the shape s, degree 8, and altitude a
of a non-associative product or power. (In the case of a power, the
shape is the same as its index.) Of these 8, a are ordinary integers,
S being simply the number of factors; while s is an integer in an
arithmetic with non-associative addition.
By the sth power of a linear algebra A, written A8, is meant the
set of all linear combinations of products having the shape s formed
from elements of A. By the 8th involute of A, written A[8), is meant
the sum of all powers of A of degree 8; in other words, the set of
all linear combinations of products of 8 elements of A.* By the ath
generation of A, written A<-a\ is meant the sum of all powers of A of
altitude <x; in other words, the set of all linear combinations of pro-
ducts of altitude a formed from elements of A. Note that
A = AW = A<°>.
By considering how products of a given degree or altitude can be
formed from those of lower degree or altitude, we find that
A!8' = AA' 8 -«+Ai i! W 8 - 2 i+...+A' s - 1 'A, (1.1)
«*-«A, (1.2)

• Cf. (8), 111.

3695.12
2 I. M. H. ETHERINGTON
equations which can obviously be simplified when multiplication is
commutative.
By grouping its factors a non-associative product of given shape
can always be regarded as a product of simpler shape, fewer factors,
lower altitude. Consequently
A[8i ^ A [S-l] ^ ^ A[2] ^ A ) (L3)

A(oc) ^ A («-l) < ... ^ A»> < A. (1.4)

Since evidently
AAft < A'8', A^A < A'8', AA<<*> < A<°>, A«*>A < A<a>, (1.5)
it follows that any involute or generation of A is an invariant sub-
algebra of A. On the other hand, a power of A, A8, is not necessarily
an algebra, since it may not be closed as regards multiplication.
2. Nilpotent algebras
A linear algebra A will be called nilpotent of degree 8 if A[8] = 0,
A[S-i] _£ 0 . a n d nUpOtent of altitude a if A<°> = 0, A^-w # 0. (Other
writers use index in place of degree in this context.)
In the first case, any product of 8 or more elements vanishes.
Consider a product of altitude a' = 8—1. Its degree 8' satisfies
8' > a ' + l , that is, 8' > 8.* Hence this product vanishes, and thus,
if A is nilpotent of. degree 8, it is nilpotent of altitude 8—1 at most.
In the second case, any product in A of altitude a vanishes. Con-
sider a product of degree 8' = 2a. Its altitude a' satisfies a' ^ Iog28',
that is, <x' ^ a* Hence this product vanishes; and thus, if A is
nilpotent of altitude a, it is nilpotent of degree 2 a at most.
Wedderburnf has stated that, if A is nilpotent of degree 8, then
A[S-1) < A[8_2] < ... < At2] < A ;

butf his proof is based on the incorrect equation§

At8] = AAl8-l]+A[8-l]A> (2.1)

The same method of proof, however, using (1.2) instead of (2.1),

yields the theorem: If A is nilpotent of altitude a, then
A(a-1) < A(a-2) < ... < A(D < A . (2.2)

3. Canonical form of a commutative special train algebra

Let X be a special train algebra with commutative multiplication.
The definition|| contains three postulates:
• See (3) (14). t (8) H I . J Cf. (6), §4, footnote.
§ Cf. equation (1.1) here. || (4), 246.
 SPECIAL TRAIN ALGEBRAS 3
(1) X is a baric algebra. Hence* we can provide by a suitable
linear transformation that one base element of X, say A, shall be
of unit weight, and the rest, say u" {a = 1,..., n), of zero weight. We
shall suppose this done. The u" form the basis of the nil sub-algebra
U, which we know is an invariant sub-algebra of X; and the multi-
plication table of X takes the form
, Au°=(u), u°uT=(u), (3.1)
where the u's in brackets denote unspecified linear combinations of
the u".
(2) U is nilpotent of a certain degree. As shown in § 2, this is equi-
valent to saying that U is nilpotent of a certain altitude, say a.
Then we know that
0 = U ^ < IK01-" < U<a-2> < ... < Ud> < U. (3.2)
Consequently, by an appropriate linear transformation of the ua, we
can separate the basis of U into sets of, let us say, Mo-elements,
itj-elements, w2-elements, etc., such that the wo's belong to U but not
to U'1', the ut'a belong to IP' but not to U<2>, the u2's belong to U<2>
but not to U<3>, and so on. Since (cf. (1.2)) UU<0> = U0+D, the multi-
plication table of U will take the form
(«0) K)
(uvu „•••) («2.«8, ...)
(u2, ua,...)
(ttg,...) • • • K - l) 0
0
(M3,...) • • • Kt-i )
0

(3.3)
That is to say, the product of any two M0-elements belongs to
U(1) and hence may involve ^-elements, u^-elementa, etc., but not
«0-elements; the product of a u0 and a ux may involve w2's, u3's,
etc., but not MO'S or u^s; and so on.
(3) U(1>, U(2),..., U'01-1' are invariant sub-algebras of X; so, of course,
is U. This gives us the further information that
A(u0) = (u0, ut,...), A(ux) = {uvu2,...), A(u2) = (u2,u3,...),....
(3.4)
• (5), Theorem II.
4 I. M. H. ETHERESTGTON
That is to say, the product of A and any ug may involve the Ug's,
ug+l's, Ug+2'a,..., b u t not the Ug_x's, v,g_2'a,... .
Consider now each ug set of base elements as forming a column
vector; and write the multiplication rules (3.4) in the form
Aug = PgUg+QgUg+1+RgUg+2+..., (3.5)
where each Pg is a square matrix and Qg, Bg,... are in general rect-
angular matrices.
A linear transformation of X affecting only the ug'a, say
Vg = Hug, Ug = H^Vg, (3.6)
transforms (3.5) into
Avg = HPgH-lVg + HQgUe+1+HRgUgf2+..., (3.7)
1
and thus induces a collineatory transformation (HPgH- ) of the
matrix Pg. Conversely, any such transformation of Pg corresponds
to a linear transformation of the us's.
Take H to be the matrix which reduces Pg to its classical canonical
form.* Let us suppose that the corresponding transformation (3.6)
has been carried out, and that this has been done for each of the
a sets of nil base elements. We shall then say that the special train
algebra X has been given its canonical form.
In what follows, it is not essential that the above reduction shall
have been carried out completely. I t is sufficient if each Pg has been
reduced to the Jacobian form.f That is to say, the essential thing
is that the latent roots of each matrix Pg shall appear down its main
diagonal, with zeros below the diagonal;
Let us return to the original notation A, u" (a = 1,..., n) for the
basis of X, without, however, changing the order in which the nil
base elements have been placed. The multiplication table of X has
now the form
, Aua = XOW+..., u"uT = O.W-\-0.ur+..., (3.8)
where in each equation the terms omitted involve only u'a with
higher suffixes than those written down. (The Xa are the latent roots,
n in all, of the a matrices Po, Pv..., Pa-i) Conversely, it is easy to
see that a commutative linear algebra whose multiplication table
has this form has the defining properties of a special train algebra.
» (7), Ch. VI. t (7), 64.
 SPECIAL TRAIN ALGEBRAS 5
4. Characteristic and rank equations of a special train
algebra
Consider a commutative special train algebra X in the canonical
(or partially reduced) form described above. Let its general element
be written as > A . ^* a IA I \
x = {A+ Z<xau", (4.1)
| being its weight. Then, using (3.8),
xA = £A + ..., (4.2)
xu< = £\uT+...,
so that
0 = (t-z)A + ... )
0= (fa-xjvH-- (4-3)
0= (£X2-x)u2+... etc. J
The characteristic equation* of the algebra is obtained by equating
to zero x times the determinant of this set of equations; it is therefore
x(x-t)'x-\J)...{x-\l€) = 0. (4.4)
Powers of x in the expanded form of this equation are to be inter-
preted as principal powers.f
If the rank equation of X is f(x) = 0, then it is knownj that f(x)
must be a factor of the left side of the characteristic equation (4.4).
The rank equation therefore has the form
x(x-£)(x-(tli)...(x-l*T£) = 0, (4-5)
where y^,..., fir are included in Alv.., A^ X has thus the essential
property of a train algebra, and the theorem has been proved.
When multiplication is non-commutative, the third defining
postulate of a special train algebra means both
XU<«<U<» (0 = 1,..., « - 1 ) (4.6)
and U0>X < U<*> (6 = 1,..., a - 1 ) . (4.7)
If only (4.6) is assumed, the.algebra may be called a left special train
algebra; the analysis as given leads to a left rank equation, and we
find that X is a left train algebra. Similarly, assuming only (4.7),
the algebra is a right special train algebra; and, using products of the
type uA instead of Au, we can show that X is a right train algebra.
When both (4.6) and (4.7) hold, the two canonical forms will not
necessarily coincide; and in fact the left and light ranks need not be
equal.
* (2), § 15, Theorem 3. t (4). § 2. % (1), § 3, Cor. II.
6 I. M. H. ETHBRINGTON
5. An example from genetics*
In (4) § 14 I considered the genetic algebra for the gametic types
depending on three linked series of multiple allelomorphs, having the
multiplication table (4) (14.3). The transformation which was applied
to this had the effect, it will now be seen, of giving this commutative
special train algebra its canonical form. Writing u, v, w, p, q, r, 8
for u, v, w, vw, wu, uv, uvw, and 6, <$>, ijt for 1— OJBC, \—u)AC, l—coAB,
the transformed multiplication table is
u v w p
I 1)U \V \w WP iAs
u 0 %o)ABr \otACq 0 0 0
v 0 1°>BCP 0 b>s 0 0
w 0 0 0 §P S 0
P> q> r> s 0 0 0 0
In the notation of §3, u, v, w form the u0 set of base elements;
p, q, r are ^-elements; and 8 is u2. The characteristic equation (4.4)
is seen to be

The rank equation is found as follows. Consider the general

e emen x
s i n p> qt r> 3>

whose square is
x2 = £ 2 /+fa%+f/to+£yu>+terms in^>, q, r, s.
We have
x(x—(;) = x2—£z = K^J-|-terms in q, r, a (say),
x{x—£)x = £ * . ^ p + t e r m s in q, r, a.
Hence x(x—£)(x—\d£) = terms in q, r, s.
Similarly, a:(a:-£)(a:-40f)(*-#£) = terms in r, a,
x{x-t;)(x-$6t;)(x—i<f>£)(z—$tl>t;) = a multiple of s,
x{x-£)(x-m){x-WZ){x-m){x-\U) = 0. (5.2)
By considering particular elements such as I-\-u, I-\-p, I+q, I+r,
7+s, we can show that no factor in this equation is superfluous, and
• The statement of (4), 247, that all the fundamental symmetrical genetic
algebras are special train algebras, refers to the gametic algebras, not to the
zygotic algebras which are derived from them by duplication. The latter are
train algebras but not in all cases special train algebras.
 SPECIAL TRAIN ALGEBRAS 7
it is thus the rank equation of the algebra. The principal train roots
are therefore 1, \9, \$>, \\\>, \\.
The plenary rank equation,* or equation of lowest degree con-
necting plenary powers (a^ml = x2""') can be obtained most simply
by use of annulling polynomials, as follows. Consider for simplicity
only a normalized element
X = I+vu+Pv+yw+8p+€q+tr+r}S. (5.3)
Let OX denote X2, so that
OX = I+<xu+pv+yw+(89+pycoBC)p+(...)q+(...)r+
+ (ijA+«8/*+j36v+y^)«. (5.4)
The operator O is to be considered as acting only on the coefficients
of X. Its effect on any one coefficient, of course, depends on the
values of the other coefficients as well. Thus
Oa = a, Oj9 = £, Oy = y \
08 = 89+PywBC, ..., ... . (5.5)
O77 = rjA+aS/t-f/Sev+y^p* J
Therefore (O— l)a = 0; or, as we may say, the polynomial
O-l (5.6)
.unuls a; it also annuls /? and y. Also (O—0)8 = f$yu)BC, which is
annulled by 0—1. Thus, of the polynomials
(<D_1)(<]>_0), (p-i)(p-fit (<D_i)(O-^), (5.7)
the first annuls 8, and similarly the other two annul respectively
eand£. Also (G-A)ij = aSjt+jfcv+yCp.
Each term on the right is.annulled by one of the operators (5.7);
hence (O—A)rj is annulled by the L.C.M. of these operators, and 17 by
(O-l)(O-0)(O-<£)(O-./»)(O-A). (5.8)
Finally, the annulling polynomial for X itself is the L.C.M. of the five
operators (5.6), (5.7), (5.8). Thus we have the plenary train equation
(O—1)(O—0)(O—^)(O-^)(O—A)X = 0, (5.9)
i.e. X[X-i][X-6][X-4>JX-$\[X-\] = 0, (5.10)
in which, after expansion, symbolic powers of X are to be interpreted
as plenary powers.
We have shown, then, that the sequence of plenary powers forms
a train, and that the plenary train roots are 1, 9, <j>, ip, A.
The genetic interpretation of train roots has already been given.f
* See (4), 246-7. t (4), 247.
8 SPECIAL TRAIN ALGEBRAS
Added 16 December 1940. It has been tacitly assumed in § 5 that
8, <f>, >p, 1 are all unequal. Genetically, we may suppose without loss
of generality that the loci A, B, C are distinct and occur in that
order on a chromosome, with
0
< ">BC ^ ^AB < WAC < i-
Then it may be shown that
i < A < 0 < 0 < 0 < 1.
Thus an exception to the tacit assumption occurs when ioBC = coAB;
in this case ip = 6, the equations (5.2), (5.10) contain repeated
factors, and the repetitions are superfluous.

REFERENCES
1. L. E. Dickson, 'Linear algebras': Trans. American Math. Soc. 13 (1912),
59-73.
2. Linear Algebras, Cambridge Tract No. 16, 1914 (reprinted 1930).
3. I. M. H. Etherington, 'On non-associative combinations': Proc. Roy. Soc.
Edinburgh, 59 (1939), 153-62.
4. 'Genetic algebras': ibid. 242-58.
5. 'Commutative train algebras of ranks 2 and 3 ' : J. of London Math.
Soc. 15 (1940), 136-49.
6. 'Duplication of linear algebras': Proc. Edinburgh Math. Soc. (2) 6.
(In press).
7. H. W. Turnbull and A. C. Aitken, Introduction to the Theory of Canonical
Matrices, 1932.
8. J. H. M. Wedderburn, 'On hypercomplex numbers': Proc. London Math.
Soc. (2) 6 (1908), 77-118.