On the Geographical Interpretation of Eigenvalues

Author(s): P. R. Gould
Source: Transactions of the Institute of British Geographers, No. 42 (Dec., 1967), pp. 53-86
Published by: Royal Geographical Society (with the Institute of British Geographers)
Stable URL: http://www.jstor.org/stable/621372
Accessed: 18-12-2015 09:47 UTC

REFERENCES
Linked references are available on JSTOR for this article:
http://www.jstor.org/stable/621372?seq=1&cid=pdf-reference#references_tab_contents

You may need to log in to JSTOR to access the linked references.

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at http://www.jstor.org/page/
info/about/policies/terms.jsp

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content
in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship.
For more information about JSTOR, please contact [email protected].

Wiley and Royal Geographical Society (with the Institute of British Geographers) are collaborating with JSTOR to
digitize, preserve and extend access to Transactions of the Institute of British Geographers.

http://www.jstor.org

This content downloaded from 202.28.191.34 on Fri, 18 Dec 2015 09:47:29 UTC
All use subject to JSTOR Terms and Conditions
 On theGeographical
Interpretation
ofEigenvalues
P. R. GOULD, M.A., PH.D.

(Associate
Professor The Pennsylvania
of Geography, StateUniversity)
RevisedMS. received
o1 March1967

(Communicated P. Haggett)
byProfessor

IN THE historyof thedevelopment of all scientificfields,a largenumberof instances can be

foundwherea man has takeninformation availableto all and has reordered it in particular
waysto illuminate an area thatwas previously darkand obscure.Numerousexamplesspring
to mind,rangingfromtheclassicalattempts of Linnaeusto orderinformation on plantsand
animalsby imposinga systemof classification, to W. Heisenberg'sorderingof knownfacts
aboutatomicstructure (B. Hoffman, 1963;Heisenberg, 1930).The latterexampleisparticularly
pertinent to the aim of thisexpository paper, for theorderingof information by Heisenberg
was ofa particular type.Taking theknown frequencies ofelectrons,whichhadbeenpreviously
expressedin theformof a listknownas theBalmer-Ritzladder,Heisenbergrearranged the
sameinformation intoa squarearraycalleda matrix.Operationson thematrixled to some
startlingand illuminating results.Not onlyweresometheoretical underpinnings providedfor
theresidualhalf-quantum ofenergythatremainedin an oscillating but
particle, some numbers
characteristic of the matrices,called eigenvalues,were interpreted as the same excitation
frequencies as theFraunhofer lines,thosecriticalvalueswe can actuallyseein thespectrum of
visiblelightas thindarkbands.
Heisenberg'sanalysis,showingpatternand orderwherenone was seen before,is not the
firstexampleofarraying datain matrixform.Mendelejeff's arrangement oftheperiodictable
in 1869,WilliamHamilton'sdiscovery in I835 of thepowerof thequaternion in mechanics,
and Laplace'sstudiesin I772 on theperturbations ofplanetary orbitsseta tradition of matrix
orderingand manipulation manyyearsbefore.In a verydifferent sense,and with a very
different purposein mind,orderingthevariouspiecesof an academicdisciplinein a matrix
arrayhas also pointedup pattern and structure thathas been usefuland thoughtprovoking.
J. Berliner'sexamination of anthropology fromthepointof view of an economist,E. Hall's
delightful 'Map' of culture,and B. J. L. Berry'sthree-modematrixof geographyare all
examples(Berliner,1962; Hall, I959; Berry,1964).
Frequently, geographicalfactsmay also be collectedor arrangedin thosesquareand
rectangular arrayscalled matrices.By manipulating such orderingsof readilyavailablein-
formation in particular we
ways, may sometimes be able to uncoverpattern and orderin very
complex situations where many variables enter a problem and compete for our attention.
Implicit in such manipulative of
exercises, course, is the belief that every scientistholds-
thatorderthereis,ifonlywe can unscramble thechaosthatliesall aroundus. Holdingto this
act offaith,metaphysical as it mustbe, we shalltryto see,first, ifwe can shakeout of square
arrays of datasome veryparticular numerical values,and then tryto givethemsomemeaning-
fulgeographical interpretation.
53

This content downloaded from 202.28.191.34 on Fri, 18 Dec 2015 09:47:29 UTC
All use subject to JSTOR Terms and Conditions
54 P. R. GOULD
Eigenvalues
The numericalvaluesconcernedaretheeigenvalues, latent roots valuesof a
or characteristic
square matrix (the termsare interchangeable) and in order to grasp what they we shall
are
seektheaid of geometry. It is no accidentthattheancientand honourablelinkagesbetween
geometry andgeography arebeingemphasizedonceagainafteran unfortunate lapseofseveral
millennia, and the geographer, with his particular outlook
spatial and mind,frequently finds
thatinsights froma fieldso closelyalliedto hisown areespecially helpful.
First,letus notethateverysquarematrixhasan equationassociated with,or characteristic
of it.The equationwillbe in theformofa polynomialofthesameorderas thatof thematrix,
so thata 2 X2 matrixwill have associatedwithit a quadraticcharacteristic equation,whilea
3 x 3 will have a cubic,and so on. It is withtheroots,and hencethe term'latentroot',of
thesecharacteristic equationsthatwe shallbe concernedthroughout thispaper.It is worth
noting that everysquare matrix alwayshas as many roots as itsorder, althoughsomeof the
rootsmaybe zero. In somecases,calledGramianmatrices, we can provethattherootswill
all be realnumbersand we can determine thisproperty by simpleinspection. Sometimeswe
can findout thatthematrixhas certainspecialcharacteristics so thattherootsof itsequation
are not onlyrealbut positiveas well. In othercases,particularly matricesthatare not sym-
metrical, we cannotdetermine theseproperties but
by inspection, onlyby trying to findall
real rootsand thenfailing.While methodsdo existforfindingtherootsor eigenvaluesof
unsymmetrical matrices,the algorithms, or the step-by-step procedureswhich lead to a
desiredsolution,breakdownwhenwe encounter imaginary rootsin anypolynomialequation
beyondthequarticand, so faras is known(1967),no computerprogramexistsforevalua-
tingnon-symmetric matrices withimaginary rootswhoseorderis greaterthanfour.
Beforewe tacklethequestionofthecharacteristic equation,we shallneeda basicbuilding
blockcalledthedeterminant. Usuallythisis an aridthingforthenon-mathematician fromwhich
a singlenumberis squeezedby muchalgebraicmanipulation (R. V. Andree,1959;P. J.Davis,
1965; W. G. Bickleyand R. S. H. G. Thompson,1964; G. Kron,I959; B. Higman,1964;
R. A. Frazer,W. J.Duncanand A. R. Collar,1963).Let us examineit witha littlemorecare,
starting witha simple2 x 2 matrixas an example.
As a matterof notation,matrices are alwaysenclosedby curvedbrackets. Determinants
are simplythesamematrices enclosedby squarebrackets so thatif:

(
/I 3 I 3
A= ofA, or detA =
thenthedeterminant
\4 6 46
We now need some rulesto calculatethevalue of sucha squarearray.For a 2 x 2 matrix,
betweentheproductofthediagonals,so that:
we simplytakethedifference
3 I
detA= = (3 x 4)-( x 6) = 6
6 4
In thisparticularcase,thevalueof thedeterminant is positive,and thematrixis calledpositive
If
definite. thedeterminantis equal to zero thematrix A deter-
is calledpositivesemi-definite.
minantmayalsobe negative.
Taking a 3 x 3 matrixas a further example,we shouldnote that,while the rulesfor
extracting the determinantvalues are simple,theybecome tediousfromthe manipulative

This content downloaded from 202.28.191.34 on Fri, 18 Dec 2015 09:47:29 UTC
All use subject to JSTOR Terms and Conditions
GEOGRAPHICAL INTERPRETATION OF EIGENVALUES 55
pointof view as matricesbecomelarger.In fact,formostappliedproblems,wherethesize
of a matrixnearlyalwaysexceeds4 x 4, a determinant
would alwaysbe evaluatedtodayby
a computer.For a 3 x 3, then,let:
3 I 4\ 3I 4
A= 2 3 so detA = I 2 3
4 I I 4 1
Thereareseveralwaysin whichwe can evaluatedetA, buttheeasiestforourpresent
purposes
is by simplyrewriting
thefirstand secondcolumnsto therightof thematrixand evaluating
thedeterminantbythefollowingrule(F. E. Walker,1963):
all a12 a13 11 a12

detA= a2
\\ a><
a / (all a22 a33)+(a12 a23 a3l)+
a22 a23 a21 a22 = (a3 a21 a32)-(a31 a22 a13)-

/ aia \\3 a32 (a32 a23 all)-(a33 a21 al2)

a31 a32 a33 a31

So thatin theexampleabove:
31 4 31
detA = I 2 3 I 2 = (6+12+4)-(32+9+I) = -20
4'' I
4 I 41
4 I

Note thatifanyrow or columnconsistsentirely is zero. This

of zerosthenthedeterminant
is obviousfromthe 3 x 3 example,sincewe would be multiplyingthroughby zero. Also,
ifanyrow or columncan be constructedfroma linearcombinationofanyother,then,again,
thedeterminantwill be zero.For example,let:
4 92
A= 2 I
6 3 3
and noticethatcolumnsi and 3 aredirectly so thatwe can write
to one another,
proportional
thelinearrelationship:
Y= bX
or Column I = 2 (Column 3)

Thereis lineardependence
present(Higman,I964, 14-16) and to confirmthiswe evaluatethe
determinant:
4 9 2 4 9
detA = 2 I I 2 I =(I2+54+I2)-(I2+I2+54) =
6 3 3 6 3
Lineardependencein thiscase impliesa simplelinearproportion. However,it does not mean
thatifwe derivecolumnI fromthetypicalY = a + bX relationship thatthedeterminant will
will it be zeroifa columnor row is in a Y = X", or non-linear,
be zero.Nor, tautologically,
withanother.
relationship
However,thesebasicrulesand properties about determinantsdo not reallytellus what
they are. Let us see if geometry can help (P. J. Davis, I965, 194-200). Suppose we have a

This content downloaded from 202.28.191.34 on Fri, 18 Dec 2015 09:47:29 UTC
All use subject to JSTOR Terms and Conditions
56 P. R. GOULD
are (3, 8), (I, 2) and (8, 4) (Fig. I). To thiswe add thevertical
trianglewhose co-ordinates
linesBD, AE and CF. Remembering
construction thattheareaofa trapezoidis:

Area= (Sumoftheparallelsides)(altitude)
2

we can findthearea of thetriangleABC by addingtheareasofAEDB and ACFE and sub-

from
tracting thetotalarea of the polygonACFDB theareaBCFD.
resulting
So ABC = AEDB+ACFE-BCFD
(2+8) (3-I) (8+4) (8-3) (2+4)(8-I)
2 2 2
= I9

(8,4)
I
I
B II
I
I
v F
v
D E F
- A
X
FIGUREI-The triangledefinedby threerow FIGURE
2-The volumeof thetriangle
withunitheight.
vectors.

(2+7=9, 6+2=8)

Ba-' x
FIGURE3-The geometrical ofa determinant.
expression

The important ABC maybe evaluatedas a determinant.

thingis thatthisareaof thetriangle
Ifwe arraytheco-ordinates andadd a columnofones,then:
ofthetriangle,
3 8 i
A= I 2 I
8 4 I

This content downloaded from 202.28.191.34 on Fri, 18 Dec 2015 09:47:29 UTC
All use subject to JSTOR Terms and Conditions
GEOGRAPHICAL INTERPRETATION OF EIGENVALUES 57
ABC willbe one-halfthedeterminant
The areaofthetriangle ofA, whichis:
38138
3 8 I 3 8 38
I/2 detA= I 2 I I 2 = I/2 [(6+64+4)-(I6+I2+8)] -= 19
2
8 4 84
Two questionsarise.First,where did the columnof ones come from,and what do they
mean in termsof the geometryof theproblem?In fact,we are evaluatingthe volumeof a
three-dimensional trianglewhoseheightis unity(Fig. 2). This,of course,is thesame as the
'square' measure. Secondly,whyshouldtheareabe ? detA? If we move ouraxesto thepoint
B on thetriangle ABC, thentheco-ordinates ofB will be (o, o), theco-ordinatesofA will be
(3-I, 8-2) or (2, 6) and thoseof C will be (8-I, 4-2) or (7, 2). We can thinkof thelinesBA
andBC as vectorsV1and V2whoseresolution is thelineBD (Fig. 3). In evaluatingthedeter-
minant, then,we arecalculatingthearea ('unit'volume)of theparallelogram ABCD, so that
ourtriangle ABC hasexactlyone-halfthearea,or I detA.
We notedearlierthateverysquarematrixhada polynomial equationthatwascharacteristic
of it, and it is necessaryto look at thismoreclosely.The characteristic equationof a square
matrixA is definedas:
det(A-2I) = o

I is simplytheidentitymatrix(thematrixofI's in theprincipaldiagonalandzeroselsewhere),
multiplied by the scalar matrixA,so we may thinkof thecharacteristic equationas simply
a determinant, with A's subtracted
fromeach diagonalelement,thathas been set equal to
zero. Thus:
all- a12 aln
a2l -- a22 - a2
a31 a32 a33-A ..* a3n
=

an 1 an2 ** ann A- A

will be a polynomialequationin n, theorder(nx n) of thematrix.Let

But thisdeterminant
us take a 2 x 2 example. Suppose we have:
2 4 2-2 4
A = then det(A-AI) = o is det = o
S5 I s5 i-A
we have:
Multiplying,
(2-A) (I-A)-(5 X 4) = o
or
;2_32-I8 = o

a quadraticwithroots,eigenvaluesor characteristic
valuesof:

31 /9+72 so
o =6.oo
2 A2 =- 3.00

Althoughtheseroots are real numbers,one of them is negative.This can be foundby

A, sinceitsdeterminant
inspecting is negative,showingthatitis nota positivedefinite
matrix.
E

This content downloaded from 202.28.191.34 on Fri, 18 Dec 2015 09:47:29 UTC
All use subject to JSTOR Terms and Conditions
58 P. R. GOULD
Frequentlywe shallbe interestedin theeigenvaluesof symmetrical matrices,wherethe
off-diagonalelementsare decimal numbers and the diagonalelements are unity.It can be
shownthatall symmetrical matriceshave realeigenvalues, and thatif theyare also positive
definite will be positive.For example:
theeigenvalues
/I.00 0.75 o.83\
A = 0.75 I.0 0.4I
\.83 o.4I I.oo/
has thecharacteristic
equation:
- A3+ 322- .58052+0.0909 = o
a cubicwithroots:
,1 = 2.3416
22= 0.5929
23 = o.o655
All thesearerealand positive,so A mustbe positivedefinite.
Noticethattheeigenvalues sum
to 3, the orderof the matrix,the orderof the characteristicequationand the sum of the
elements along themaindiagonal known as the of
trace thematrix.
Untilnow we haveconfined ourselves Let us turnto geometry
to algebraicmanipulation.
again and deal with2 2 matrices forsimplicity.Supposewe have thematrix:
/4 84
A(=
8 4
We could thinkof therow or columnelements(we shalldeal withrowsin theseexamples)
ofpointsin a planewhichdefinetwo vectorsfromtheorigin(Fig.4).
as co-ordinates
The matrixA hasthecharacteristic
equation:
2_ 82-48 = o
with roots 21 = I2
22 = -4

-~/ ,'~ -What do thevaluesofA2and22 really mean?

#/
/ Perhaps the best to
way grasp their meaning
,/,,,"/ , is to thinkof themas 'stretchability coeffi-
/ 'cients'.
(8,) Imagine our two vectors as a small
/ / .'cluster' surrounded by an ellipse. If the
\-^s~ '~/
/,,/
/,^ vectors are close together we could 'fit'a
"
',- ,"'____ very elongated ellipse along the primary
-, /axis, whose widestpointwould be shown
FIGURE4-The 'stretchability' from
(dottedlines)resulting by the lengthof the secondaryaxis. In a
two row vectors(solidlines). crudesense,2Ais theamountofstretch along
thefirstaxisand 22 istheamountof stretch
alongthesecond.
Ifwe move V1closerand closerto V2 (or viceversa)we would endup with:
/4 8
A=I
4 8

This content downloaded from 202.28.191.34 on Fri, 18 Dec 2015 09:47:29 UTC
All use subject to JSTOR Terms and Conditions
GEOGRAPHICAL INTERPRETATION OF EIGENVALUES 59
equationis:
whosecharacteristic
A2- I22+0 = 0
SO 21 = 12
22= 0

of our ellipseis stillI2 along theprincipalaxis,but is o along thesecond,

The stretchability
becauseour 'ellipse'has collapsedto a linewhich,of course,has zerowidth.Supposewe had
two vectorsorthogonal, or at rightanglesto one another(Fig.5). Say:
-4 8
A=

equationis:
whosecharacteristic
,2 +o--8o = o
witheigenvalues:
A1 = + 8.95
22 -8.95
Y

(-4,8) \

J.\
\ (8,4)

X---
0
figureor 'ellipse'fromorthogonal
FIGURE5-The circular
vectors.

As we mightexpect,whenthevectorsof equal lengthareat right-angles to one another,our

ellipseis stretched
the same amount in both and
directions, so becomesa circlewiththesame
stretchabilityindicesforits axes.

TheEigenvectors Associated withtheEigenvalues

Associatedwitheacheigenvalue, or stretchability
coefficientofa matrix,is an eigenvector.
Mathematicallyit is a vectorwhosedirection is notalteredby beingmultiplied by theoriginal
matrix,whichis thoughtof in thiscontextas an operator(Higman,1964,I6-I7). A linear
operatoris simplya matrixthatcan alterthevalue of a vector.For example,if a groupof
Polynesianfarmers pioneering in Alaskachooseto growbreadfruit and rejectrye,we might
their
describe choice the
during year of colonization (to) as a vector:
breadfruit-rye
I.0
V
Vbr=
\I.o0
0.0

If MotherNaturerewardsthechoiceof rye,but punishestheculturally choiceof

influenced

This content downloaded from 202.28.191.34 on Fri, 18 Dec 2015 09:47:29 UTC
All use subject to JSTOR Terms and Conditions
6o P. R. GOULD
we mightexpectsome of thefarmers
breadfruit, to learna lessonand makea betterchoice
thenextyear(R. Bushand F. Mosteller,I955). Thus,letA be thematrixoperator:
/o.6 o.3
A=
0.4 0.7
whichmultiplies
thebreadfruit-rye
vectorat successive
periods(Fig. 6).
to tl t2 t3
/o.6 0.3 /I.o Breadfruit o.6o /o.48\ 0.444\

\o0.4 0.7/ \o.o Rye \0.40/ \0.52 \0.5556/

t4 t5 t6
/o.4332\ /0.42996\ /o.428988\

\o0.5668/ \0.57004/ = \o0.57IoI2!

Herewe arechangingthedirectionof thevector,so theoriginalVb,was notone of theeigen-

vectorsoftheoperatorA. But noticehow theelements of Vbrareconvergingon thevalues:
o0.428\

Vbr = 0.72
o.572
andhow thesuccessive vectorsaregradually
swingingtowardsthisfixedpoint.Eventu-
ally,accordingto thisrathernaivemodel,
thefarmers will grow 57 per centrye,but
therewillalwaysbe a groupinthePolynesian
farming profession whichresistschangeand
continues to pineforbread-fruit.Suchresis-
tanceto new ideasis notunknownin other
professions. Let us take some furtherex-
oY^~ -~o
v?
o?, ~ Breadfruit br 1.0 amples, and continue to consider the
FIGURE6-A simplelearningmodel: thebreadfruit-rye as
geometrical, well as thealgebraic,aspects
vectorconverging
upontheprincipal
eigenvector. (Fig.7). If:
4 8
A=
\8 4/
equationis:
thecharacteristic
2_ 8,-48 = o
withroots:
AI = I2

Associatedwith each of theseeigenvaluesis an eigenvector:once again, a vectorwhose

by thematrix,althoughitslengthmaybe scaledin
is not alteredby beingmultiplied
direction

This content downloaded from 202.28.191.34 on Fri, 18 Dec 2015 09:47:29 UTC
All use subject to JSTOR Terms and Conditions
GEOGRAPHICAL INTERPRETATION OF EIGENVALUES

of thesesimpleexamplesby substituting
or out. We can evaluatetheeigenvectors theeigen-
valuesin turn.Thus,when2i = 12:
4V1+8V2 = I2V1
8v1 +4v2 = I2V2
or 8v1 = 8v2
v1 8 I
so -
v2 8 I
Associatedwiththelargesteigenvalueis a vectorwhoseelementsarealwaysin theratioI: I.
This is a line fromtheoriginwitha slope of I, and in thisparticular
case it bisectsthetwo
vectors(4, 8) and (8,4). Similarly,associated with the smallesteigenvalue (A = -4), we have:
4V1+ 8V2 = -4v1
8v/+4/2 = -4V2
or 8v1 = -8v2
so 21 -I
V2 I
Y

(4,8)

v1 -?-
(8,4)

FIGURE7-The slopesof the eigenvectors

(heavylines)
fromtherowvectors.
resulting
This is a vectorwitha slopeof- I, so it is clearlyat right-angles
to thefirst.This property
oforthogonality willalwaysholdforsymmetric matrices.
Returning to ourPolynesian
farmers in Alaska,we havetheoperator:
0o.6 o.3
A=
0.4 0.7
equationis:
whosecharacteristic
22-1.302+0.30 = o
witheigenvalues:
a" = I.0
12 = 0.3
are:
The eigenvectors
o.6v1+o.3/2 = i.oVi o.6v1 +o.3v2 = 0.3V
0.4v1+0.7v2 = I.ov2 and o.4v x 0.7v2 = 0.3V2
v1 0.3 v1 0.3
so -- -so
s
/2 0.4 v2 0.3

This content downloaded from 202.28.191.34 on Fri, 18 Dec 2015 09:47:29 UTC
All use subject to JSTOR Terms and Conditions
62 P. R. GOULD
Notice,withrespectto thefirsteigenvector,thatifwe take3/7and 4/7of Ioo percent,we
obtain42.8 percentand 57.2 percent,the'steady-state'
proportionof thebreadfruit
and rye
growers.

The Geographical Interpretation

ofEigenvaluesandEigenvectors

Havingindicatedthateverysquarearrayof information maybe considered as a matrix,

forwhichthereis a characteristic polynomialequationwhoseroots,or eigenvalues, maybe of
morethanpassinginterest, letus returnto theoriginal,expository intentofthispaperand see
how thesesummary valuesmightbe interpreted in a geographicalcontext.
As geographers we arecontinually dealingwithmaps.As a compression, abstraction and
simplificationof reality(D. Hawkins,1964; M. Polyani,1963; S. Toulmin,I960), themap is
our basicmodel,a way of selectingcertainfeatures fromthechaoticworldaroundus while
stillkeepingthemin theirspatialcontext.Yet themap allowsus fewopportunities formani-
pulating itsinformation directlyso thatwe may extend our understanding about the pheno-
menaor theregionportrayed. To makesuchextensions we can alwayspose questionsabout
thepatterns and structureswe observeon themap and speculateabouttheunderlying spatial
processes.But we may need morethansimpleand directobservation to shakeout thoserele-
vanteducatedguesseswhoseconfirmation or denialextendsourknowledge.It maybejust as
important to transform pieces of cartographicallyportrayedinformation into numbers,
locatedon a varietyofscales,capableofbeingmanipulated mathematically (Fig.8).
VectorSpace

MAP Measurement
System t Numbers I i) Characteristic
Equation

Matrix
Array

FIGURE 8-The flow of informationfromthe map to the characteristicequation.

Though theyvary greatly,the examplesthatfollow all have the proceduralscheme

portrayed abovein common.Information readilyavailableon a mapis extractedin numerical
form.It isthenarrayedas a squarematrix,easilytranslated
into thegeometry of itsappropriate
vectorspace.The eigenvalues,andsometimes theeigenvectors ofthematrix,arethenexamined
to see ifsomeusefulinsights are obtainedby interpretingthesecharacteristic
values,literally
valuesthatare characteristic of theinformation
of thestructure we extractedfromour original
map.

Eigenvaluesand Terrain Roughness

While thenotionof roughnessis difficult to define,we intuitively recognizedifferent
degreesofterrain convolution, fromthe'flat'plains,to the'rolling'hillsto the'jagged'peaks.
Suppose we abstractfromtopographicalsheetsinformation about the relativeheightsof
pointsin a triangular samplespace. Quite literally, our lower triangular matrixis simplya
recording of the contour levels underlying the locations of the matrix elements(Fig. 9).
Thus,if T is a triangular'terrainmatrix,'we can define A = T+ Tt+ a
I, squaresymmetrical
matrixwithits characteristic equation,eigenvaluesand eigenvectors. We entera trace,or
principaldiagonal of ones so that we have a common standard for comparative purposes.
Extracting information in thistriangular way, ratherthanby overlayinga square matrix
is simplya mathematical
directly, convenience, bothon thegroundsthatwe shallwantreal

This content downloaded from 202.28.191.34 on Fri, 18 Dec 2015 09:47:29 UTC
All use subject to JSTOR Terms and Conditions
GEOGRAPHICAL INTERPRETATION OF EIGENVALUES 63
eigenvalues (thoughimaginary onesshouldnot be dismissed too quickly),and because
computer routinespresently available can handle symmetrical matrices moreeasily.The
matrix records
actually thecontour values asdecimal numbers relativetothe highestpointafter
reductiontosomebase,perhaps sea-level.
Let us considerthepossibilities.If a triangular matrix wereplaceduponan absolutely
planesurface,theneach elementwould be the same. Upon reduction to base-level,
we could
thinkoftheresulting nx n matrix as a seriesofn vectors ofsimilar unitlength all lyingon
topofoneanother. Theusualenclosing oblatespheroid inn-space wouldcollapseto a single
line,andwe wouldhaveoneprincipal eigenvalue of n,and (n- i) remaining eigenvalues of
zero.Supposeourplanesurface isdotted withdrumlins, orsome
equallyconvenient morphological features. Our matrix would /
reflect
theincreasingdeparturefrom the plane('roughness'?), and
decreaseourfirsteigenvalue tosomevalueslightly smaller thanx
n. At the otherextreme, we mightexpectthecharacteristic /,
equationofa verymountainous terrain matrix tohavea relatively xxxxxxx
smallprincipalroot.Ifwe hada split-image moving camera, we xxxxxx /
couldshowthatas a planesurface becamemoreandmorecon-
voluted,itsvector
representationwouldbecomemoreandmore
likean n-dimensional porcupine whose
liker1 ~spines
anr~
.~
1.matrix'
were
. changing
in FIGUREand
9-A triangular 'terrain
its overlay-
transpose
in angulardirection.
lengthanddispersing inga contourmap.
areasin theeastern
Threecontrasting UnitedStateswere
sampledforillustrative
purposes:(i) theNewJersey coastalplain;(2) an areaon therolling
plateau;and(3) an adjacent
Allegheny areato thesouth-east
in theRidgeandValleysection
ofcentralPennsylvania.
TABLEI
of Terrain
Eigenvalues Matrices

40 Al
Location lt t2i
I

New Jersey
CoastalPlain 0.6628 8.34

Plateau
Allegheny 0.2451 3.84

Ridgeand Valley 0.2096 3.I2

Examinationof the firsteigenvalueof the threecharacteristic equations(Table I) seemsto

give some support forthe intuitive notionsabove. The flatNew Jerseyplainhas by farthe
as a
highesteigenvalue(expressed proportion of thesum of thetraceelements)and thereis the
plausibleprogressionthrough the Plateauareato the Ridge and Valleysection.
Five thingsshouldbe notedas possibleleads forfurther exploration.First,it mightbe
worthwhile the
rotating matrix in thex, y plane to see if the questionof terrainorientation
could be illuminated.Withtheplateauand coastalplainexamples,we mightnotexpectvery
differenteigenvaluesto result.This mightnot be trueof theRidgeand Valleyexamplewith
linear,and stronglyorientated,
its distinctive pattern.Secondly,some investigation of the
generaltrendsof thesurfacemightbe incorporated. Is a tilted
planesmooth?Extracting major

This content downloaded from 202.28.191.34 on Fri, 18 Dec 2015 09:47:29 UTC
All use subject to JSTOR Terms and Conditions
64 P. R. GOULD
trends(rotationsin thex, z and x, y planes),and examiningtheeigenvaluesof the'residual'
surfacemight be fruitful.
Thirdly,we shouldremember thatourmatrixis alsoa terrain 'filter.'
the
Separate sample elements in
(points) space and we pickup large'waves'; closethe distance
and increasetheresolutionof our matrixlens,and we pay a priceforpickingup thesmall
features(W. R. Tobler, I966a and I966b; F. W. Preston,1966; W. C. Krumbein,1966).
Next,supposewe had takena square,unsymmetrical matrixdirectly fromtheterrain, rather
thanconstructing a symmetrical one fromourtriangular its
sample, transpose and the identity
matrix.The eigenvectors associatedwiththerowswould not be thesameas thoseassociated
withthecolumns.We could,in fact,represent theoriginalterrainmatrixin termsofitseigen-
valuesand row and columneigenvectors (Bickleyand Thompson,1964, 94-IOI). Breaking
down,or decomposing, sucha matrixintoadditivecomponents is termedspectral resolution,
and suchan analysisof terrainmightwell lead to somefruitful that
insights are trulymor-
and
phological geographical ratherthan simplygeological. Finally, recent
investigations into
topographical eigenvector have
analysis indicatedthatwe may have some fruitfulleads into
the questionof the compactionof cartographical data. That is, whilehuge matricesof the
orderof thousandsmake impossibledemandsupon the storageof even giantcomputers,
highlyeffective approximations by the additionof eigenvectors reducethe requirements
by manyordersof magnitude(M. A. Holley,1966).

EigenvaluesandConnectivityMatrices

Anynetworkmaybe considered as a numberofpointsornodesjoinedbylines.Considered

in thisverysimpleway,we may translate a stream,road, or airlinenetworkon a map,for
example, intoa matrix
connectivity by recordingon a binaryscale (I or o), whetheror not
TABLEII
associated
Eigenvectors withConnectivity
Matrix
UgandaRoadNetwork,I92I

Associated
Eigenvector With:

Node 2i = 3.65 A2 = 2.83 A3= 2.62 ,4 = 2.45

I. Kampala 64 7 o 20
2. Luweru 24 4 o 14
3. Mityana 29 8 32 -3
4. Mubende 13 8 51 - 25
5. Kyenjojo 6 6 5I -33
6. Fort Portal 2 3 32 -23
7. Entebbe 24 4 o I4
8. Jinja 24 5 o 27
9. Busenbatia 11 3 o 19
io. Masaka 35 14 -32 -36
II. Mbirizi I5 II - 32 - 48
12. Mbarabara 6 6 -20 - 33
I3. Bukata I3 8 -20 -25
I4. Kiboga 29 -22 o 13
5. Hoima 14 -47 o -I
I6. Masindi 7 -64 o -15
I7. Butiaba 3 -35 o -Io
I8. MasindiPort 3 -35 o -Io

This content downloaded from 202.28.191.34 on Fri, 18 Dec 2015 09:47:29 UTC
All use subject to JSTOR Terms and Conditions
GEOGRAPHICAL INTERPRETATION OF EIGENVALUES 65
directconnectionsexistbetweenpoints(W. Garrisonand D. Marble,1965).In thefollowing
examples we shalldefineeach place as connectedto itself,so we shallbe dealingwithsym-
metricalconnectivitymatriceswithunityin thediagonalelements.

TheRoadNetwork ofUganda,1921and1935
Considerthebasicoutlineof themainroadnetworkof Ugandain 1921(theconnectivity
matrixis too cumbersome
to reproducehere),translated
fromthemap intoa polynomialof

LAKE
VICTORIA
Roads1921
1935 --------
Extensions
Bukoba
0

FIGUREIo-The schematic ofthemainroadnetsin Ugandain I92I and I935.

representation

order 18 (Fig. Io). At least the firstfourrootsof the characteristic equationare positive,
and associatedwitheach of theseeigenvaluesis an eigenvector interest
thatis of considerable
fromthegeographer's pointof view (Table II). The firsteigenvectoris a linefromtheorigin

This content downloaded from 202.28.191.34 on Fri, 18 Dec 2015 09:47:29 UTC
All use subject to JSTOR Terms and Conditions
66 P. R. GOULD
locatedin an eighteen-dimensional space.Vectorsrepresenting
well-connected townswillnot
only lie in the middle of a large number of dimensions but will tend, turn,to lie
in
close to the principalaxis of our envelopingoblate spheroid.Towns thatare moderately
well connectedwillnotlie in themiddleofso manydimensions as thewell-connectedtowns,
and will tendto formsmallstructural clusterson theirown. These,of course,maylie close
to othereigenvectors, whichare orthogonalto therest(we have a symmetrical connectivity
matrix), and form theother axesofourhyper-figure.

TABLEIII
associated
Eigenvectors withConnectivity
Matrix
UgandaRoadNetwork,1935

Associated
Eigenvector With:

Node 2i = 4.64 ,2
= 4.08 23 = 4.00 24 = 3.43

I. Kampala 44 I 7 28
2. Luweru 23 9 3 27
3. Mityana 34 13 I 27
4. Mubende 37 12 -II -12
5- Kyenjojo i6 -4 -13 -9
6. Fort Portal 5 -I -5 -II
7. Entebbe 21 8 3 23
8. Jinja i8 -Io 30 5
9. Busenbatia II -20 38 -7
Io. Masaka 35 5 -5 -22
I. Mbirizi 23 II -7 -33
12. Mbarabara 12 8 -5 -46
I3. Bukata I0 5 -2 -9
I4. Kiboga 27 o -9 7
I5. Hoima 17 -22 -22 2
16. Masindi I -43 -29 3
17. Butiaba 8 -2I -17 2
18. Masindi Port 3 - 4 -0o I
19. Nakasongolo 6 3 I II
20. Ntungamo 4 3 -3 - 3
21. Busia 12 -23 45 -9
22. Mbale 9 -26 39 -12
23. Bukoba 13 8 -3 -28
24. Ibanda 3 3 -I -19
25. Katunguru 3 I -3 -I7
26. Kabale I I -I -13
27. Gulu 6 - 35 -20 2
28. Lira 7 -40 -i8 I
29. Tororo 6 -16 28 -9
30. Soroti 4 -22 7 -5
3I. Kitgum 4 -25 -I3 I

It is interesting
to seethewayin whicheachtownis valuedon thesuccessive eigenvectors,
remembering thatwe are examiningonlythepurestructural of our network.
characteristics
The firsteigenvector is clearlytheKampalanode,and withthevaluesdecreasing as we move
awayfromKampala,we mightinterpret aroundthecapital.Thisis nothing
it as a fieldeffect
new in thehistoryof Uganda,forKing Mutesasetthesame structural patterns long before
withhis systemof couriersto the paramountchiefs.Thus the British,by 1921, had made

This content downloaded from 202.28.191.34 on Fri, 18 Dec 2015 09:47:29 UTC
All use subject to JSTOR Terms and Conditions
GEOGRAPHICAL INTERPRETATION OF EIGENVALUES 67
onlysuperimpositions andextensions on anearlier system oflinkages, andwe might think of
themainroad system as theadministrative hierarchy literallyetched into thelandscape.
Thereis someintellectual comfort, too,thatthevaluesofeachtownareintuitively plausible.
NoticethewayKampalaitself, by farthebestconnected townowingto thenumber ofits
direct linkages and itscentral location,has the value
highest (64). Entebbe is alsohigh(24),but
because ofitsterminal nature itisbelowMasaka(35),Mityana (29)andJinja (28)which aremore
nodalin character. Thisis repeated at a lowerorder,too,foralthough Mbirizi(15) and
Bukata(I3) arebothone connection awayfromMasaka(35),Bukata'sterminal natureon
LakeVictoria lowersitsvaluebelowMbiriziontheroadtotheMbarabara terminalpoint.
Thesecondeigenvalue andeigenvector 'pullout'a smallregional network ontheMasindi
node,whilethethird contrasts thelinearpatterns oftheMityana to FortPortalandBukatato
Mbarabara routes. Finally, the fourth vector could be
perhaps interpreted as 'inner-connected-
ness'contrasted with'peripheralness'. Ina sensethisnotion repeats, andis,therefore, related
to,
thefirst component which was interpretedas a field effect.
Thus, we may have herethe non-
linear pieceofthefieldeffect, plausibleperhaps in viewofwhatwe knowaboutthenon-
lineardecline ofa city'sfieldwithdistance. Itis,perhaps, worthnoting thatifwe weight each
element ofoureigenvectors bythesquarerootoftheeigenvalue (vi /), so thattheelements
oftheeigenvectors associated withthesmaller eigenvalues aresuccessively reducedin scale,
we areinthemiddleofa conventional principal components analysis.
Ifwe nowexamine thesameareain 1935,we canseehowthestructural characteristics
of
1921havebeenstrengthened bytheextensions ofthenetwork intheintervening years(Table
III). Thefirst eigenvector canstillbe interpreted as theKampalanodeand,whilethesecond
is stillmainly theMasindisub-region, thenewlylinkedtownsofLiraandGuluareclearly of
structural importance. In lateryearstheybecomeincreasingly competitive andimportant in
a space-organizing senseas nodesofan areaopenedup by thecottondevelopments ofthe
1940sand 1950s.The eastern townsalsohavesomewhat higher valueson thiseigenvector,
andifthese areincluded, thesecond root
largest and its associatedeigenvector seemtobebreak-
ing out a northern sub-network that contrasts strongly with the south and south-west. Somecon-
firmation comesfrom thethirdeigenvector, whichclearly breaks thelarger northernnetwork
intotwodistinct andcontrasting north-west andeastern sub-regions. Only the fourtheigen-
vectorfinally pullsouta southern network focused upontheMbarabara node,andin a sense
thisisa space-organizing development springing fromthepurely linearpattern ofconnections
I4 yearsbefore.

TheRoadNetwork ofSyria
The transport network of Syriarepresents
an interesting
exampleofthewayin which
or connectivity
structural characteristics
may be disturbed
bypolitical changes (N. Kanaan,
I965). Syriawas out
carved of theTurkishEmpire after
theFirstWorld War andthe oldroutes
to themajorportsofBeirutandAlexandretta weresevered by new boundaries,
political so
thattheminorportofAl Latiquiya hadto beartheexportburden. thespatial
Significantly,
reorganization thatsuchdislocatingchangesinducedis stillunreflected in the structural
characteristics
oftheroadnetwork nearlyhalfa century
later(Fig.I I).
The firsteigenvectorbreaksout two strongnodalregions focused uponAleppoand
Hama,whilethesecondandthird pulloutthenodesofDamascus andHoms.Thusthemain
fociarestilltheoldtownsatthebaseofthemountains thatparallelthecoast;townsthathave
beenofimmense importance asbreak-of-bulk
historically pointsordesert portsforthecamels

This content downloaded from 202.28.191.34 on Fri, 18 Dec 2015 09:47:29 UTC
All use subject to JSTOR Terms and Conditions
68 P. R. GOULD

TABLIE IV
associated
Eigenvectors withConnectivity
Matrix
RoadNetwork,
Sy-ria 1963

Associated
Eigenvector With:

Node A,1= s.I' ). = 4.26 )L3 = 3.90 )L4 =3.67

i. Ain Deiwar 0.00 -0.00 0.00 0.02

2. Karatchok 0.00 - 0.00 0.00 0.01
3. Tell Kotchek 0.00 -0.00 0.00 0.02
4. Dmir Kabar 0.00 - 0.00 0.01 0.05
S. Al Kamishli 0.00 -0.00 0.02 o.o8
6. Al Aasakeh 0.02 -0.02 0.05 o.i6
7. Ras-el-Ain 0.00 -0.00 0.01 o.o6
8. Abou Kamal 0.01 -0O.01 0.04 0.11
9. Deir-ez-Zor 0.07 -0.05 0.11 0.30
i0. Rakkah 0.19 -0.00 0.00 0.27
i i. Tell Abyad 0.o6 0.00 - 0.02 0.I4
12. Ain-el-Arab o.o6 0.01 - 0.09 0.10
I3. Manbij 0.07 0.02 -0.12 o.o8
14. jarablus o.i8 0.05 -0.23 0.14
I5. Al-Bab 0.I2 0.03 -0.I2 0.07
i6. Aleppo 0.45 o.o8 -0.25 0.11
17. Izaz o.i8 0.05 - 0.21 o.o8
i8. Ifrin 0.11 0.04 -0.14 -0.03
i9. Rajon 0.02 0.01 -0.04 -0O.01
20. Harem 0.13 0.03 -0.05 -o.i8
21. Iblib 0.24 o.o6 -0.15 -o.i6
22. Jisr-esh-Shagour 0.31I 0.04 0.00 - 0.32
23. El Ma'arra 0.19 0.04 -0.10 0.00
24. Maarrat-en-Nouman 0.33 0.04 - o.o6 -0.08
25. Al Lataquiya 0.08 0.01 0.02 -0.17
26. Al-Haffeh 0.09 0.00 0.08 - 0.23
27. Jebleh 0.02 o.oo 0.05 - 0.I4
28. Banias 0.02 -0.00 0.15 - 0.2I
29. Kadmous 0.07 -0O.01 0.24 - 0.28
30. Misyaf 0.13 - 0.03 0.24 -0.13
31i. Hama 0.40 -0O.01 0.17 - 0.02
32. Salamiya 0.19 -0.05 0.19 o.i6
33. Horns 0.20 -0.15 0.37 0.19
34. Safita 0.04 - 0.02 0.23 -0.I7
35. Tartous 0.01 -0O.01 0.13 -0.14
36. Tell Kalakh 0.09 - o.o6 0.29 -0.04
37. Palmayra 0.08 -0.13 0.23 0.25
38. Douma 0.08 -0.22 0.18 0.17
39. Damascus 0.03 -0.45 -0.05 0.02
40. Zabadani 0.00 -0O.14 -0O.01 0.00
41. Sab-Biar 0.00 -0.14 -0O.01 0.00
42. Shahba 0.00 -0.23 -0.07 - 0.02
43. Suweida 0.00 - 0.32 -0.15 -0.10
44. Salkhad 0.00 -o.i6 - 0.09 -0.07
45. Bosra 0.00 -0.22 -0.12 - 0.09
46. Dar'a 0.00 - 0.25 -0.11 - 0.07
47. Feek 0.00 -0.17 -0.07 -0.04
48. Onietra 0.01 -0.32 - 0.09 - 0.03
49. SheikhMiskin 0.01 -0.41 -0.14 -0.07

This content downloaded from 202.28.191.34 on Fri, 18 Dec 2015 09:47:29 UTC
All use subject to JSTOR Terms and Conditions
GEOGRAPHICAL INTERPRETATION OF EIGENVALUES 69
thatfollowedtheold caravantrailsto theeast.Here theywere unloadedand theshipments
brokendownintosmalllotsfordonkeytraveloverthesteepmountaintrailsto theseaportson
the Mediterranean.Only thefourtheigenvector disclosesan inner,desertsub-network, and
Al
significantly Latiquiyaneverappears as a focusof evenminor nets.
The spaceorganization,
as reflected
by thepurestructural
characteristicsoftheroadnetwork,stillseemsfocusedin the
age-oldway despiteefforts
overthepast46 yearsto reorganize theareaand pointit to a new
export outlet.

FIGUREIi-The simplified
roadnetof Syria.

MeasuresofOrientation
and TensorFields
Dealing with spatialpattern,shape and structure, we frequently findourselvesfacing
questions oforientation (A. E. Scheidegger,I965). Oftena largenumberofdirectional measures
is available,and fromthesewe maywishto calculatethe'best'or meandirection. In geomor-
phology,measuresof the orientation of beddingplanes,sand grains,streampebblesand so
on maybe of interest; in humangeographytheexamplesmayrangefromculturalstudiesof
dwellingorientation in China by geomancy,currently beinginvestigated by thegeographer
B. Boxer,to orientational components of transportnetworks.
FollowingA. E. Scheidegger's notation,we assumeeach axis ni is a unitvectorwith
componentsnix,niy and ni_ in an orthogonalsystemof co-ordinates. In a two-dimensional
spaceni will onlyhave the components ni andniy(Fig. 12),andifwe measuretheorientation
alwaysas an azimuth(northor south'so-many-degrees' eastor west),then:
ni = Cosa
ny
= sin a

This content downloaded from 202.28.191.34 on Fri, 18 Dec 2015 09:47:29 UTC
All use subject to JSTOR Terms and Conditions
70 P. R. GOULD
Note threethings:(I) our orientationvectorsare assumedto be of unitlength(we do not
careifourstreampebbleis 3 or 4 incheslongas we areinterested onlyin direction);
(2) though
it is convenientto measureour azimuthfromthenorth-south axis,it would notmatterfrom
whichdirection we actuallytook our bearingspro-
East viding,of course,we were consistent;and (3) we
I mustremember to watchthesignsof our sinesand
cosines,thatis,if the bearingsare in thenorth-east
,---- n-- and south-east quadrants,the signsof our sineswill
'ni /..... be positive negativeelsewhere;and if the bear-
and

i~~/ ,
ingsare in the south-eastor south-westquadrants,
~thesignsof thecosineswill be negativeand positive
elsewhere.So foreveryazimuthalbearingwe shall
have two piecesof information, an ni, thatis, plus
X-South or minuscos cxdependingupon the quadrant,and
nix
an niy,thatis,plusor minussincx.
FIGURE
FIGUREI2-The components
I2-The components in two-spaceof
in two-space of 1 * * 1
an orientation
vector.
For the threedimensionalcase we have to take
intoaccounttheplunge,
whoseanglewe shallrepre-
sentbyp. Then:
ni = cos a Cosp
niy= sin c. cos p
nz = -sin p

Assumingthatthedeviationsof theindividualaxesfromthemeanarenormallydistributed,
themeanvectorX is theeigenvectorassociatedwiththelargesteigenvalueofthesymmetrical
matrixA whose elementsare thesumsof squaresand cross-products of thenix,niyand niz
Let
components. us consider
an example.
TheMean Orientation
ofStream Pebbles
Supposingwe had the measuresof thelong axesof streampebbles
followingorientation
(Table V).
Then:
a1 = Z ni = 2.5950
a22 = 'n22 = 2.7370
a33 = ? ni2 = 0.5353
a12 = a21 = = 0.7359
n nniy
a13 = a31 = nix iz = -0.4000
a23 = a32 = niy niz = -0.6789
So:
2.5950
A = 0.7359 2.7370
-0.4000 -0.6789 0.5353
-a symmetrical
matrixwitheigenvalues:
i = 3.6005 / o.6342\ / 0.7678\ /o.09o0\
2 = 1.9409 and eigenvectors o 0.7333 o.6346 0.244I
23 = 0.3260 -0.2452 -o.o880/ \o.9655

This content downloaded from 202.28.191.34 on Fri, 18 Dec 2015 09:47:29 UTC
All use subject to JSTOR Terms and Conditions
GEOGRAPHICAL INTERPRETATION OF EIGENVALUES 7I
or of unitlength:
arenormalized,
Sincetheseeigenvectors

3
norm 11eA - = e2
i=I

the firstelement of the eigenvectorassociated with A2 is ni, = cos a. cos p. Since -sin p =
-0.2452, cos p = o.9694. Thus:
0.6342
cos o = = 0.634
0.9694
And since both cosine (o.6342) and sine (0.7333) are positive, the mean orientationvector
quadrant,or N. 50?42' E. The lastone is orthonogonal
mustlie in thenorth-east to both.The
plunge component is -0.2452.

TABLEV
ExampleoftheMean Orientation
Hypothetical ofStream
Pebbles

nix= nly = nz
Pebble Azimuth
(a) Plunge(p) cos a sina cos p -sin p cos a. cosp sina . cosp - sinp

I N. 3o E. 24 o.8660 o.5ooo 0.9135 -0.4067 0.7911 0.4568 -0.4067

2 N. 20 W. 8 0.9397 -0.3420 0.9903 -0.I392 0.9306 -0.3387 -0.I392
3 S. 40 W. i6 -0.7660 -0.6428 0.9613 -0.2756 -0.7364 -o.6i8o -0.2756
4 N. 40E. I8 0.7660 -0.6428 o.95II -0.3090 0.7285 0.6114 -0.3090
5 S. 85 E. 13 -0.0872 0.9962 0.9744 -0.2250 -0.0850 0.9707 -0.2250
6 N. 65 E. 21 0.4226 0.9063 0.9336 -0.3584 -0.15I5 0.846I 0.3584

The Orientation Networks

ofTransport
Littleattentionseemsto have been givento thequestionof theorientation of transport
networks,althoughit mightbe worthwhile investigating such generaland thoroughly
geographical particularly
properties, on a comparative basis.For example,a well developed
country,whose contemporary transportnetworkreflects verylong historicalprocessesand
large movementassociatedwitha highlevel of internalexchange,mighthave onlya very
weak overallorientation. On theotherhand,a countrywhosepresenttransport networkis a
of a colonialregimeor an extractive
reflection 'draining'economymighthave a stronger
focus.Takingfromthemap a sampleoftheorientation
overalldirectional ofmajorroadlinks
andreducing to
these the2 X 2 component matrixcould provide some insightintothisquestion
by comparingthe ratioof the eigenvalues.In some countries, the differencebetweenthe
eigenvaluesmightbe quitesmall.The figureenclosingthevectorsnixand niywouldbe almost
circularwithtwo axes of approximately thesamelength.However,countriessuchas Tan-
zania and SierraLeone,withhistories of colonialismand economiesgearedto theexportof
bulkraw material,mightwell have a networkwitha ratherstrongoverallorientation. This
would be indicatedby quite disparatepairsof eigenvaluesassociatedwiththeirorientation
matrices.
For illustrative
purposes,the main road networksof SierraLeone and Tanzania were
sampled with twenty-seven and nineteennon-overlapping circularquadrats(Fig. I3). Within

This content downloaded from 202.28.191.34 on Fri, 18 Dec 2015 09:47:29 UTC
All use subject to JSTOR Terms and Conditions
72 P. R. GOULD
each quadrat,road segments weremarkedoffbetweenmajortowns,terminals on majorroad
branches andboundary points wherethe major roads crossedthecircular
quadrats.The orienta-
tion of the road linkswas measuredon the straight linesjoiningnodes so definedand the
nixand niycomponents obtained.
Even when measuredat a verygrossscale,
"nix plausible measuresemergeto whichsomespecu-
^s^lative
\^f"---*-*-i\ interpretationmaybe given.If theratioof
the two eigenvaluesis takenas our measureof
,/ , \ ttheorientational strengthof the road network
n,
/n----y; \ within a quadrat,high values generallyappearin
f l9y8 ^^^ ^"' those areas where the roads serve mainlyas
^> I~\~'draining,' or feederlinks(Fig. I4). Thus, along
'^^/ i j the railwayand in areassuch as the Kilombero
",s ^~
. ( ' \ ~ I valley,manyofthequadratshavea strongnorth-
southorientation, while the high values in the
\ '* south reflecta strongeast-westfocus. On the
other hand, 'islandsof development',like the
Bukoba (I.II), Mwanza (1.38), Tanga (I.29),
Lindi(I.3o) and Mbeya (1.05-I.36) regions,have
FIGURE 13-A samplequadratused in calculating low eigenvalue ratios reflectingtheir weak
theorientation roadnet.
of theenscribed 'directionality'.
In SierraLeone, too (Fig. IS), rathersimilar
interpretations be
may given. Most of the quadratsalong themainrailwayline have high
values,reflectingthe feederfunction of the roadswithstrongorientations. The majorexcep-
tion is the Bo quadrat(1.41), whichis certainly a much more developed'island'thanthe
surrounding areas.Generally,the less developed,peripheralquadratshave a ratherstrong
orientation,while themoredevelopedcentralareashave low values.
The evidenceis weak,butenoughperhapsto explorethequestionofnetworkorientation
further. Certainlytheratiosof theeigenvaluesseemto reflect thedegreeto whichareashave
enteredintotheinternalexchangeeconomyof thecountryand,ifwe acceptthenotionthat
theroadsarethefinaloutcomeofnumerousprobabilistic decisionsetchedintothelandscape,
we mayhavea worthwhile measureofthefunction servedbyan areaandthedegreeto which
ithasabandonedthetraditional economyto enterthemodernworld.
Threeotherpointsseemworthmentioning as leadsforfurther exploration along these
lines.Clearly,thescaleofobservation is goingto be criticaland,as we decomposea transport
networkintoitsdirectional components, it would be well worthwhilechangingthesize of
our quadratfilterto see whatwe pickedup at various'wavelengths' (L. Curry,1966; W. R.
Tobler,I966c; W. R. James,I966; J.L. Holloway,I958). Secondly,someweighting system
is needed,themostobviousand easiestbeingthemultiplication ofeachroadsegment (orienta-
tionvector)byitslength,and perhapsincorporating flowdatatoo. Finally,and thismightbe
quiteintriguing, we could considerthematrices of sumsof squaresand cross-products when
the elementshave been dividedby N as orientational tensors(Scheidegger,I965), sinceit
can be shownthattheyconformto thedefinition of a secondranktensor:

far= ;ifji
x. x.
_-
Xi Xj

This content downloaded from 202.28.191.34 on Fri, 18 Dec 2015 09:47:29 UTC
All use subject to JSTOR Terms and Conditions
GEOGRAPHICAL INTERPRETATION OF EIGENVALUES 73
Or, usingtheEinsteinsummation
convention:
fap = fij * ija
By movingthesamplingquadratby smallincrements overthemap surfaceand plottingthe
to
tangent
projections themain of
directions thetensorat everypoint,we could thinkof the

14-Tanzania: theratiooftheeigenvalues
FIGURE in sampleorientation
quadrats.

resultas a tensorfieldindicatingthe spatialvariationin an area's orientation

or directional
'pull'. Again,by opening and closingour quadratlens, we could examine the components
of sucha fieldby a sortofintuitive'Fourieranalysis'.

TheEigenvalues ofMeanInformationFields
modelsis themeaninformation
A criticalcomponentofnearlyall MonteCarlo diffusion
F

This content downloaded from 202.28.191.34 on Fri, 18 Dec 2015 09:47:29 UTC
All use subject to JSTOR Terms and Conditions
74 P. R. GOULD
field(M.I.F.),thebasicspatialexpression
oftheprobability
thatan innovatorwillcommunicate
a new idea overparticulardistances I957
(T. Hagerstrand, and I965; R. Yuill, I964). Usually,
though not always (L. Bowden,
1965; R. Colenutt, 1966), the con-
tinuousinformation fieldis handled
/-2GUINEE > as a square symmetrical arrayof
00
co 2
(155 )y V Jfigures totalingunity. Considered
as a matrix,the determinant of a
symmetrical M.I.F. is always zero,
00
JX.
\362 ^ A S / becausetherows and columnsare
g1421v ^ ^^ -^not linearlyindependent, and the
largest,non-vanishing minorwill
134 )/ have the order + (n- )/2. Thus,
Freetown 128
124 k1J\^ \172
associated with the characteristic
0< \-
>
s --z/^T;
/B equationof such a M.I.F. will be
a" K Y ~\
141 Y 179 A I+(n-I)/2 real eigenvalues, with
^222 )\ 196 \I J theremainderequal to zero. This
' mayalsobe trueof meaninforma-
205 1 tion fieldswhere there is a direc-
m(^ 212
21<2 Ai
\?^^^^^^ LIBERIA
tional bias that is symmetrical
ATLANTIC OCEAN in form. Unsymmetrical,and
thereforebiased, mean informa-
' tion fieldsmay not hold to this
FIGURE I5-Sierra Leone: the ratio of the eigenvaluesin sample rule.
orientation
quadrats. Let us consider the following
3 x 3 fieldusedbyYuill:
/0.05 0.09 0.05\
Fto = 0.09 0.44 0.09
\o.o 0.09 o.05/
whoseeigenvalues
are:
21 = 0.48237
22 = 0.05763
23 = 0.00000

0.54000

summingto 0.54,thetotaloftheelements on theprincipaldiagonal,or thetraceofthematrix.

If we visualizethe geometrical representation of the M.I.F. in three-dimensionalspace,we
have one largevectorand two muchsmalleronesthatcoincidewithone another.For com-
parativepurposes,let us considerthe 'reach' or 'power' of the M.I.F. as the value of the
principaleigenvalueexpressedas a proportionof thecentralelementin thefield,say a. As
communication in theregionbecomeseasier,and thefriction of distancedeclines,theperi-
pheral elements of the M.I.F. will gain at the expense of the innervalues,and our major
vectorwill shrink,while the other,linearlydependentones will both grow in lengthand
changedirection, comingcloserto themajorvectorcontaining thecentralvalueat theorigin
of thefield.Forexample,as thefriction of distancedeclines,thefieldmightchangeto:

This content downloaded from 202.28.191.34 on Fri, 18 Dec 2015 09:47:29 UTC
All use subject to JSTOR Terms and Conditions
GEOGRAPHICAL INTERPRETATION OF EIGENVALUES 75
/o.o6 o. o.o6\
yFtl = o.II 0.32 O.II
o.o6 0.II o.o6/
witheigenvalues:
2 = 0.40493
22 = 0.03 507
23 = 0.00000

0.44000

Thus, the a ratio(Table VI), has increasedto 1.265 fromI.095. If the frictionof distance
declinesstillfurther
then:
/o.07 0.13 0.07\
yFt = ( .13 0.20 O.I3J
\0.07 o.13 0.07/
witheigenvalues:
1 = 0.35628
22 = 0.01628
23 = 0.00000

0.34000
so that:
a = 1.78

This is clearly approaching the limitingcase (yFi,,), when distance has no effectover the
fieldspace,and eachelementis identicaland approximately equal to o.1II. In sucha case,all
the vectorsare identicaland caequals 3.00, theorderof theM.I.F. Thus theeigenvaluesare
themselvesa measureof thefriction of distance,withpossiblysome utilityforcomparative
purposeson a cross-cultural
basisas well as acrosstime.For example,it is quicklyapparent
thattheYuillM.I.F. is rather
similarto Hagerstrand's 5 x 5 field,for:
/ .o00 o.oi4 0.017 0.014 o.oIo\
.o014 0.030 0.055 0.030 .o014
HF = 0.017 0.055 0.443 0.055 0.017
.o014 0.030 0.055 0.030 o.o04
\o.oio 0.014 o.oi7 0.014 O.OIo/
witheigenvalues:
21 = 0.46001
22 = 0.05804
23 = 0.00445
23 = o.oo445
24 = 0.00000
= 0.00000
25
0.52250

so the cxratiois I.037, differing

littlefromYuill's I.095.
The sameinsightsholdforreversefieldeffects,
forexample,wherea processis competitive
in and overspaceratherthancontagious,suchas thespreadof centralplaceson thefrontier,
or the spreadof cottonmarketing co-operativesocietiesin theLake Provinceof Tanzania.
In thelattercase,thefieldsoftheco-operatives,
whichareconsidered as singlefunction
central

This content downloaded from 202.28.191.34 on Fri, 18 Dec 2015 09:47:29 UTC
All use subject to JSTOR Terms and Conditions
76 P. R. GOULD

places,would forceaway othersocieties

to somedistance
where'marketing
serviceviability'is
reached.A simplereversefieldmightbe:
/0.15 0.09 o.5\
RF= 0.09 0.04 0.09
\.I5 0.09 o.I5/
witheigenvalues:
21 = 0.35193
22 = --o.II93
=3 = 0.00000

0.34000
and a ratio:
= 8.8i
Notice theway in whichthereversalof theM.I.F. producesa considerable extensionof the
'reach'.
Finally,let us considerthe asymmetrical case. Suppose it could be established
thata
consistentdirectional biaswas presentin thespreadofinformation overan area.Forexample,
in theRidgeand Valleysectionof Pennyslvania, withstronglineartopographical it
barriers,
might be instructiveto handlethe distinct
directionalbias by warping theinformation field
baseduponknowledge,insteadof,or in additionto,varyingthebarriers whosepermeabilities
werebasedupon guesswork.Such knowledgecould be established, forexample,frompairs
of socialmovementsso avidlyrecordedin thelocal newspaper.As an example,let us warp
Hagerstrand's fieldin a north-east-south-west
direction.We have:
/0.004 0.009 0.013 0.017 .oI2 \
0.009 0.027 o.o60 0.034 o.o17
HFB = o.oI3 o.o60 0.443 o.o60 o.o13
0.017 0.034 o.o60 0.027 0.009
\o.oI2 0.017 o.oI3 0.009 0.004/
witheigenvalues:
2, = 0.46223
2= 0.05468
23 = -o.0I552
=4 = 0.00000
= 0.00000
25

o.50I39
The ratio(I.043) hasscarcelychangedfromthevalueof theunwarped(1.037), confirming the
intuitive
notionthatwarpingdoesnotrepresent an extensionofthefield's'reach'butonlyan
orientationalbias.
Summarizing (Table VI), we can seehow thea ratioreflects
thefrictionofdistancein the
area,and theway in whichit illustrates themarkedextensionof the'reach'of a fieldthathas
beenreversed.

andtheClassification
Eigenvalues Problem
Problemsof classification,
whichappearfrequentlyin youngsciencesgropingfororder
andstructure
in massesofempiricaldata,havereceivedsomenumericalandtheoretical
under-

This content downloaded from 202.28.191.34 on Fri, 18 Dec 2015 09:47:29 UTC
All use subject to JSTOR Terms and Conditions
GEOGRAPHICAL INTERPRETATION OF EIGENVALUES 77
pinningsonlyfairlyrecently. Althoughtheearlyworkin discriminant analysiscan be traced
to archeologicalwork in classifyingskulls(M. M. Barnard,1935),and subsequentwork in
thefieldsof botanyand genetics(R. A. Fisher,1936; M. G. Kendall,I957), mostof thead-
vancessincetheSecondWorld War have come fromthebiologicalsciences.For example,a
typicaldiscussion,
listingfifty-six containsonlysix publishedbefore1948(W. T.
references,
WilliamsandJ. M. Lambert,1961).It is interesting to notethatthecomplexproblemshave
to
only begun yield with the of
development high-speedcomputers,for the sheercom-
putationaltaskof and
handling sorting largeamountsofdatain anything remotely resembling
a sophisticated
way has been much too severeuntilrecentyears(R. Sokal and P. Sneath,1962
and 1963; SokalandJ.Camin,I965a and I965b; D. F. Merriam,1966).
TABLE VI
andAssociated
Eigenvalues ValuesofMeanInformation
Fields

M.I.F. 21 a = -1/Central
Element a/Order
M.I.F. a/RankM.I.F.

yFto 0.482 I.095 0.365 0.548

yFqt 0.405 1.265 0.422 0.633
yFt2 0.356 1.780 0.594 0.891
YFlim ?-333 3.000 1.000 1.500
HF 0.460 I.037 o.2o8 0.346
RF 0.352 8.8io 2.935 4.41I
HFR 0.462 I.043 0.209 0.348

Though problemsof classification raiseexplicitlythe questionof measurement, the

criteriaupon whicha classificatory schemeis based need not, indeedsometimescannot, be
measuredon thepowerfulratioand intervalscales(W. S. Torgerson,1958; C. Churchman
and P. Ratoosh,1959). Sometimeswe mustbe contentwiththeweakerordinaland binary
scales,particularly when our criteriaare essentially attributes, or qualitativemeasures.It is
no longernecessary to belabourthefalsedichotomy betweenthequalitativeand quantitative,
fortheincreasing numberof non-metric studiesin thesocialand behaviouralsciencesmakes
it obviousthatin applied,as well as in puremathematics it is logicratherthannumber thatis
critical(L. Guttman1966; R. McGinnis,1965; C. H. Coombs, 1964; Coombs and R. C.
Kao, 1955;R. L. Carneiroand S. Tobias,1963).Furthermore, thereis frequently a highdegree
of metricinformation containedin simplermeasurement scales, such as theordinal,whichcan
be extracted whenpropermethodsare used(J.B. Tukey and R. P. Abelson,1959; Coombs
and Kao, 1960; R. N. Shepard1962and I963;J. B. Kruskal,I964a and I964b; J. C. Lingoes
1964and 1966).
Classification a divisiveprocess:we assignan objectto two or more'pots'
is essentially
according to itsrelativelocation upon a scale,or scales.Repetitionof thepluralform(scales)
is crucial:as geographers we mayfrequently findourselvesdealingwiththingsthataremulti-
dimensional in character so thattheycannotbe assignedaccordingto theirlocationson just
one rulerbut on many.This is assuming,of course,thatwe wishto use all theinformation
availableto us. Ifthereareonlyfewrelevantrulers,or dimensions, in a particular
classificatory
problem, then we may be ableto divideour into
objects large,internally homogeneous groups;
but a largenumberof dimensions, whichis judged to be relevant,may implyconsiderable
heterogeneity in theobjectsto be classified.
Problemsofuniform regionalizationareessentially
taxonomic(M. G. Kendall,1939; M. J. Hagood, 1943; B. J.L. Berry,1960, 1961; Berryand

This content downloaded from 202.28.191.34 on Fri, 18 Dec 2015 09:47:29 UTC
All use subject to JSTOR Terms and Conditions
78 P. R. GOULD
M. Ray, 1964; Gould etal., I96I; E. Soja, 1966),andregionalschemesbaseduponfunctional
maybe handledin thesameway(Berry,1963,I966 and 1967).
organization

TheHomogeneity ofArchitecturalFeatures
If thearchitectural of churches,
characteristics as well as othercriteriasuchas sites,dates
of buildingand degreesof wealth,reflect theculturalbackgrounds of thebuilders,we might
usesuchrelatively permanent aspectsofthevisuallandscapeas surrogate measures in a cultural
scheme.Considering
classification eachchurchas a row vector,whoseelements areeitherone
or zero dependinguponthepresenceor absenceof a criterion judged to be relevant,our geo-
metricalrepresentationcould rangefromtotalcoincidenceof thevectors(identicalchurches
reflectinga completelyhomogeneousculturalbackground), to greatdispersionin thespace
whosedimensions werethesamenumberas thecriteria used.It is difficult
to workwithsuch
rectangular matrices,however,so we can translate theinformation into a squarematrixof
cosineswhichmeasurethe anglesbetweenunitvectors,each one of whichrepresents one
criterionused in the analysis.A morefamiliardescription would be simplythematrixof
correlation whichis precisely
coefficients, whatthecosinesbetweentheunitvectorsare.
In CentreCounty,Pennsylvania, thirty-eightmeasureson a binaryscalewere recorded
for 192 churches.The characteristicequation of the resulting38 x 38 matrix(too large to
reproducehere), has the six
following largest roots:
2l = 6.I378 24 = 1.9907
22 = 3.4421 25 = 1.6380
23 = 2.3418 26 = 1.5782
Obviously,iftherewerean extremely highdegreeofculturalhomogeneity reflected
through
architecturaland otherchurchfeatures, the eigenvalueassociatedwiththefirsteigenvector
would be verylarge.In thisparticular instance, thefirstvalueis not impressive, and thereis
onlya verysmalldeclinein thevaluesofeachsuccessive one. The elementsofthenormalized
eigenvectors can be plausiblyinterpreted, however,and thefirstscaleor dimensiondescribes
thedominant'arch-type' of thearea,an open-gabledchurchwithbelfry, builtbetweenI850-
I900, with colouredglasswindows,a parsonageattached,a town locationand relatively
wealthymembers.A seconddimension,quiteindependent of the first,contraststhepoor
woodenrectangular churchwithreturns, builtbetweenI8oo-50, withtheT- or L-formbrick
edificewithGothicwindows.
Thereisno doubt,though,thatthereis littlehomogeneity in thisportionofPennsylvania
today, and the principaleigenvalueexpressed percentage thesumof thetraceelements
as a of
reflectsthis.Furtherextensionsseem promising.It would be interesting to samplesimilar
featuresalong hypothesized 'corridors'of advancefromthe culturalhearthson the eastern
seaboard.Plottingeigenvalueratioswoulddisclosethepresent pattern ofcultural homogeneity,
probablydecliningaway from theold hearths, but withmuch 'noise',so thattrendswouldbe
blurred.Restricting theanalysesto churchesbuiltin theearlierperiodsmightilluminatethe
wayinwhicharchitectural features deterioratefrom'pure'formsthrough timeandoverspace.

PrincipalComponents
It shouldbe clearby now thatsomeof theexamplesconsidered
so far are closelyallied
to principal components analysis(L. L. Thurstone, I947; G. H. Thompson, I95I and I956;
H. H. Harman,I960; R. B. Cattelland W. Sullivan,I960; Cattell,I965a and i965b; P.

This content downloaded from 202.28.191.34 on Fri, 18 Dec 2015 09:47:29 UTC
All use subject to JSTOR Terms and Conditions
GEOGRAPHICAL INTERPRETATION OF EIGENVALUES 79
Horst,1966),a powerfulmultivariate tool thatis closelyrelated,in turn,to canonicalforms
(H. Hotelling, 1936; W. W. Cooley and P. R. Lohnes,1962; T. W. Anderson,1958)and the
moregeneralfieldsof factoranalysis,molarcorrelation analysisand simplextheory(M. B.
Jones, 1958 and 1959). Indeed,ifwe examine the characteristic
equationofa correlationmatrix,
thenthecomponent, or factorloadingsarenothingmorethantheelements ofthenormalized
eigenvectors whichhavebeen'weighted'bythesquarerootoftheireigenvalues. Forexample,
if the 3 x 3 matrixon page 58 is consideredas a correlationmatrixbetweenvariablesV1,
V2 and V3,then:
Vt V2 V3
I.o000 0.750 0.830\
R = V2 0.750 I.000 0.4Io
\.830 0.4o1 I.000/
witheigenvalues:
21 = 2.3416 22 = 0.5929 3 = 0.o655
and corresponding,
normalized eigenvectors: 0.6403 0.0503 -0.7665
0.5283 -0.7332 o.3919
0.5576 0.6559 0.5088
so thatthefactor
loadings are: 0.9798 0.0387 -0.1962
0.8085 -0.7599 0.1003
0.8532 0.5050 0.1302
If we consider the geometry of the problem (Fig. 16), it is obvious that the loadings are, in
fact, the orthogonal projections of the unit vectors (indicatingeach variable or measure) on

Component
1.98'

.85
.81'
....81 . ;..
. 7

3
Component
:Vi
i

V2

-.19

2 '
Component 58 *3 5

FIGURE ofprincipal
I6-The geometry analysis:unitvectors'loading'upontheeigenvectors
components
in three-dimensional
space.

This content downloaded from 202.28.191.34 on Fri, 18 Dec 2015 09:47:29 UTC
All use subject to JSTOR Terms and Conditions
 80 P. R. GOULD
the normalizedeigenvectors. Thus, all the variablesload high on the firstcomponentor
dimensionof theproblem,whilevariablesV2 and V3 arecontrasted on thesecond,to which
V, is virtuallyunrelated.Moreover,if components(eigenvectors) I and 2 are considered
important dimensions, eachitemusedto calculatethecorrelation coefficients maybe 'scored'
on eachcomponent or 'ruler.'A mapofscoresmayfrequently be illuminating whenintuitively
valid notionslike 'economic development,''economichealth,''politicalmodernization,'
'social health'and 'space preferences' are being investigated. Two thingsshouldbe noted,
however:first,we are imposingan orthogonalstructure; secondly,we are examiningand
interpreting only the major axes to our vectorclusters.Imposingan orthogonalstructure
implies thatwhen we interpret our 'dimensions'theseshouldbe independent conceptually
as well as mathematically. We could,of course,make thecomponentsoblique,so thatthe
basic dimensionsof the thingwe were investigating wereinterrelated; but we mightthen
examinetheinter-correlation matrixof thecomponents and look forhigher orderdimensions.
Anotheralternative would be to rotateour orthogonalaxes to some otherposition,perhaps
to producethemaximumdifference betweentheloadingson our dimensions, so thata few
variablescontributed heavily to eachdimension and thenvery littleto all theothers.

ThePerception ofResidential
Desirability
As thelinesbetweenthesocialand behaviouralsciencesblur,and theonce-tenuous links
thickenandbindmorestrongly, itis apparentthatmanyoftheanswers toquestionsnow being
posed by humangeographers will lie in areassharedby a numberof the 'humansciences'
(Gould, I965a). A subjectof considerable importance and interest,thoughbarelyinvestigated
in a systematicfashion,is thatof thepreferential imagesof geographical space.Investigations
in thebroadarea of spatialbehaviourhave been undertaken in architecture,
sociology,psy-
chology, cityplanning,anthropology, economicsand geography,but few have triedto
describetheimagesof geographical spacethatmenhold,and littleattempthas beenmadeto
measurethe varyingintensities of theseperceptualimages-to surveyliterallythe mental
mapsin themindsofmen(C. C. Trowbridge,I939; Gould,I966aand i966b; R. White,I967;
K. Bassett,I967). Yet it is precisely
thedifferential spaceimagesheldby menthatmayform
-crucialpiecesof our ultimateexplanations of thosepatterns,structuresand processesinvesti-
gatedby thehumangeographer. Forexample,it maybe thatthebroadpatterns ofmigration
in thiscountrycanbe trulyunderstood onlyin termsofthementalimages that people have
of thespacecalledtheUnitedStates(Gould, I967). Certainlyin theunderdeveloped world,
wherea tinyeducatedeliteholdsenormouspower of decisionforgeographical assignments,
an understanding of theoverallspatialview thatis heldby themajoritymay be criticalfor
planningpurposesandunderstanding future developments. Furthermore, ifroaddevelopment
in an Africancountryis considered as a searchprocess,thendecisionsto expandand contract
trunkand feederinvestments may also be shapedand mouldedby thementalmaps of the
decision-makers (Gould, i965b and I966c). Moreover,as industries becomemoreand more
foot-looseand 'non-economic' factors weigh more and more heavilyin thedecisionsto locate
branchplants,thenon-optimizing theories of locationwill be forcedto considerthemental
mapsof consensualspaceimages(A. Wilson,I966).
EuropeLooksatEurope
In a pilotinvestigation, students
university in Sweden,England,France,Germanyand
Italywere askedto providerankorder ofthecountries
preferences ofEuropeconsidered
from

This content downloaded from 202.28.191.34 on Fri, 18 Dec 2015 09:47:29 UTC
All use subject to JSTOR Terms and Conditions
GEOGRAPHICAL INTERPRETATION OF EIGENVALUES 81

theviewpointof residential desirability. While a portionof thementalimageof Europeheld

by eachstudentin a particular samplegroupwas quiteuniqueand peculiarto him, based as
it was upon his unique experiences, education,prejudices,etc., anotherportioncould be
consideredas beingsharedto formpartof an overall'mentalmap'. Thus two peoplein one
ofthenationalsampleswithverysimilarviewpoints wouldhaverankorderliststhatcorrelated
very highly; two others with diametrically oppositeviewpointswould have listswith a
strong inverse relationship between them, while the viewpointsof othersmightbear little
relationshipat all. Thus each national sample can be tabulatedas a data matrixin whichthe
countriesof Europeare orderedby thestudents, and thismay be translated intoa matrixof
rank correlationcoefficients. If the principaleigenvectors are extractedfromeach of the
matricessummarizing thepreferences of a nationalgroup,thecosinesof theanglesbetween
each of thesemakeup, in turn,a final,or higherordermatrixwhoseprincipaleigenvector is
an overallsummarypreference scale (Fig. I7). Such a scale seemsto reflectmixedconcern
MostPreferred LeastPreferred

.//
- 5b
/ /////c1
@ '-? ? o o-?Na <? >o <
1 40"
/ ///
rr S g 2 3 S 3>
o
^ 2 2 ? o ?0

FIGURE17-The summary preferencescaleforresidential in Europe.SamplesfromSwedish,

desirability
French,West Germanand Italianstudents.
British,

forthestandardof livingand thedegreeof authoritarianism

inpoliticallife.Amongthemost
countries
preferred are Sweden and Switzerlandwith democraticgovernments,
liberal, high
standards
of livinganda traditionof neutralism.
Nextcomethedemocratic socialistcountries
withpredominantly Protestant
churchaffiliation,
withtherestof WesternEuropescattering
down thescaleto thedictatorialregimesof Spainand Portugal.A considerable gap separates
thecountriesofEasternEurope,whichareled by Poland.Albaniabringsup thetail.

The Viewofan African Elite

The samesummarizing powerof theprincipaleigenvector of a correlationmatrix may
be usedto construct thementalmapsof peoplefortheirown country.In Ghana,as in other
Africancountries, universitystudents are likelyto entergovernment service,wherethepace
of advancement maybe veryrapid.The overallmentalmapsthatsuchpeoplehave of their
countrymay be of morethanpassinginterest, forinvestmentfunds,whetherforeconomic,
educational,medicalor generalsocialpurposes,mustbe assignedspatially as well as sectorally.
The majorview,derivedfroma sampleofstudents attheUniversity ofGhana,Legon(Fig.18),
scoreseachdistrictof Ghanaon theprincipalcomponentthataccountsfor67 percentof the
inter-correlation
inthesample.Thereisa highdegreeofspatialregularity anda steadylowering
of spacepreferencesas one movesnorthward away from thecoast. On thesecondnormalized
eigenvector, which becomes the second component when weightedby the square root
of theeigenvalue,regionalclustersarebroughtout as thecoreareasof Ashantiarecontrasted
againsttheeasternEwe bloc (Fig. 19). However,thisis muchlessimportant thantheprincipal
mentalmap illustrating theway in whichtheurbancentresof thecoastcastbrightimagesin
themindsofthesestudents. Yet one ofthecritical needsinnearlyall under-developed countries
is to encouragetheeducatedmenandwomento moveout ofthecitiesandintotheruralareas,

This content downloaded from 202.28.191.34 on Fri, 18 Dec 2015 09:47:29 UTC
All use subject to JSTOR Terms and Conditions
82 P. R. GOULD
as eitherteachers
or administrators,
andto catalyseand mobilizethepopulation.Thus,perhaps
thereare some implicationsherefora schemeof incentiveallowancesthatwill distribute
theseskilledpeoplein moreeffective
ways.

HIGH

LOW

FIGUREi8-Ghana: Major spacepreferences

forgovernment
assignment.

andSpeculationsfor
Thoughts Tomorrow
As geographers,
we arenot theonlyoneswho have to handlemultivariate problemsof
greatcomplexity.But, forsome reason,we seem to to
try compound the usualdifficulties
by
addingthesecond,or even thenthdimensionto problemsthatby anyreasonablestretch of
the imaginationare sufficiently
taxingin one. Under thesecircumstances, theory,always
partofa discipline
themostdifficult to develop,is boundto be hesitant,
ploddingandperhaps
untestableformanydecadesafterit is initiallypostulated.Despitethatmostmarvellousof

This content downloaded from 202.28.191.34 on Fri, 18 Dec 2015 09:47:29 UTC
All use subject to JSTOR Terms and Conditions
GEOGRAPHICAL INTERPRETATION OF EIGENVALUES 83
devices,themap,we maystillfinditdifficulttoextend ourunderstandingevenafter we have
abstracted, and
compressed simplified the realworld and modelled
it in such a waythatthe
spatial are
juxtapositions still
true.Yet, ifwe can a
go step further
and utilizeto thefullthe
informationthata mapportrays, it is possiblethatthesortofextensions thatwe cantruly
termgeographical
explanationswillcomewithgreater frequency.

-_ HIGH

[ii.:ii.ii:::
tiiiiiiiLiOi:
i I LOW

FIGURE
19-Ghana: Subsidiary
spacepreferences
contrasting Ashantiandeastern
thewestern Ewe regions.

One way of doing thisis by transformingthe informationon a map into formsthatare

more tractablefromthe mathematicallymanipulativepoint of view. In thisway, we can bring
to bear a varietyof calculi that may be able to break problems open, or provide measuresof
thingsand relationshipsthatwere difficultto grasp before. One of the many calculi available
to us is thatof linearalgebra and group theory,and withinthisvast fieldthe eigenvalue and its
associated eigenvectormay be usefulstaves that help us along the long road to spatial order

This content downloaded from 202.28.191.34 on Fri, 18 Dec 2015 09:47:29 UTC
All use subject to JSTOR Terms and Conditions
84 P. R. GOULD
and understanding. Indeed,perhapsmorethananyothersinglebranchofmathematics, linear
a
algebrarepresentskey thatcomes the closestto beinga skeleton.
We have seenthatifwe arrangedatain theformof squarearrayssomeusefulsummary
values,withdirectgeometrical as wellas algebraicmeaning,maybe forthcoming. Manymore
examples could be cited.Plantecologists, through the factor analyticapproach, have notonly
tackledthe problemof multivariate classification,
but, by manipulating plant association
matrices, have obtainedformalinsights intotheecologicalstructure underlying suchassocia-
tions(D. W. Goodall,I955). This typeof analysiscouldprovidea lead,perhaps,formedical
geographers wheredistinctive spatialpatterns of diseaseappearwhendata,normallyshown
in tabularform,are cartographically portrayed.Certainly,recentmaps of diseasein Great
Britainlead to manyspeculations abouttheunderlying 'dimensions' ofthedistributions.Flow
matrices,too, when analysedin thisway, may providesome insightsinto the functional
structureofa region,andthesemaybe extracted fromcommodity flowmapsbytheeconomic
geographer(Berry,1963,1966,1967),or frommapsof telegramtraffic representingadminis-
trativetiesbythepoliticalgeographer (B. Witthuhn, I967). Finally,we canextendourordering
of mapinformation byhandlingseveralmapsat once; forexample,ifwe areinvestigating the
sequentialnature of an unfolding spatialpatternthrough time (L. Tucker,1963,I964; J.Levin,
I963, I965). In such cases,the squarearraysof information extractedfromeach map are
'stacked'toforma cubicmatrix, so thatthecomponents, ordimensions, ofchangeareextracted.
While computational are and
problems severe, interpretation almost as increasesin
difficult,
thecapacitiesof computers arenow makingrealistic programs available.

ACKNOWLEDGEMENTS
The authorthanksMr. T. Leinbachforundertaking muchof thearduouscomputing work;Mr. B. Riddellfor
writingthe quadrateigenvalueprogram;Mr. A. Kardosfor allowingthe use of his fielddata on architectural
and Professor
characteristics; J.Kiusalaasforexplaining
elementary to the
aspectsoftensoranalysis.He is also grateful
PennsylvaniaStateUniversityfora granttowardsthecostof illustrations.

REFERENCES
ANDERSON,T. W. (1958)Introduction toMultivariate Statistics
(New York).
ANDREE, R. V. (I959) Selections fromAbstract Algebra(New York).
BARNARD, M. M. (I935) 'The secularvariations in fourseriesof Egyptianskulls',Ann.Eugen.,
of skullcharacteristics
352-71.
BASSETT, K. (1967) Socio-Economic andSpacePreference RegionsoftheSouth;A Comparative Study,unpub.M.S. Thesis,
University Park,Pennsylvania.
BERLINER, J.(1962)'The feetofnativesarelarge:an essayon anthropology by an economist',
Curr.Anthrop., 3, 47-77.
BERRY,B. J.L. (I960) 'An inductive approachto theregionalization ofeconomicdevelopment', in N. GINSBURG(I960).
BERRY,B. J.L. (I96I) 'A methodforderiving multi-factoruniformregions',Przegl.Geogr.,33,263-82.
BERRY,B. J.L. (i963) StructuralComponents ofChanging FlowNetworks
Transportation (FortEustis,Virginia).
BERRY,B.J. L. (I964) 'Approaches Ann.Ass.Am.Geogr.,54,2-II.
to regionalanalysis:a synthesis',
BERRY,B. J.L. (I966) Essayson Commodity FlowsandtheSpatialStructure oftheIndianEconomy (Chicago).
BERRY,B.J. L. (1967) 'A synthesisofformalandfunctional usinggeneralfieldtheoryofspatialbehavior'
regionalization
in B. J. L. BERRYand D. MARBLE(I967).
BERRY,B.J.L. andRAY,M. (I964) Multivariate Socio-Economic A PilotStudyinCentral
Regionalization; Canada(Charlotte-
town,PrinceEdwardIsland).
BERRY,B.J.L. andMARBLE, D. (I967) SpatialAnalysis(Englewood Cliffs,
NewJersey).
BICKLEY,W. G. andTHOMPSON, R. S. H. G. (I964) Matrices;
TheirMeaning andManipulation(London).
BOWDEN, L. (I965) Diffusion of theDecisionto Irrigate; oftheSpreadofa New Resource
Simulation Management Prac-
ticein theColoradoHighPlains(Chicago).
BUSH, R. and MOSTELLER,F. (I955) Stochastic Modelsfor Learning
(New York).

This content downloaded from 202.28.191.34 on Fri, 18 Dec 2015 09:47:29 UTC
All use subject to JSTOR Terms and Conditions
GEOGRAPHICAL INTERPRETATION OF EIGENVALUES 65
CARNEIRO, R. L. and TOBIAS, S. (1963)'The applicationofscaleanalysis to thestudyofcultural evolution',Trans.N.Y.
Acad.Sci.,I96-207.
CATTELL, R. B. (I965a) 'Factoranalysis:an introduction PartI: the purposeand underlying
to essentials. models',
Biometrics, 21, I90-215.
CATTELL, R. B. (I965b) 'PartII: theroleoffactoranalysis in research',
Biometrics,21,405-35.
CATTELL, R. B. and SULLIVAN, W. (I960) 'The scientificnatureof factors:a demonstration by cupsof coffee',Behavl.
Sci.,7, I84-93.
CHURCHMAN, C. and RATOOSH, P. (1959)Measurement; DefinitionandTheories (New York).
COLENUTT, R. (I966) LinearDiffusion unpub.M.S. Thesis,University
inan UrbanSetting, Park,Pennsylvania.
COOLEY, W. W. andLOHNES, P. R. (I962) Multivariate
ProceduresfortheBehavioral Sciences
(New York).
COOMBS, C. H. (1964)A Theory ofData (New York).
COOMBS, C. H. and KAO,R. C. (I955) Nonmetric FactorAnalysis (AnnArbor).
COOMBS,C. H. and KAO,R. C. (I960) 'On a connection betweenfactoranalysisand multidimensional unfolding',
Psychometrika, 25, 219-31.
CURRY, L. (I966) 'A noteon spatialassociation', Prof.Geogr.,18,97-9.
DAVIS,P.J. (I965) TheMathematics A FirstBookon MatrixTheory
ofMatrices; andLinearAlgebra(New York).
FISHER,R. A. (1936)'The use ofmultiplemeasurements in taxonmicproblems', Ann.Eugen.,7, 179-88.
FRAZER, R. A., DUNCAN, W. J. and COLLAR, A. R. (1963)Elementary MatricesandSomeApplications to Dynamicand
Differential Equations(Cambridge).
FREDRICKSON, N. andGULLIKSEN, toMathematical
H. (1965)Contributions Psychology (New York).
GARRISON, W. andMARBLE, D. (1965)A Prolegomenon totheForecasting ofTransportation
Development(Evanston, Illinois).
GINSBURG, N. (I960) Essayson Geography andEconomic Development (Chicago).
GOODALL, D. W. (1955) 'Objectivemethodsfortheclassification of vegetation:Part3, an essayin theuse of factor
analysis',Aust.J. Bot.,2, 305-24.
GOULD,P. R. (I965a) 'Joshua'strumpet:the crumbling walls of thesocialand behavioralsciences',Geogr.Rev., 55,
599-602.
GOULD,P. R. (I965b) A Bibliography ofSpaceSearchingProceduresfor
Geographers
(University Park,Pennsylvania).
GOULD,P. R. (I966a) On MentalMaps(AnnArbor,Michigan).
GOULD,P. R. (I966b) 'Questionsarisingfromthedescription ofmentalmapsofBritishschoolleavers',NSF Conference
New Trends in Cultural Geography(Columbus,Ohio).
GOULD,P. R. (I966c) SpaceSearching inGeography
Procedures andtheSocialSciences
(Honolulu,Hawaii).
GOULD,P. R. (1967)'Structuring informationon spacio-temporal reg.Sci. (forthcoming).
preferences',J.
GOULD,P. R. etal. (I96I) 'Towardsa geography ofeconomichealth:thecaseofNew YorkState',Ann.Ass.Am.Geogr.,
52, 1-20.
GUTTMAN, L. (I966) 'The non-metric breakthrough in the behavioralsciences',Automatic Data ProcessingConference
(Jerusalem, Israel).
HAGERSTRAND, urkorologisk
T. (I953) Innovationsforloppet synpunkt (Lund,Sweden).
HAGERSTRAND, T. (I957) 'Migrationandarea',Migration inSweden(Lund,Sweden).
HXGERSTRAND, T. (1965)'A MonteCarloApproachto Diffusion', Archives
Europe'annesdeSociologie,
6, 43-7.
HAGOOD, methodsforthedelineation
M. J.(I943) 'Statistical ofregionsappliedto dataon agriculture andpopulation'
SocialForces, 2I, 287-97.
HALL,E. (I959) TheSilentLanguage (New York).
HARMAN, H. H. (I960) Modern FactorAnalysis(Chicago).
HARRIS,C. W. (1963)Problems ofMeasuring Change(Madison,Wisconsin).
HAWKINS,D. (1964) TheLanguage ofNature;AnEssaytnthePhilolosphy ofScience(San FranciscoandLondon).
HEISENBERG, W. (I930) ThePhysical PrinciplesoftheQuantum Theory (New York).
HIGMAN, B. (1964)AppliedGroup-Theoretic andMatrixMethods (New York).
HOFFMAN, B. (1963) TheStrange StoryoftheQuantum (Harmondsworth).
HOLLOWAY,J. L. (I958) 'Smoothingandfiltering oftimeandspacefields', Ad. Geophys.,4, 351-89.
HOLLEY,M. A. (I966) Topographic Eigenvector Analysis(Downey,California).
HORST,P. (I966) FactorAnalysis ofData Matrices (New York).
HOTELLING, H. (1936)'Relationsbetweentwo setsofvariates', Biometrika,28, 32I-77.
W. R. (I966) TheFourier
JAMES, SeriesModelinMap Analysis (Evanston,Illinois).
M. B. (I958) MolarCorrelational
JONES, Analysis (Pensacola,Florida).
M. B. (1959)SimplexTheory
JONES, (Pensacola,Florida).
KANAAN, N. (1965)Structure andRequirements oftheTransport Network ofSyria(Washington).
KENDALL, M. G. (I939) 'The geographical distribution ofcropproductivity inEngland',J.Roy.statist.Soc.,I02, 21-62.
KENDALL, M. G. (1957)A CourseinMultivariate (Londonand New York).
Statistics
KRON,G. (1959) Tensors forCircuits
(New York).
KRUMBEIN, W. C. (I966) 'Classificationof map surfaces basedon thestructure of polynomialand Fouriercoefficient
matrices',in D. F. MERRIAM (1966).

This content downloaded from 202.28.191.34 on Fri, 18 Dec 2015 09:47:29 UTC
All use subject to JSTOR Terms and Conditions
86 P. R. GOULD
B. (I964a) 'Multidimensional
KRUSKAL,J. goodnessoffitto a nonmetric
scalingbyoptimizing Psychometrika,
hypothesis',
29,
I-27.
KRUSKAL,J.B. (I964b) 'Nonmetric scaling:a numerical
multidimensional method',Psychometrika,
29,II15-29.
LEVIN, J.(I963) ThreeModeFactor Analysis,unpub.Ph.D. dissertation, Urbana,Illinois.
LEVIN, J.(I965) 'Threemodefactoranalysis', Psychol. Bull.,64,442-52.
LINGOES, J. C. (1964)'New computerdevelopments in patternanalysisand nonmetric tehniques', Symposium on Com-
putersin Psychological Research (Blaricum,Netherlands).
LINGOES, J. C. (I966) 'Recentcomputational advancesin nonmetric methodology forthebehavioralsciences', Interna-
tionalSymposium onMathematical andComputational Methods in SocialSciences (Rome).
McGINNIS,R. (1965)Mathematical Foundationsfor SocialAnalysis (New York).
MERRIAM, D. F. (1966) Computer Applicationsin theEarthSciences;Colloquium on Classification Procedures (Lawrence,
Kansas).
POLYANI, M. (1963)The StudyofMan; TheLindsayMemorial Lectures I958 (Chicagoand London).
PRESTON, F. W. (1966)'Two-dimensional powerspectraforclassification oflandforms', in D. F. MERRIAM (1966).
SCHEIDEGGER, A. E. (1965)'On thestatisticsoftheorientation ofbeddingplanes,grainaxes,andsimilar sedimentological
data',Geol.Surv.Res. CI64-CI67.
SEWELL, W. R. (I966) HumanDimensions ofWeather Modification(Chicago).
SHEPARD, R. N. (1962) 'The analysisof proximities: multidimensional scalingwithan unknowndistancefunction',
Psychometrika, 27, I25-39 and 219-46.
SHEPARD, R. N. (1963) 'Analysisof proximities as a techniqueforthestudyof information processing in man',Hum.
Factors, 5, I9-34.
SOJA, E. (1966)'The spatialdimensions of modernization in Kenya',African Stud.Ass.(Bloomington, Indiana).
SOKAL,R. (I966) 'Numericaltaxonomy',Scient.Am.,215, io6-I6.
SOKAL, R. and SNEATH, P.(I962) 'Numericaltaxonomy',Sci.,4818,855-60.
SOKAL, R. and SNEATH, P.(I963) PrincipalsofNumerical Taxonomy (San Francisco andLondon).
SOKAL, R. and CAMIN,J.(i965a) 'A methodfordeducingbranching sequencesin phylogeny', Evolution, 19,3II-26.
SOKAL, R. andCAMIN, J.(I965b) 'The twotaxonomies:areasofagreement andofconflict', Syst.Zool.,I4,I176-95.
THOMPSON, G. H. (I95I) The Factorial Analysis ofHumanAbility (London).
THOMPSON,G. H. (I956) The Geometry ofMentalMeasurement (London).
THURSTONE,L. L. (I947) Multiple FactorAnalysis (Chicago).
TOBLER,W. R. (I966a) Numerical Map Generalization andNotesontheAnalysis ofGeographical Distributions(AnnArbor).
TOBLER, W. R.(I966b) 'Of mapsandmatrices', (mimeo).
TOBLER,W. R. (I966c) 'Spectralanalysis ofspatialseries',Conference onUrbanPlanning, Information Systems andPrograms
(Berkeley,California).
TORGERSON, W. S. (I958) Theory andMethods ofScaling(New York).
TOULMIN, S. (I960) ThePhilosophy ofScience(New York andEvanston).
TROWBRIDGE, C. C. (I939) 'On fundamental methodsof orientation andimaginary maps',Sci.,38, 888-97.
TUCKER,L. (I965) SomeMathematical Noteson ThreeModeFactor Analysis (Urbana,Illinois).
TUCKER, L. (I963) 'Implications of factoranalysisof threeway matrices formeasurement of change'in C. W. HARRIS
(I963).
TUCKER,L. (I964) 'The extension offactoranalysis to threedimensional matrices' in N. FREDRICKSON andH. GULLIKSEN
(I965).
TUKEY,J. B. and ABELSON,R. P. (I959) 'Efficient conversionof nonmetric information into metricinformation',
Proc.Am.statist. Ass.,226-30.
WALKER,F. E. (I963) SomeElementary ConceptsofMatrixAlgebra anditsApplication (The Ohio StateUniversity).
WHITE,R. (I967) TheMentalMapsofBritish SchoolLeavers, unpub.M.S. Thesis,University Park,Pennsylvania.
WILLIAMS, W. T. andLAMBERT, J.M. (I96i) 'Multivariate methodsin taxonomy',Taxonomy, I0, 205-II.
WILSON,A. (I966) 'The impactofclimateon industrial growth:Tucson,Arizona:a casestudy'in W. R. SEWELL (I966).
WITTHUHN,B. (I967) TheSpatialIntegration ofUgandaas a Process ofModernization, unpub.Ph.D. dissertation, Univer-
sityPark,Pennsylvania.
YUILL, R. (I964) A Simulation StudyofBarrier Effects in SpatialDiffusion Problems (Evanston, Illinois).

This content downloaded from 202.28.191.34 on Fri, 18 Dec 2015 09:47:29 UTC
All use subject to JSTOR Terms and Conditions