PSYCHOMETRIKA~VOL. 5, NO.

2
JUNE, 1940

A COMPUTATIONAL PROCEDURE FOR T H E METHOD

OF P R I N C I P A L COMPONENTS
MERRILL M. FLOOD
PRINCETON U N IVEP~SITY

An n-rowed correlation matrix may be thought of as an ellipsoid in n-dimensional space with its center at the origin. The principal components of the matrix are essentially the semi-axes of the
ellipsoid. A direct and simple method of computing the lengths and
directions of these semi-axes is presented.
This p a p e r makes no a t t e m p t to p r e s e n t a comparison of the
m a n y computational methods which have been proposed for use in thc
application of the method of principal components. Hotelling introduced (1) this method of statistical analysis and has offered a simplified computational m e t h o d (2). O t h e r w r i t e r s , including F r i s c h (3),
B u r r (4), and T h u r s t o n e (5), have considered the general problem of
" f a c t o r analysis," have discussed the method of principal components
in its mathematical, computational, and statistical aspects, and have
c o m p a r e d this method with o t h e r s suggested f o r use in applied problems of " f a c t o r analysis." 0 n l y the question o f computational procedure f o r the method of principal components is considered here.
Stated briefly, the problem with which we are concerned is the
practical d e t e r m i n a t i o n o f the latent roots and i n v a r i a n t vectors* o f
a m a t r i x M . In the p a r t i c u l a r case o f statistical i m p o r t a n c e the mat r i x M is the real, symmetric, positive definite correlation m a t r i x R of
o r d e r v . T h e l a t e n t roots x~, and i n v a r i a n t vectors d~ are the solutions of the m a t r i x equation
(R--x)d=O,

(1)

w h e r e x is a scalar m a t r i x and d is a m a t r i x (vector) with only one

column.
The c h a r a c t e r i s t i c function of R is the scalar polynomial f ( x ) defined by
/ ( x ) --- ix - - RI = / / ( x

- - x,) = x v -

i=l

E a~x h

(2)

h'-o

For the matric terminology, notation, and theorems used in this paper, see
W'edderburn's Lectures on Matrices (6), particularly Chaps. 1, 2, 3, and 6. An
elegant presentation of the method of principal components, using the algebra of
matrices, has been given by Householder and Young (7).
--169--

170

PSYCHOMETRIKA

It is well known that R satisfies its characteristic equation, t h a t is

U-1

(3)

f ( R ) = R~'--F~ a~R h = 0 .
h=o

If the roots x~ of R are distinct, then f ( x ) is the only polynomial of

degree v , and leading coefficient unity, having R as a matric root.
This characteristic function m a y be computed, therefore, by solving
the v ~ linear equations of (3) above, for the values of the v coefficients
ah. The roots x~ m a y then be found by applying Horner's method, f o r
example, to the polynomial f ( x ) . Finally, the components of each inv a r i a n t vector d~ m a y be computed by solving the set of v linear homogeneous equations
( R - - x ~ ) di = 0 .

(4)

This completes the mathematical exposition of the computational

method proposed in this paper for the practical determination of the
latent roots and invariant vectors of a correlation matrix.
A n example will illustrate the method of computation. F o r this
purpose, I shall analyze the m a t r i x R* below, which was used by Hotelling (2) to illustrate his simplified method for the calculation of
latent roots and invariant vectors (principal components).
R* =

I 1.
.9596
I .7686
.5427

.9596
1.
.8647
.7005

.7686
.8647
1.
.8230

.5427
.7005
.8230
1.

(5)

The calculations may be simplified at the outset by considering the

m a t r i x R = R e - - 1 which has the same invariant vectors as R* but
latent roots each reduced by unity.
Of the v 2 linear equations (3) it is only necessary to use v , say
those shown as the first column of f ( R ) . These equations m a y be
found without actually raising R to successive powers since only the
first column of each power is required; thus R'el = R ( R e l ) , R3el =
R (R2e,), and so on. The result is the set of linear equations
el

1
0
0
0

Rel

0
.9596
.7686
.5427

R2el

1.80610141
1.04476977
1.27640822
1.30475760

R3e~

2.69170038t[
3.75082780 ~|
3.36539747 I]
2.76251642 JL

R4e~

ao ~
[ 7.68515651
a~ __ | 7.42815763
a~ ]
]7.58573273
a3
[ 6.85796278

(6}

Actually, it is unnecessary to include R3e~, for a3 must be zero since

MERRILL M. FLOOD

171

the trace of R is zero. This leaves three equations to be solved for

the coefficients ao, a,, and a2 ; the complete solution being
[

ao- l
a~

1@ 0.96314191
@ 3.68871936

The greatest root of the polynomial

f (,~:) = x ~ - - ~ x ~ - - %x ~ ~o
is found to be x, = 2.33972. The invariant vector d~ corresponding to
the root x~ is determined by the equations

dl

R~xl
--2.33972
.9596
.7686
.5427

.9596
--2.33972
.8647
.7005

.7686
.8647
--2.33972
.8230

@.5427 7
@.7005
1
@.8230 |
--2.33972 ]

d..,1
d, 1

d,,1

= 0,

A solution of these four linear homogeneous equations is found to be

__

d4~

1.000000

.977399
.858967

These components of d~ agree, to four decimal places, with those

found by Hotelling, and the value 1 @ x, is the same as that found by
Hotelling for R*. The other latent roots and invariant vectors m a y be
found in similar fashion.
A number of questions arise concerning this method of computation; for example, under w h a t conditions:
1. Is it f a s t e r than other methods?
2. Is the matrix (el Re~ ... Rv-~el) non-singular?
3. Are the results accurate to the required number of
decimal places ?
There is the f u r t h e r question concerning the case in which some of
the roots are equal, or nearly equal. I hope to examine these questions in another paper and extend the method for use in analyzing any
normal* matrix M .

* A matrix M is said to be "normal" if it commutes with its transpose.

172

PSYCHOMETRIKA
BIBLIOGRAPHY

1.
2.
3.
4.
5.
6.
7.

Hotelling, H. "Analysis of a Complex of Statistical Variables into Principal

Components," Journ. Edue, Psychol., 1933, 24, pp. 417-441, 498-520.
Hotelling, H. "Simplified Calculation of Principal Components," Psychometrika, 1936, l, pp. 27-35.
Frisch, Ragnar. Statistical Conflue~tce Analysis by Means of Complete Regression Systems, Oslo: 1934.
Burt, C. "The Unit Hierarchy and its Properties," Psychometrika, 1938, 3,
pp. 151-168.
Thurstone, L. L. The Vectors of Mind, Chicago: University of Chicago Press,
1935.
Wedderburn, J. H. M. Lectures on Mat~'ices, New York: Amer. Math. Soc.
Coll. Publications, 1934.
Householder, A. S. and Young, Gale. "Matrix Approximation and Latent
Roots," Amer. Math. Monthly, 1938, 45, pp, 165-171.