VOL. 20 NO. 2 (1997) 383-392
RESEARCH
NOTES
RANDOMPROCESSES
WITHCONVEX
COORDINATES
ONTRIANGULAR
GRAPHSJ.N.BOYD P.N. RAYCHOWDHURY
Department
of Mathematical SciencesVirginia CommonwealthUniversity Richmond Virginia 23284-2014 U.S.A.
(Received September 28, 1994)
ABSTRACT. Probabilitiesforreaching specifieddestinationsandexpectation valuesfor lengths for randomwalksontriangulararrays ofpoints andedges arecomputed. Probabilities and expectation values aregivenasfunctionsof theconvex(barycentric)coordinatesof the
starting point.
KEYWORDS AND PHRASES. Randomwalk,expectationvalue, trap,convex(barycentric)
coordinates.
1991AMSSUBJECT CLASSIFICATION CODES. Primary60J15, Secondary 60J05.0
1. INTRODUCTION.
Barycentric coordinateswereintroduced into mathematicsbythegeometerA.F. MSbius
(1790-1868).
SupposethatVz,V2,V3,...,Vn,V,+
arethelinearlyindependentverticesofan n-dimensional Euclidean simplex. The barycentric coordinates of point Pin the Euclidian spacen+l n+l
aregiven by theunique
(n
+ 1)-tupleof realnumberscqsuch thatP=
c,Viwithy
c,=l.[1]
==1
The points of the closedsimplex
S,
VzV2V
...VnV
+ arethose points for which allbarycentriccoordinatesarenonnegative. The barycentric coordinatesofvertex
V,
aregiven by a, 1,aj 0 for i. Inourpaper,weare concernedonlywith points of the closedsimplex, andwewillreferto(a
z,o2,a3a,)
astheconvexcoordinatesofPeS,
witha’especttothe vertices intheorderV,V2,V
3Vn,V,
+1"However,
ourprimaryconcernwillbe withtheclosedtriangularregion
VzV:V
3.Convex
(or
barycentric)coordinates admit ofphysicalinterpretation. The pointPis the center ofmassforthe discrete distribution ofpointmassesa,atV,
for 1,2,3, n,n+l. Thenotionsof balanceand mechanicalequilibrium permitusto discover theconvexcoordinatesofvariousspecialpointsof triangles and tetrahedra andtorelate the geometryoftrianglesto mechanics.
[2]
Convex
coordinatesaxequite versatile mathematical objects. Their invarianceunderaffine transformations has manyapplicationsingeometryand theirinterpretationaslengths,
areas,andvolumesleadstoawaytocount partitions of theintegers.
[3]
Since 0 < a, < for each pointof
S,,
ithasalsoseemednaturaltoseek probabilisticinterpretationsand applicationsfor convex coordinates. Suchinterpretations andapplications
A
RANDOM
WALK ON ALINE
SEGMENT. Letus consideraline onwhichthereisdefinedaCartesiancoordinatesystemsothat pointsV andV of have Cartesian coordinates x 0andx 1,respectively.
Next,
letusconsidern- points,AI,A2,...,A
n 1, equallyspacedbetweenV and
V2
asshowninFigure1. For convenience in notation, letusgiveV andV2
thealternatenamesA0and
A,,
respectively. The collection ofnsegmentsA,A
+1, 0,1,2,...,n-1, andn
+
1 pointsAs,
0,1,2 n,constitutes a onedimensionalgraph. Theconvex coordinatesof pointA,
withrespecttoVI=A
oandV=A,
are(al,
)=
(1-
,
i)for
0,1,2 n.
FIGURE1. TheOneDimensional Graph.
Nowletussupposethatarandom walk takesplaceontheone-dimensionalgraph. If the walk has reached
A,
1,2,...,n- 1, thenextstepmustbeeithertoA,_
with transitionprobability p(Ai-* Ai_
1)
ortoA,
+ alsowith transitionprobability p(A,A,
+1)
End points
V1=A
oandV=A,
serve astraps. That is,p(AoA)
p(A,A,_
1)
0. The walk mustcontinueuntil itreacheseitherVI=A
oorV=A,.
Letp(Ai) denote the probability thatawalk starting from
Ai
reaches endpointV.
Then p(Ai)mustsatisfy theequationp(Ai) p(Ai Ai_
1) p(A) +
p(AAi
+1)P(A,
+1)
21-
((A,_,)
+
((A,+,))
()
for 1,2 n-1, subjecttotheboundaryconditionsp(Ao)=O, p(An)=l.
A
functionsuchasp(A)is anexample ofadiscrete harmonicfunction;and,ifp(A) satisfiesboth1and the boundary conditions,itdoessouniquely.
[7,9]
Itissimpletoshow that p(A)=a2,theconvex coordinateof
A
withrespect toV2=A,
meets the requirements given above. Thustheprobability thatawalk startingat
A
with terminate atV
isa.
Likewise, the probability that the walk willterminate atV1
isq(A)=l- p(A)
1-a
a1.THE EXPECTED LENGTH OF THE
ONEDIMENSIONAL RANDOM WALK.
Each walkbeginningatA
eventuallyterminates atVA0
orV
An.
LetE(A)
denote the expectedlengthofawalkasmeasuredby thenumberof steps taken from startingpointA
toeitherendpoint. Sincethe walk beginswithasinglestep frominteriorpoint
A
toeitherA,_
or
A
+1,wecanwriteE(A)=
p(A-A_I)-(1
+
+
p(A-,A+I).(1
+
E(A+))
(2)
1/2
(E(A_1)+
E(A,
+))
+
1.Sinceawalkbeginningat eitherendpointwillhavezerolength,ourboundary conditionsare that
E(A0)
0andE(A)
0.Supposethatboth
E(A)
andE(A)satisfy 2 and the boundary conditions. Then F(A,)E(A)
E’(A)satisfies1 withboundaryconditionsE(Ao)
E(A,)=0andF(A) 0. ThusF(A)
isadiscreteharmonicfunctionand must be uniquely given byF(A)
0 for 1,2,n.Therefore
E(A)
E’(A),
and any functionsatisfying2 while vanishingatthe endpointsisSince E(Ai)
n2ala2
satisfies 2andthe convex coordinatesofV1,V2
are(1,0), (0,1),
respectively,wehave theexpectedlength. That is, theexpectedlengthofarandomwalk
starting from
A,
ontheonedimensionalorlineargraphofn+
points andnsegmentsisE(A,)n2ac
wherec,1,c
aretheconvexcoordinatesofA,.
A
RANDOMWALK
ON A TRIANGLE. Consider theclosed triangular regionnamedbyitsvertices as
V1V2V
3. Forconvenience,wetake the triangletobe equilateral. Weconsidereachsideto beacopy oftheonedimensionalgraphofn
+
1equallyspacedpointsandn connecting segmentsdescribed above. Ifaline is drawn withinVV2V
paralleltosideViVa
then each point of thelinehas thesameconvexcoordinatea,,k i,j
Suppose,
then,that threelinesaredrawnsothat each isparalleltoadifferent side ofthe triangleandsothatthe threelinesareconcurrent. Eachisalineof constantconvexcoordinate withrespect to a different vertex. Thustheconvexcoordinatesfor the pointPcommontothe threelinescanbe read from the pointsat whichthelines intersectthesideof the triangleasindicated inFigure2.FIGURE2. TheConvexCoordinates ofP:
(a,b,c).
Nowletus"drawthree setsofn- lineseachparallel tooneofthe three sidessothat the lines dividethesides inton congruent segments.
In
Figure 3,weshowthe three setsoflinesfor thecasein whichn 6. Theinterior intersectionpoint
P
hassixadjacent points.FIGURE3. The Triangular Graphforn 6.
Thearrayof points andsegmentsconstitutesatwodimensional, triangular graphupon
which wewillconsiderarandom walk. Each stepis amovebetween adjacent points of the
graph alongthe segment connecting them. Ifwedenote the points adjacentto interiorpoint
P
by
A,
for 1,2,3,4,5,6,wecanstatethe followingrulesforthe random walk.1 1. The probability ofamovefrom
P
toA
isgiven byp(PA)
.
2. Oncethewalk reachesaboundary point, the walkmustremainontheboundary. 3. Theverticesofthetriangleserveastraps. The walk mustterminateonceitreaches
The probabilityofa movefromany boundarypoint otherthanavertex toeitherof
its twoneighboring boundary pointsis
1/2.
That is,oncethe walkhasreachedasideof thetriangle, the walk becomesaonedimensional walkasdescribedabove.
Letusnowsuppose that Figure3 represents thegeneralcaseforthetriangulargraph
with n > 3. Then,iftheconvexcoordinates of interiorpoint
P
area(P)
(al,a2,
as),
the convexcoordinatesof its nearestneighboringpointsare(A,)
(,-
,
+
,
),
()
(,-
,
,
+
),
()
(,,,-
,
+
),
(,)=
(,
+
,
,-
,
),
Byect
computation,westhatH
P
is udypoint shown()
((,)
+
(n,)).
(3)
FIGURE4. Point
P
onSideViVj
Letp(P)denote theprobabilitythatarandomwalkstarting frompointPofthe
triangulargraphwillreachvertexV1. Thefunctionp(P)isagainadiscrete harmonic function.
Itsatisfiesthecondition
6 6
for eachinterior pointP. Ontheboundaryp(P)
1/2
(p(A1)+ P(A2))withthenotationtkea fromFigttre4.Equation 3 and the results for theone dimensionalwalk imply that p(P) al. That is, the probability ofreaching
V1
starting from pointP
is al. Likewise, the probabilities of reachingV
andVsarea and as, respectively.The walkcanbeextended tographsontetrahedron
V1V2V3V
4. Each facecontainsatriangular graphand eachedgecontainsalineargraph. The probability ofarandom walk’s
reachingvertex
V,
1,2,3,4 from starting pointP
isa, theconvex coordinate ofP
withrespectto vertex
V,.
The walkcanbeextendedinductivelytographson the k-simplex, each face of whichisa(k-1)-simplex.ELECTRICAL APPLICATION
1.For
eachrandom walk thereexists anelectrical potentialproblemhaving thesame solution withtheonlydifferencebetween theproblemsbeing thephysicalmeaning of the variables.[4]
Imaginethatthe triangular gridofFigure5is anelectricalnetworkof identicalresistors in whicheach segment connecting adjacent points hasresistance
R.
LetthepotentialatV1
bemaintained at 1voltwhileall pointsof
V2Vz
areheldatzeropotential. The potentialsonsidesVIV
andV1V
3 decrease linearly from to0 volts. Thus the boundary conditionsonpotential match those oftherandomwalk.
orv(P)= v(A,). Thus theaveragevalueproperty given by Equation3 issatisfied. We
can
conclude’’ht
theelectrical potential atpointP
isa.
Figure5. AnElectricalNetwork.
THE
EXPECTED LENGTH OFTHE
TRIANGULAR RANDOM WALK. Letus supposethatarandom walkstartsfrom pointP
of the triangulargraphand terminates atone ofthevertices. Letusdenote theexpectedlengthofthewalk byE(P).If
P
isaninterior pointhavingnearestneighborsA,,
1,2,3,4,5,6,itfollows thatE(P) (I+E(Ai))
+
E(A,).=1 =1
()
If
P
isabounda.,Tpointwithbounda
neighborsA1, A
onV,Vj,
wealready know thatE(P)
n2aiaj
satisfiesE(P)=
1+1/2
(E(A1) +(E(A2)).
Furthermore,E(V)
E(Vi)
0.SinceE(P)
n2(la
+
aa
3+
aas)
satisfiesEquation4and also reducestotheboundarysolution,itfollows thatwehavediscoveredthe expectedvalueof thelengthof the randomwalk.
That is, the expectedlengthofawalk starting from point
P
of the triangulaz graphisE(P)
n2(ala2
q-ala q-a2a3)
wherea, a,azeconvexcrdinatesof withrt
toV,
V,
Vz
inthat order.AN EXAMPLE.
Letn 12. Thecentroidof the trigle hconvexcrdinates(,
,
)
,disapoint of theaph.The probability thatawMk stating
om
thecentroidwill__reach
aptic..x
vtexis]steps.
COMMENTS ON BOUNDARY CONDITIONS. A nst
problemconcernsthe cotructionofdiscretehocfunctionsontheinteMorpoints of thetrigul aphwhen the functionserequiredtomt bitrlysetboundyconditions. Weconsider theproblemwitha
sci
ce.Letuschgethe fourth ofourroles for rdomwksontMglesby supsing thatthe entireboundyof
V]VVz
serves trap.In
otherwords, the rdom wk mustcome to aht wheneveritrhbound
point. Letusfindorindicatehowtofinddiscretehonicnctionsintesof], a,d
z
inclosed fowhichwill give theprobabityofdenote the desiredprobability by p(P).
Since thevertices
V1,V,V
3cannot be reached directly fromany interiorpoint of thearray, wecanexclude themfromourcalculations.
In
Figure 6,weindicatethe triangulargraphwithV
1,V2,V
3 removed.v!
Figure6.
AV1V2V
3{VI,V,V3}.
The boundaryconditions forp(P)arep(P) 0ifPe
V1V2
UVV
{V,V,V3}
andp(P) ifPe
VV
{V2,
V3}.
Oncewefindafunction p(P} satisfyingboththeseboundaryconditionsand the average value property given by 3, that probability function will beunique.
However,
the forms of thefunctions will notbeunique. Forexample, p(P) ot+
a
isharmonic,yet ithasthedifferent, but equivalent,form p(P) 1 03. Furthermore,sincethe
functionsneedbe evaluated onlyat a finitenumber of points,variouscombinationsof
continuousfunctionsmaytakeon thesamevalues.
A
bitof calculationindicatesthatp(P) cosnra cosnrct co3rtrct3satisfiesthe average value property, but the value ofp(P}is at allinteriorpoints of thegraphif nis evenandifnisodd.
Thisexample involvingthe trigonometric functionssuggeststhatweinvestigateproducts of exponentialfunctions inoursearchfor other,moreuseful, harmonicfunctions. Therefore,
suppose that
f(P)
analbna2c
ha3for nonmegative real numbers a,b,c. Wecan rewritethe na nl a n2function asf(P)=
ana’bnC’2c
n’nc’1"na=c()
()
implyingthatweneed only considerproductsof theformp(P) ana,b
na=.
If p(P)
analb
ha2isto satisfy the average value property,werequire that
Simplifyingthisequation,wefindthat
a’Ib+a-+b
"l+ab
"+a+b=6
or(a+l)b
2+(a
6a+l)b+(a
2+a)=0.
In
order thatbtakeonrealvalues,it isnecessarythat the discriminantD
be nonnegative. The discriminant takes the simplified formD
(a- 1) (a
-
14a+
1)
implyingthatashouldbe chosen from
(-cx,
?-4x/r]
U[7
+
4V/,
cx). In
addition,neithera norb may bezero.Sinceweintend tousesumsoffunctionsof the form
(constant)-a’mlb
ha2tosatisfy theboundaryconditions of ourproblem, letuschooseaand btoyieldassimple calculationsas
possible. Tothat end,welistall of the admissible values ofathat give discriminants whichare
-6 14 15 20
5929
169
3136
43681
S or 15 21 10
-
or"T
-6 or -12-35-
orTable 1. Values of a,D,b.
Since therolesofaand b may beinterchanged,wecanconstructfifteen different functions ina anda whichhaveoneintegral basewhilethe other baseis rational.
In
addition,we can usethepairs of convex coordinatescq,c,3orc,2,aaasindependentvariables. Of
course,aand b need not be rational, butourimmediate purposeis toproduce simple examples.
In
Figure 7,weshowthe simplegraphforn 3. The probability that awalkstartingatP,
the onlyinteriorpoint,willreachapointofVV
aisclearly1/2.
However,
we usethisgraph as atestcasetodisplaythe processfor constructingadiscrete harmonic functiongiving thedesiredprobability.
Figure 7. TheCaseforn 3.
Theproblemis symmetric withrespectto
a
andaa.
Sincesumsof harmonic functions areharmonic, letusfind constantsA,B,C, andD
sothatp(P)
[z,(-)
+
*(-)
+
Bo,
+
C,
+
D
(S)
vishesif
(a,a,a)
0, 0, d h thevMue if(a,
a3)
0,,
(0,
,
)}.
Notethatsyetryimphes thatifp(P) vishesonVV,
it willMso
vhonVV.
Substitution of appropriate values ofa,a,aintoEquation 5 yields thefollonglineequations:
-3375A+2B
+
D=0, 1665A+ B +2D=0, 90A / C+2D=3,90A
+2C+
D=3.Solvingthese equations,weobtain
A
1/2205, B
187/147,
CD
149/147.
Checkingourwork byevaluationof the probabilityfunction withthesecoefficients atFigure 8. TheCaseforn 4.
probabilities ofreaching
V2V
3 from,
thatprobabilitybyx. The averagevalueproperty impliesthat
0+0+l+l+z+y 0+0+0+0+x+z
6 =x and 6 =Y"
Thus
Nowletusconstructp(P)andthencompareourfunction values with x and y. Recalling that theform of p(P)is notunique,wesimply chooseourformforreasonsofconvenience. Let
p(P)
A[154a’(_6)
’In,
+
154a’(-6)
4%
+
Oa
+ E% +
Fa.
Using theconvexcrdinat of points from
VV
vV
{V,V,V}
d noting thatE
F,
we cwritethe equations-16874A 8780.8B 5062.5C
+
.75D+
.25F 0 8325A+
3653.44B+1631.25C
+
.5D+
.5F 0-3225A -1023.232B- 219.375C
+
.25D+
.75F 022A -78.288B 18.125C
+
F
172A
+
35.28B+
12.5C+
F
A
computer programtoimplement aGaussianeliminationprocedureyieldsA
1.425 x 10"4, B
1.029 x 10-4 C -8.189 x 10"4, D
-1.449, andE
F
0.996. Evaluatingeasilyexplainedbyround-offerrorsincomputation.
Weconclude thatlinearcombinations offunctionsofthe formahal b
ha2,
ahal bn%,
an% bn%,
andc1,c2,o3willproducediscrete harmonic functionssatisfyingarbitrarilyset boundaryconditionsfor triangular random walks. Unfortunately, the computations do notseemanesthetically pleasing, but they dogeneralizethe subject.
ELECTRICAL APPLICATION
2.Let
usconsider thenetworkbelow whichisderivedVIV
{Vl, V}
andVlV
{Vl,
V}
aremaintained atapotential of+1volt.Figure9. AnotherElectricalNetwork.
1 3
Itfollows fromourlastcomputationsthat the potentialsatpoints
P,
and
,
respectively.THE
EXPECTED LENGTH OFTHE
RANDOMWALK WHEN THE BOUNDARIES
ARE
TRAPS.In
thecasethat all points of the boundary oftriangular graphV1V2V
3serveastraps, the expectedlengthofarandom walk frominteriorpointP:
(ch,c,,a3)
isgiven byE(P)
3nala2a3
Thisresult follows from the fact thatE(P)=
Again,theaverageistakenoveradjacentpoints
A,,
1,2,3,4,5,6. Theclaimcanbe verifiedquite easily bydirect computationoverthesix pointsadjacent to
P
in thegraphhavingn+
pointsoneachside. Itisalso clear thatE(P)vanishesontheboundaryof thetrianglesince at
leastone convex coordinate mustbezeroat anyboundarypoint. The resultisunique by argumentssimilar tothose givenin theonedimensionalcase.
In
the simplecasewithn 3, E(P) 1forthe walkstarting fromthe centroid. Ifnn 12and
P
isthecentroid, thenE(P)
16.FURTHER COMMENTS
ON EXPECTED LENGTHS. Again,it isthe result ofstraightforwardcomputationthat, forinterior point P:
(a,a,c,),
each of thefollowingfunctions
n2aa,
n22a3, n2al3,
andn2l2a3
satisfies theaverage valuepropertyf(A)+1/2
f(P).
Itfollows thatE(P)
n(aaa
+
bac,
+
caa3+
da=a3)
satisfiesEquation4 andbecomesacandidateforanexpectedlengthifthereal coefficients,a,b,c,d,sumto 3.
For
example, ifa b c 1, d 0,wehave theexpectedlengthfor the walk in which the verticesaretheonly trapsand for whichoriginal boundarycondition 4holds.Ifa b c 0, d 3,wehavethe expectedlengthif allboundary pointsserve as traps.
Suppose
thatallpoints ofVV=
UV1V
serveastrapsbut that boundarycondition 4holdstrueon
VV3.
ThenE(P)isgiven bya 0,b 1. c 0, d 2sincesuchafunctionuniquelysatisfiestheboundaryconditions.
REFERENCES
GRAY,
J. "MSbius’ GeometricalMechanics," inMSbius.and
His Band(J.
Fauvel,R. Flood, andR. Wilson,Editors),OxfordUniversity
Press,
New York,pp.78-103, 1993.
HANSNER,
M. "TheCenterofMassand AffineGeometry," Th._.eAmerican MathematicalMonthly69(October 1962),
724-737.BOYD, J.N.,
andRAYCHOWDHURY,
P.N. "CountingPartitions oftheIntegers
withConvexCoordinates," Bulletinof Number TheoryXV
(1991),
28-37.DOYLE,
P.G.andSNELL, J.L.,
"Random WalksandElectrical Networks"