HAL Id: jpa-00246941
https://hal.archives-ouvertes.fr/jpa-00246941
Submitted on 1 Jan 1994
HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.
Linear operators on correlated landscapes
Peter Stadler
To cite this version:
Peter Stadler. Linear operators on correlated landscapes. Journal de Physique I, EDP Sciences, 1994,
4 (5), pp.681-696. �10.1051/jp1:1994170�. �jpa-00246941�
Classification Physics Abstracts
0250 0550
Linear operators
oncorrelated landscapes
Peter F.
Stadler(*)
Santa Fe Institute, 1660 Old Pecos Trail, Santa
Fe,
NM 87501, U-S-A-(Received
26 October 1993, received in final form 12January
1994,accepted
28January 1994)
Abstract. In this contribution we consider the elfect of a class of
"averaging operators"
on
isotropic
fitnesslandscapes. Expbcit expressions
for the correlation function of theaveraged landscapes
are derived. A new class oftunably rugged landscapes,
obtainedby
iterated snlooth-ing
of the random energymodel,
is established. The correlation structure of certainlandscapes,
among them the
Shernngton-Kirkpatrick
model, remainsunchanged
under the action of ailaveraging operators considered here.
1 Introduction.
Evolutionary adaptation
as well as combinatorialoptimization
takeplace
on"landscapes"
that result frommapping (micro)configurations
to scalarquantities
like fitness values orenergies,
or, more
generally,
to nonscalarobjects
hke structures. Theassumption
of aparticular
mech- anism formutation,
or the choice of aspecific
move-set forinterconverting configurations,
introduces trie notions of
neighborhood
and distance betweenconfigurations. Viewing
trie set of ailconfigurations, 1e.,
trieconfiguration
space, as a non-directedgraph
r bas tumed out to beparticularly
useful. Eachconfiguration
is a vertex ofr,
andneighbonng configurations
areconnected
by edges.
Whenconfigurations
are sequences ofequal length
and trieneighborhood
relations are defined
by
the number ofdiffering positions,
one obtains trie sequence space with theHamming
metric.Among
the mostimportant properties
of alandscape
is its"ruggedness",
which influences thespeed
ofevolutionary adaptation
as well as triequality
of solutions of heunsticoptimization strategies.
Whileruggedness
bas never beenprecisely defined,
a number ofempincal
measures bave beenproposed,
such as trie number of localoptima
or trie averagelength
of up- or downhill walksil, 2],
and thepair-correlation
as a function of distance mconfiguration
space[3, 4].
The very notion of
"uphill"
or"optimum" implies
a mapping to a scalarquantity.
Trie definition of apair correlation, however,
con be extended to mappmgs toarbitrary objects,
as(*)
Address forCorrespondence:
Institute for TheoreticalChemistry, University
ofVienna,
Wàhringerstra17,
A-1090Vienna,
Austrialong
as there is a metric distance measure defined for them. Extensive studies oufolding
of RNA sequences intosecondary
structures have been based on this fact[5,
GI.Trie
landscapes
of a number of well known combinatorialoptimization problems
bave beenstudied,
such as trieTraveling
Salesman Problem[7],
theGraph Bipartitioning
Problem[8],
and theGraph Matching
Problem[9].
Detailed information on the distribution of localoptima
and the statistical characteristics of downhill walks have been obtained for the uncorrelatedlandscape
of the random energy model[10-12]. Furthermore,
twoone-parameter
families oftuuably rugged landscapes
have been studied in detail: the NÉ model aud its variauts[2, 5, 13, 14]
aud thep-spiu
mortels[15, 16].
Here we will be concerned with the eifect of
averaging procedures
related to runuing averagesou
laudscapes.
After some formol definitions(Sect. 2)
we discussbriefly
the Fourierdecompo-
sitiou of
laudscapes (Sect. 3)
which is used in section 4 to derive the basicproperties
of liuear operatorsapplied
tolandscapes.
In section 5 webriefly
address the relation of averages over ensembles oflaudscapes
aud averages overconfigurations
insingle
instances oflandscapes.
In section 6particular
operators,namely weighted
averages overone-step neighborhoods,
are usedto coustruct a uovel tuuable
family
oflaudscapes.
In section 7 the behavior ofp-spin
aud NÉlandscapes
under the iterated action of theseoperators
isiuvestigated.
2. Some definitions.
Let r be a
graph
with N < cc vertices. Agraph autoniorphism
a is a one-to-one mappmg of r onto itself such that the verticesa(~)
auda(y)
are connectedby
anedge
if andonly
if~ and y are conuected
by
anedge.
Two vertices of r are said to beequivaleut
if there is anautomorphism
o such that y=
a(~).
Agraph
is uerte~ transitive if ail vertices areequivalent.
A vertex transitive
graph
isalways regular,
1-e-, each vertex has the same number ofneighbors.
Trie distance
d(~,y)
of two vertices ~,y E r is defined as the minimum number ofedges
thatseparates
them. This distance is a metric. Agraph
is distance transitive if for any two pairs of vertices(z, y)
and (~1,u)
withd(~, y)
= d(~1,u)
there is anautomorphism
a such thata(~)
= u
and
o(y)
= u
(see,
e-g.,[17]).
Agraph
r is definedby
itsadjacency
matrizA,
withAxy
=1 if the vertices ~ and y are connectedby
anedge,
andAxy
= 0 otherwise.In the
following
we willonly
considergraphs
that are at least vertex transitive.Following
Bollobàs
[18]
we may thenpartition
the vertex setby
thefollowing procedure:
Choose anarbitrary
vertex as referencepoint,
which in thefollowing
will be labeled 0. Then constructpairwise disjoint
sets of verticesV(0)
=
(0), V(1), V(2),
..,
V(M)
such thatâ~~
:=£ Axy
isindependent
of y EV(v). (l)
x~V(p)
In other
words,
thispartitioning
is chosen such that each vertex y mV(v)
isadjacent
toâ~~
vertices in
V(p).
For someexaniples
seefigure
1 in section 3. It is mducedby
thesymmetry
of r: the vertices in each of the setsV(~1)
arepermutated by automorphisms
that leave 0invariant. In the worst case each set contains a
single
vertex. Ingeneral
we will usegreek
letters to index trie sets of such a
partition,
while roman indices are reserved for trie vertices off.For
highly symmetric graphs
triecollapsed adjacency
matriz=
(â~~)
cari be much smaller than trieadjaceucy
niatrix A. lu case of distance transitivegraphs,
for instance, trie distance classes(1.e.,
the sets of ail vertices with commou distance from the reference vertex0)
form apartition fulfilling equ.(l),
and thecollapsed adjacency
matrix comcides with the intersectionmatrix as defined
by Biggs [19] (see
also [20]).
We will use d instead ofgreek
letters formdexing
the distance classes.
The
following property
ofcollapsed adjacency
matrices will be used in thesubsequent
sec-tions.
Lemma 1.
£ (A~)xy
=
(Â~)~~
for ail ~ EV(p),
and ailnon-negative integers
m, and y~v(~)hence for each
polynomial
p holds£~~~~~~lp(A)]xy
=[p(Â)]~~, i-e-, jJ(A)
=p(Â).
Proof. We
proceed iuductively:
The assertion is trueby
definitiou for m = 1. For m= 0 it
holds
trivially.
Now suppose it is true for m 1; theu£ £ (A~~~)xzAzy
"
&<)~ [ (~[[
~~~ii~i~lll z(~mi)«~Àv<
-
(Â~~«~
~ ~~~~~, ~~v~~~ v
The
geueralizatiou
topolyuomials
isstraight
forward.~
Definition 2. A
landscape f
r - lit is a raudom field F ou the set of vertices of r definedby
the distribution fuuctionP(yi,
y2>UN)
" Prob
( f(~~) §
j,§1 § N) (2)
where N is the number of vertices of r.
Let
El-j
denote the mathematicalexpectation.
Forexample,
theexpected
value of theproduct
of the valuesf(~k)
andf(~i)
at the vertices ~k and ~i isEifl~k)fl~i)i
=
/vkvidPlvi>v2>. >vN). (3)
Definition 3. A
landscape
isisotropic
ifIi) El f(~)]
=f
for ailconfigurations
~ Er;
(ii)
for any twopairs
ofconfigurations (~, y)
and (~1,u)
such that there is agraph
automor-phism
a witho(~)
= ~land
a(y)
= v halas
£[f(~) f(y)]
=
El f(~1) f(v)].
The second condition nleans that any two
pairs
ofconfigurations
that areequivalent
inconfig-
uration space have the same
pair-correlation.
The covanauce matrix of F is definedby
Cxy
=Eif(~)f(v)i Elf(~)iElf(v)1 14)
By
vertextransitivity
there is anautomorphism
a such thata(~)
= 0.
Isotropy
henceimplies C~y
=Co,~(y)
=c(p),
whereV(~t)
is thepartition
to whicho(y) belongs.
The variance is givenby
Varjfj
=
Cxx
=Coo
#c(0) (5)
The a~ltocorrelation
f~lnction
is defiued asp(p)
=
c(~1)/c(0).
3. Fourier series on
landscapes.
Let C be the covanauce matrix of a raudom field F ou some
graph
r. Deuoteby (#~)
acomplete
set of orthonormaleigenvectors (which
existsby symmetry
ofC). Although (#~)
can be chosen
real,
we will admitcomplex
vectors as well. Then thelandscape f
on r cari berepresented
in mean square sense asf(~)
=£ a(v)4y(~) (6)
(The labeling
of theeigenvectors #y
with vertices isarbitrary.)
This seriesrepresentation
ta known as the Karh~lnen-Loèueexpansion.
Remark. The coefficients
a(y)
are uncorrelated. For limite sets the Karhunen-Loèveexpansion
coiucides with the well known
principal compouent aualysis
mtroducedby Hotelliug [21] (see,
e-g.,
[22]).
For au
arbitrary graph
r withadjaceucy
matrix A thegraph Laplacian
is definedby
/h=
D~~A E,
where D is thediagonal
matrix of vertexdegrees (see,
e-g.,[20])
For vertex transitivegraphs
this becomes /h=
(1/D)A E,
where D is now the constant vertexdegree
of r. Denote
by (Ù~)
acomplete
orthonormal set ofeigenvectors
of /h. A serresrepresentation
of the formf(~)
=£ b(v)Ùy(~) (7)
y~r
is called Fourier expansion of the
landscape
[2].Definition 4. Let
(G,o)
be a limite group, aud let 4l be a set ofgeuerators
of G such that the groupidentity
is flot contained in4l,
and for each ~ E 4l the inverse group element~~~
is also coutained in 4l.
(4l
is a set ofgenerators
meaus that each group element z E G can berepresented
as a truiteproduct
of elemeuts of4l,
withmultiplication
defiuedby
the groupoperatiou o,)
Letr(G, 4l)
be thegraph
with vertex set G aud auedge couuectiug
two vertices~ aud y if and
ouly
if~y~~
E4l,
1-e-, two vertices are couuected if andouly
there is a't E 4l such that y= ~/~.
r(G, 4l)
is called aCayley graph (see,
e-g., [23]).
Remark. A
Cayley graph
is vertex transitive.A commutative group
(G, o)
can be writteu as trie directproduct
of a finite uumber L ofcyclic
groupsCj
of orderNj.
A group element cau berepreseuted by
~ =(~i,
~2,~L)
with 0 < ~j <Nj;
theu the groupoperatiou
becomesz=~oy -
zj=~j+yjmodNj j=1,. ,L. (8)
It is easy to check that the functions ep, p E
G,
defiuedby ep(z) IL
= exP
2«1£ (~~ (9)
~=i j
fulfill
ep(~)eq(~)
=
epoq(x).
Furthermorethey
areorthogonal
withrespect
to the scalarproduct
£~~~ eq(~)eq(~)
=
ôpq.
Asimple
cousequence of these facts is trie well knowuLemma 5. Let
r(G,4l)
be aCayley graph
of the commutative groupG,
aud letHi
=h(j
o1-1) depeud ouly
ou the "diifereuce" of the group elemeuts1 audj.
Theu ep as defiuedin
equatiou (9)
is aueigeuvector
of H. Inparticular,
the set(ep)p
EG)
forms au orthouormal basis ofeigeuvectors
of theadjaceucy
matrix ofr(G, 4l).
Remark. Lovàsz
[24]
used this result to show that that theeigenvalues
of theadjacency
matrix of a
Cayley graph r(G, 4l)
with commutative group G aregiveu by
Àp=
£~~~ ep(k).
Weiuberger [14]
showedusiug
lemma 4 that Karhuueu-Loève expansion aud Fourier expan- sion coiucide forisotropic landscapes
ouCayley graphs arisiug
from comm~ltatiue groups.(He
used a
slightly
more restrictive defiuitiou ofisotropy, requinug
the same covanauce for ailpairs
ofconfigurations
with the samedistance.)
We will show that auaualogous
result holds for distance transitiveconfiguration
space.01o 0110
Â=
3 01Â-
21010 2 2
~0121
10 01Fig.
1. Petersen'sgraph
is distance transitive, but it is not aCayley graph (see,
e-g.,[23]).
It has three classes of vertices: the refereuce vertex(top),
itsueighbors,
aud ail other vertices. Its collapsed adjacency matrix is given below thegraph.
Thetrigonal
prismgraph
T is theCayley graph
of C2 x C3 with set of generators 4l=
((~,e), (e,~),(e,~~~)),
with C2#
le, ~)
andC3
=(e,~,~~~)
and edenoting
the groupidentity.
Anedge connecting
the twotriangles
isclearly
notequivalent
to anedge
within atriangle,
hence T is not distance transitive. Thereare 4 symmetry classes of vertices: the reference vertex
(top),
its twoneighbors
m the uppertriangle,
itsneighbor
in the lowertriangle,
and the remammg two vertices, which have distance 2 from the reference.Remark. There are
graphs
that are distance transitive but flotCayley graphs,
aud vice versa, there areCayley graphs
of commutative groups that are not distance transitive. The two mostsimple examples,
the Petersengraph
aud thetrigoual
pnsm are shown mfigure
1.Lemma 6. Let
a(~
= ifd(1, j)
=t,
and 0otherwise,
forail1, j
E r. Denote thecorresponding
matrixby A(~)
Let r be distance transitive. ThenA(~)
is a linear combinatiou of the first t powers of A.Proof. The assertion is
trivially
true for t= 0 aud t
=
1,
asA(°)
= E aud
A(~)
= A.
Now calculate A
A(~~~)
Thetriangle mequahty
assures thatonly
pairs of indices(1, j)
withdistances
t,
t-1,
aud t 2 can have non-zero entries. Distancetransitivity
of thegraph guarantees
that ail entriesbelonging
to the same distance Mass areequal,
and heuceA
A(£-1)
= c ~
A(£)
+ c~A(£-1)
~~~
~(t-2) (~~)
where c+ > 0 as
long
as t ares flot exceed the diameter of r.~
Theorem 7. Let
f
be auisotropic laudscape
ou aCayley graph r(G, 4l)
with a commutative groupG,
or on a distance transitivegraph
r. Then trieadjacency
matrix A of r and thecovariance matrix C of the
landscape f
commute, AC= CA.
Proof.
(a)
ForCayley graphs
of commutative groups theproposition
is a direct cousequeuce of leninia5,
see[14].
(b)
Iff
isisotropic
aud r is distancetransitive,
theuCxy depeuds ouly
ou the distance of the verticesd(~, y).
Heuce C cau be written as a hnear conibiuatiou of the matricesA(~),
andby
lemnia
6,
it cari be writteu as apolyuoniial
in terms ofA,
C=
p(A).
In both cases C aud A commute because
they
bave trie sameeigeuvectors. ~
4. Linear
operators
onlandscapes.
Consider
landscape
g obtainedby
someweighted averaging procedure
from alandscape f,1
e.,g(~)
=£~~r flxy f(y).
In thefollowing
we will use the obvious matrix notations g =Qf.
Lemma 8. Let C be the covariance matrix of the
landscape f,
and let fl be a hnearsymmetric operator.
Then the covariance matrix of g = flf
is givenby
C~
= flcfl
(11)
Proof. We assume without
loosing generality
thatEl fp]
= 0 for ail p E r.
c)~
=Ej(n1)~(n1)~j
= E
£ n~~ i~ £ n~~ il
p q
=
L "xpElfpfqlflqy
=(flcfl)xy
p,q
Corollary
9. Letf
beisotropic
and assume that theconfiguration
space r is distance transi- tive or theCayley graph
of a commutative group.Suppose
fl is apolynoniial
of theadjacency
matrix. Then we have
C~
=
(fl~)C.
Proof. From theorem 7 we know that C and the
adjacency
niatrix A of r comniute. ThusC comniutes with ail powers of A and
consequeutly
with ailpolynomials
of A. Thecorollary
follows now
imniediately
from lemma 8.~
Theorem la. Let
f
beisotropic
with autocorrelatiou functiou p aud let r be distance tran- sitive(or
aCayley graph
of a comniutativegroup).
Furthermore let fl be somepolynomial
inA,
Q=
P(A).
Then g= Q
f
is agamisotropic
on r andp~
= c
ù~p (12)
where the constant c =
1/(ù~p)(0)
is chosen such thatp~(0)
=1.Proof. From
corollary
9 we kuowCg
=
(fl~C)ox.
WithQ
"
fl~
we may wnte this as~~ ~Î0 ~ ~ QxyCy0 ~ ÔÀpCp0
P y~V(p) P
for ail ~ E
V(À),
as a cousequence of lemma 1.Dividing by
the variaucecompletes
theproof.~
There exists a set of
nght eigenvectors
of that areorthogonal
withrespect
to theweight
function
w(~t)
=
)V(~t)) (for
details see[25]).
We may thereforeexpand
the autocorrelation function withrespect
to thisbasis, p(~t)
=£~ a~p~(~t).
Corollary
ii. Uuder theassumptions
of theorem 10 we havep~(~t)
= c£ a~P~(À~)p~(~t) (13)
where À~ denotes the
eigenvalue
ofbelonging
to p~. Inparticular,
if p= pj for some
j,
thenwe have
p~
= p.
5.
Empirical landscapes
and ensembles ofIandscapes.
The
empirical
vanancea~
of anisotropic landscape f,
1-e-, of aparticular
instance of the random field F is measuredby picking
asample
of values atconfigurations randomly
distributed on r.Denote
by p(~t)
=)V(~t)) IN
theprobability
that tworandomly
chosen vertices ~ and ybelong
to
partition
~t, 1-e-, that there is anautomorphism
a witha(~)
= 0 anda(y)
EV(~t).
In caseof a distance transitive
configuration
space we canmterpret p(d)
as theprobability
topick
at random twoconfigurations
with distance d. Theempirical
vanance cari be written as [5]a~
=
jr (f(~) f(Y))~ (14)
~,y
and its
expected
value cari be calculated:Eia21 =j ~l (2Eii(~)21 2Eii(~)i(y)1)
=varifi
LP(JL)jv)~~j L Eif(°)f(Y)i
=
varifi Ii P(~1)P(~1)j
~~~~~ y~v~~> ~
=varlfl(i à)
Hence the
expected enipirical
variance of alandscape f
is related to the ensemblevariance,
i e., the vanance of the randoni field F at a
particular vertex, by
E[a~]
=Var[f](1 fl) (16)
The
parameter
p nieasures the average correlation of thelandscape.
As shown m[25]
fl is the contribution to pcorresponding
to thelargest eigenvalue
ofA,
Ào " D.The ensenible-autocorrelation function p and the
expectation
of theenipincal
correlation function fi are hence relatedby
P(11) =
~l~ j~ (17)
Equation (17) iniplies
that theoninipresent 'back-ground'
correlation p has to be subtracted.We reniark that for
practical
purposes it is theempincally
observable correlation function fi(~1) that niatters. The autocorrelation function is said to beself-averaging
if fi= p
(see,
e-g.,[26]).
Corollary
12. Under theassumptions
of theorem 10 it follows from p= fi that
p~
=
fi~,
1-e-,if the autocorrelation function of
f
isself-averaging,
then the autocorrelation function of thelandscape
Qf
is alsoself-averaging.
6. Iterated
smoothing
maps A tunablefamily.
By N(z)
we will denote the set ofneighbors
of ~plus
~itself,
i-e-, ail vertices y E r withd(~,y)
< 1. We derme thesmoothing
operator itby
it =
~
(A
+E) (18)
The snioothed
landscape
is then definedby
g =itf.
Let r be distance transitive. Then we obtain as an immediate consequence of theoreni 10f a~(à)p(à+k) p~(à)
=
k=-j (ig)
La~(o)p(k)
~mo
where ah
(d)
=
(Î~)d,d+k.
The coefficients have asiniple
combiuatorialiuterpretatiou:
ah(d)
is the uumber of pairs (~1,
v)
with u EN(~),
v EN(y)
with d(~1,v)
= d + k where
d(~,y)
= d is
fixed.
For
particular graphs
r it isfairly
easy to calculate the coefficients ah(d) explicitly.
Cousidera
geueral hypercube
of diameter n over aualphabet
of size ~. Theconfiguration
space has N= ~" vertices.
By
distancetrausitivity
we may choose~
=(000...0000000....0000)
y
=(111.. 1111000....0000)
~~~~with
d(~, y)
= d. Wepartition
the set of pairslu, v)
with ~1 EN(~),
v EN(y)
iuto urne subsetsdepending
on whether u= ~, u is
neighbor
of ~ with the diiferenceoccurring
m one of the first dpositions,
or u is aneighbor
with the diiference occurnng in the latter n dpositions,
andthe same for v. If u
= ~ aud v
= y then
d(u,v)
= d. If u diifers form ~ in oue of the latter
n d
positions
and v= y then d(~1,
v)
= d
+1;
there are(n d)(~c -1)
suchpairs.
If ~t differs from ~ in the firstsection, however,
there are d(~c2)
pairs with distance d aud dpairs
with distance d1,
the latteransiug
frommutatiug
a 0 to a 1. The combinatious of mutations mboth z and y cau be treated
similarly. Collecting
ail termsfiually yields
a~~(à)
=
à(à i)
a-i(d)
= 2d + (~c
2)(2d 1)d
Où
(d)
= 1 + (lG1)(n
+2d(n d))
+ (lG2)(lGd
+d(d
1)(lG2)) (21)
a+i
(d)
=
2(~ l)(n d)
+(~ 2)(~ l)(n d)(2d
+1) a+2(d)
= (lG
1)~(n d)(n
d1)
Let
fo
be the random euergy mortel[27, 28].
The autocorrelation function isp(d)
=
ôo,à
andfi(0)
=1, fi(d)
=-1/(N -1)
for d >0, respectively,
where N denotes the number ofpoints
m the
configuration
space. Definefr
= itfr-i
Let fi(~) deuote the autocorrelation fuuctiou offr.
We will refer to thisfamily
oflaudscapes
as iteratedsmoothing landscapes,
ISLS.Numerical evaluatiou of
equation (19)
with the initial condition fi ou BooleanHypercubes
and their ~c-letter
generalizatious
shows that(fr)
iudeed forms afamily
oftuuably rugged landscapes (Fig. 2).
Inparticular,
we haveTheorem 13. Let r be a
generalized hypercube
over analphabet
with ~c > 2 letters. Thenlim
fi(~'(d)
= 1-
Îd (22)
Ou a Boolean
Hypercube
(~c=
2)
we havelim
fi(~)(d)
=
~(l ~~)
+(-1)~ (23)
Proof. The
eigenvalues
of theadjaceucy
matrix of thegeueral hypercubes
are[29,30]
ÀJ "
n(~t 1)
~cj(24)
aud heuce the
eigeuvalues
of thesmoothiug operator
it are 1j
~ forj
=
0,
,
n(~c -1)
+ n.The
largest eigeuvalue
ofit~, apart
from1,
which is obtaiued fromj
=0,
is fouud besettiug j
= 1 for ~c > 2. In the case ~c= 2 the second
largest eigenvalue
ofit~
isdegenerate,
thecorresponding eigenspace
isspanned by
pi and p~. As an immediate cousequence ofequation (28)
in[25],
the autocorrelation function of the random energy mortel con be written as>(°) (à)
= c
f(~ i)£ (])
p~(à) (25)
tmi
By corollary ii,
the iterates of fi will converge to a vector tu the eigenspace of thelargest eigenvalue
ofÎ~
that is notorthogonal
to the initial condition. Theeigeuspace correspoudiug
to Ào has beeu omitted
by
construction offi. By equatiou (25) p(°),
is flotorthogonal
to auy of theeigenvectors
pithrough
p~. For ~ > 2hence,
p converges to pi while for ~c= 2 we obtaiu
a mixture of pi aud p~. The ratio of their contributions does flot
change
uuder trie action offlÎ as a consequence of
coroiiary
II..
o
0.8
îl.
~ ~,
>i ~
O.G ~ ',
i1,
"~'i
~
'~
0,4 ~ ',
', i '~
i
1,
"~,°.~
",, ',,~ ",_~
~
ooÎ""-=.-m-ÎÎÎ-=~Î~
~ --_
0.2 '~~~"--
0.4
0.6
0.8
0
0 10 20 30 40 50 60 70 80 90 100
Hamming
Distance da)
0
;,
°'8 ' ',
' ',
' ',
0.6 1 ~,
[
i '~', ' '~~
°'4
', "~
' , ',
' , ',
~ '
0'2
' '~ "~
" ~~ '~,
~~ ~,
~o ',_ ~-__ ~
6 O.O '~~'"'~~~~~~
~~ '- "~-,
-0.2
-0.4
0.6
o.8
1.o
0 10 20 30 40 50 60 70 80 90 100
Hamming
Distance db)
Fig.
2. Autocorrelation function of ISLS ongenerahzed hypercubes
witha)
<= 2, and
b)
<= 4
for with n
= 100. The number of
smoothings
area)
r = 25(dotted),
r= 50
(short dashed),
r = 75(long dashed),
r = 100(dot,dashed),
r= 200
(sohd),
andb)
r = 50(dotted),
r = 100
(short dashed),
r = 150
(long dashed),
r= 200
(dot,dashed)~
and r= 400
(sohd), respectively.
We remark that trie contribution of p~ decreases with the size of the
configuration
space.The
hmiting
autocorrelation function(22) corresponds
to both the trivialspin glass
mortel with Hamiltonian7i(a)
=
£~ Ajaj,
and to the Nk,mortel with k= 0. We also note that for
r
In
< 1the
landscapes
areessentially uncorrelated,
1-e-, the correlationlength
iso(n).
Numerical estimates show that the nearestneighbor
correlation fi(~)(1) depends only
on the ration rIn
forlarge
n(See Fig. 3).
As a consequence of theorem13,
we find thatlimr-m pli
= 1
~ ~
n ~c 1
for ~c > 2. For ~c
= 2 we find
lim
p(~)(d)
= lim
p(~)(d 1)
= 1-
~~
for even d
(26)
r-ce r-ce 7Î +1
The data in
figure
3 indicate that this limit isalready
obtained forr/n
> 2 to agood approximation.
7.
Averaging operators
onspin glass
and Nklandscapes.
We first consider the p-spm
mortels,
1-e-, thelong
rangespin glasses
with Hamiltonian7ip(a)
=£
Aj~j~__,j~ aj~aj~ aj~
(27)
31<32<...<3P
with a~ E
(-1, +1)
andcoupling
constants A~j___k drawn from some common distribution. Theruggedness
of thelandscapes
increases as the order p of thespin-interactions
increases. We have thenTheorem 14. Let fl be any
polynomial
of theadjacency
matrix of the BooleanHypercube.
Then the
landscapes 7ip
andQ7ip
have the same autocorrelation functions.Proof. As shown in
[16]
the autocorrelation functions of thep-spin
mortels arePV
id)
=
(il
~(ii)~ Ill Ii Il (28)
These are
smtably
normalized Krawtchoukpolynomials,
which are known to be theeigenvectors
of thecollapsed adjacency
matrix of the BooleanHypercube (see,
e-g.,[25, 29]). Application
of
Corollary
iicompletes
theproof. ~
In Kauifman's Nk-mortel the fitness of a
binary string
~ is defined as the average over"Site fitnesses"
f~,
whichdepend
on ~~ and k additionalpositions ii through ik.
The site fitnesses f~ =f~(~~,
~~~,..,
~~~)
are assumed to be uncorrelated random variables drawn froma common distribution
iii.
Theruggedness
of thelandscape
mcreases as the number k ofpositions contributing
to asurgie
site fitness mcreases. Diiferent variants of theselandscapes
have been
investigated, depending
on whichpositions
contribute to a given site fitness(see,
e-g.,
[31]).
In contrast to the p-spm
mortels,
thelandscapes
of NÉtype
are aifectedby
thesmoothing
operator
it.Theorem 15. Let
f
be anarbitrary
NÉlandscape.
Then the autocorrelation function of thelandscapes it~f
converges top(d)
=
((1- 2d/n)
+(1- ()(-1)~
with o <(
< 1 as r tends towardsinfinity.
o
0.8
0.6
0.4
n=25 n=50 n=100 n=200 0.2
ù-ù
ù-ù 0.5 1-Ù 5 2.0
r/n
a)
i o
o.8
0.6
0.4
n=25 n=50 n=100 n=200
o-ù
o-o 0.5 1-o 1.5 2.0
r/n
b)
Fig.
3. Nearestneighbor
correlation for ISLS ongeneralized hypercubes
as a function of the number of iteratedsmoothings. a)
~= 2,
b)
«= 4.
Proof. It is necessary and sufficient to show that the autocorrelation function p of
f
and piare not
orthogonal,
1-e-, that£ p(d)pi (d) (")
=( #
0(29)
"
Î~
d
Note that pi
(d)
= -pi(n d). Equation (29)
cari hence be rewritten asf
=£ (()
pi(d) lp(d) p(n
d)1(30)
d<n/2
If
p(d)
>p(n d)
for ail d <n/2,
and if theinequality
holdsstrictly
for at least one d <n/2,
we have
(
> 0. It followsimmediately
from the very definition of the NÉ mortel(irrespective
of how theneighboring
sites arechosen)
that p ismonotonically decreasing
and non-constant-'
Remark. The vanants of the NÉ mortel do trot behave
identically
under the action of it.Whenever the autocorrelation function is a
polynomial
ofdegree
less than n, then we have(
= 1 in theorem 15. For exactexpressions
of the autocorrelation functions of varions NÉniodels see
[2, 5]. Recently
Nk-mortels on sequences ofarbitrary alphabets
have been considered[2].
Theorem 15 holds with(
= for ailalphabets
with three or more letters.As a final
exaniple
consider anisotropic landscape
on the circle oflength
n > 4(Millonas,
pers.
comniunication).
Theconfiguration
space hence is theCayley graph
of thecyclic
groupC~
=l't~)0
< k <n)
with set ofgenerators
4l =16't,'t~~).
Its vertexdegree
is D= 2. The
eigenvectors
of theadjacency
niatrix are eh(~)
=exp(2grk~ In),
as an immediate consequence of lemma 5. Theeigenvectors
of thecollapsed adjacency
matrix arepk(d)
=
cos(2grkd/n)
for 0 < k <
n/2,
associated with theeigenvalues
Àk #2cos(2grk/n).
Theeigenvalue
ofÎ~
corresponding
toÀ~/2
is niuch smaller that the onecorresponding
toÀi
for n >4,
hence p converges to pi under the iterated action of it whenever p is notorthogonal
to piSuppose f
has a monotonenon-increasing
autocorrelation function p. As aparticular
ex-ample
consider the random energy niodel on the circle.By
the same reasoning as in theproof
of theorem15,
p is notorthogonal
to pi, and hence iteratedsmoothing eventually
leads to alandscape
with autocorrelation function pi Note that the smoothedlandscapes
on the BooleanHypercube (and
on relatedgraphs) eventually
have autocorrelation functions of the formp(d)
= 1-
)d
+O(d~)
while for the circlegraphs
we findp(d)
= 1
ad~
+O(d3)
Thelandscape
on the circle hence becomesexceptionally
smooth [16] as a consequence of thespecial geometry
of theconfiguration
space.This
example
also shows that thedegeneracy
of it on the BooleanHypercube
is trot anecessary consequence of the fact that the
configuration
space isbipartite,
1-e-, that there isan
eigenvalue
-D of theadjacency matrix,
since trie circlegraphs
arebipartite
for even n, and it is notdegenerate
there.8 Discussion.
A
general theory
of the action of linear(averagmg)
operators onisotropic landscapes
is devel-oped
which allows topredict
their eifect on the correlation structure of thelandscape.
We have shown that there areparticular Iandscapes (characterized by
autocorrelation functions that areeigenvectors
of thecollapsed adjacency
matrix of theunderlying configuration space)
which arenot aifected
by "smoothing" operators.
Thep-spin models,
and inparticular
theSherrington- Kirkpatrick [32],
model are members of this dass on the Booleanhypercube.
Anotherexample
is the
landscape
of thegraph-bipartitioning problem.
A new
family
oftunably rugged landscapes
isestablished,
which arisesby
iteratedsmoothing
of the random
landscape.
These "iteratedsmoothing landscapes" (ISLS)
have anadvantage
over the other two known
tunably rugged
families: Thep-spin
mortels are definedonly
for Booleanhypercubes,
and I<auifman's Nkfamily
lives ongeneral
sequence spaces, while ISLS can be constructed onarbitrary configuration
spaces. Theirsimple
structuresuggests
furthermorethat
analytical
solutions of these mortelsmight
be feasible. Iteratedsmoothing
of the random energy morteleventually
leads to alandscape
withlinearly decreasing
autocorrelation functionon
general
sequence spaces withalphabets
of size ~c >2,
while on Booleanhypercubes,
~c =2,
one obtains a
superposition
of a linear function and an oszillation of the form(-1)d.
On a circlegraph
trie sameprocedure
leads to an autocorrelation function of trie form 1-ad~
+.This shows
again
that trie structure of alandscape
cannot beseparated
from thegeometry
of theunderlying configuration
space.Evolutionary adaptation
cari be descnbed in thesimplest
case as the motion of apopulation
in
configuration
space. Almost ailanalytical work, however,
has been doue withadaptive
walks
(or
relatedprocesses)
whichneglect
thepopulation aspect.
For a recent review see, e-g-,[31].
TO a firstapproximation
apopulation
cari bereplaced by
its "center of mass"(consensus configuration),
which moves in thelandscape "experienced" by
the entirepopulation: again
to a first
approximation,
thislandscape
is aweighted
average of the fitnesses ofconfigurations
centered around the consensus.
Consequently,
the width of thepopulation
shouldstrongly
influence thedynamics
of evolution: broaderpopulations
"feel" ingeneral
smootherlandscapes,
and henceoptimization
should be easier.Usually
the broadness of thepopulation
is controlledvia the mutation rate
[3,33]
which at the same time controls theexploration
rate. Furtherinvestigations
are necessary in order to separate the eifects of enhancedexploration
andimplicit smoothing
of theunderlying landscape
causedby
an increased mutation rate.It should be
kept
in mina that the resultspresented
heredepend
on theassumption
ofisotropic landscapes.
While this requirement is often fulfilled forsimple
statisticalmortels,
it isunlikely
to holà for realbiophysical
orphysical systems.
It bas beenshown,
forinstance,
that trie
landscapes
of RNA freeenergies
are notisotropic
for natural sequences[34].
Much work on random
landscapes
such as Derrida's REM bas been motivated with thenotion that a
sufficiently
"coarsegrained"
version of anyrugged landscape
would look random.More
precisely,
this is to say that it should bepossible
to construct renormalization groupshaving
randomlandscapes
as attractive fixedpoints.
While theaveraging procedures
discussedm trie this contribution are trot
exactly
trie kind of coarsegraining
one would have in mmd for such aconstruction,
the methodsdeveloped
here seem to be useful for such a task.The
theory
outlined in this contribution is trotapplicable
for thetraveling
salesmanproblem
since the raturai
configuration
spaces of the TSP areCayley graphs
of thesymmetric
group with the set of ailtransposition,
or the set of ail inversions(20pt-moves),
asgenerators
[7].It is easy to check that these
graphs
are trot distancetransitive,
and of course thesymmetric
group is non-commutative-Acknowledgements.
thank the Max Planck
Society, Germany,
forsupporting
my stay at the Santa Fe Institute.Stimulating
discussions with P-W-Anderson,
W.Fontana,
S-A-Kauifman,
M.Millonas,
R.Palmer,
and M. Virasoro aregratefully acknowledged.
References
iii
Kaulfman S-A- and Levin. S., Towards a GeneralTheory
ofAdaptive
Walks onRugged
Land-scapes, J. Theor. Biot, 128
(1987)
11-45.[2]
Weinberger
E-D-, LocalProperties
of the Nk,Model, aTunably Rugged Landscape, Phys.
Reu.A 44
(1991a)
6399-6813.[3]
Eigen M.,
Mccaskill J-S- and SchusterP.,
The MolecularQuasispecies,
Adu.Phys.
Chem. 75(1989)
149-263.[4]
Weinberger
E-D. Correlated and Uncorrelated FitnessLandscapes
and How to Tell theDilference,
Biot.Cybern.
63(1990)
325,336.[Si Fontana W, Stadler P-F-,
Bornberg,Bauer E-G-,
Griesmacher T., Hofacker I.L., TackerM.,
Tara-zona
P., Weinberger
E-D- and SchusterP.,
RNAFolding
andCombinatory Maps, Phys.
Reu.E 47
(1993)
2083-2099.[6] Fontana W.,
Konings
D. A.M-, StadlerP-F-,
Schuster P., Statistics of RNASecondary Structures, Biopolymers
33(1993)
1389,1404.[7] Stadler P-F- and Schnabl W., The
Landscape
of theTraveling
SalesmanProblem, Phys.
Lent. A 161(1992)
337-344.[8] Stadler P-F- and
Happel R.,
Correlation Structure of theGraph Bipartitioning Problem,
J.Phys.
A: Math. Gen. 25
(1992)
3103-3110.[9] Stadler P-F-, Correlation in
Landscapes
of CombinatorialOptimization Problems, Europhys.
Lent. 20
(1992)
479-482.[loi
Macken C.A.,Hagan
P-S- and PerelsonA-S-, Evolutionary
Walks onRugged Landscapes,
SIAMJ. Appt. Math. 51
(1991)
799-827.[iii Flyvbjerg
H. andLautrup B.,
Evolution in arugged
fitnesslandscape, Phys.
Reu.Af15j
46(1992)
6714-6723.[12] Bak
P., Flyvbjerg
H, andLautrup B.,
Coevolution m arugged
fitnesslandscape, Phys.
Reu.Af15j
46(1992)
6724-6730.[13] Kaulfman S-A-,
Weinberger
E-D- and PerelsonS-A-,
Maturation of the ImmuneResponse
ViaAdaptive
Walks onAflinity Landscapes,
TheoreticalImmunology,
Part I, Alan Perelson Ed.(Addison-Wesley,
Redwood City~1983).
[14]
Weinberger
E-D-, Fourier andTaylor
Series onLandscapes,
Biol.Cybern.
65(1991b)
321-330.[15] Amitrano C.~ Peliti L, and Saber M.,
Population Dynamics
inSpin
Glass Model of ChemicalEvolution,
J. Mol. Euol. 29(1989)
513-525.[16]
Weinberger
E-D- and Stadler P-F-,Why
Some FitnessLandscapes
are Fractal, J. Theor. Biot.163
(1993)
255-275.[17]
Buckley
F. andHarary
F., Distance inGraphs (Addison-Wesley, Reading,
MA,1990).
[18] Bollobàs
B., Graph Theory
AnIntroductory
Course(Springer-Verlag,
New York~1979).
[19]
Biggs
N.,Algebraic Graph Theory (Cambridge University
Press~Cambridge~ 1974).
[20] Cvetkovié D.M.~ Doob M. and Sachs
H., Spectra
ofGraphs Theory
andApplications (Academic Press~
New York~1980).
[21]
Hotelling
H.,Analysis
of acomplex
of statistical variables intoprincipal
components~ J. Educ.Psych.
24(1933)
417-441 and 498-520[22] Rao C.R., Linear Statistical Interference and Its
Applications.,
2nd Ed.(Wiley,
New York,1973).
[23]
Biggs
N-L- and White A.T., PermutationGroups
and Combinatorial Structures(Cambridge
Univ.
Press, Cambridge, 1979).
[24] Lovàsz L.,
Spectra
ofgraphs
with transitive groups, Pemodica Math.Hung.
6(1975)
191-195.[25] Stadler P-F-, Random Walks and
Orthogonal
Functions Associated withHighly Symmetric Graphs~
prepnnt 93-06-31(1993),
Disc. Math.(1994)
m press.[26] Binder K. and Young A.P.
Spin
Glasses:Experimental
Facts~ TheoreticalConcepts~
and OpenQuestions.
Reu. Mod.Phys.
58(1986)
801-976.[27] Derrida B. Random
Energy
Model: An exactly solvable model of disordered systems~Phys.
Reu.B 24
(1981)
2613-2626.[28] Gardner E. and Derrida B.~ The
probability
distribution of the partition function of the random energy model, J. Phys. A 22(1989)
1975-1982.[29] Dunkl C.F.,
Orthogonal
Functions on Some PermutationGroups.
D.K.Ray-Chaudhun Ed., Proceeding
ofSymposJs
in Pure Mathematic 34(Amencan
MathematicalSociety,
New York~1979).
[30]
Rumschitzky D., Spectral Properties
ofEigen's
evolutionmatrices,
J. Math. Biot. 24(1987)
667-680.
[31] Kaulfman
S-A-,
TheOrigins
of Order(Oxford University
Press, NewYork, Oxford, 1993).
[32]
Sherrington
D, and S.Kirkpatrick,
Solvable Model of aSpin Glass, Phys.
Reu. Lent. 35(1975)
402-416.
[33]
Eigen
M.,Selforganization
of matter and the evolution ofbiological
macromolecules, Naturwis-senscha/ten
58(1971)
465-523.[34] Stadler P-F- and W. Grüner,
Anisotropy
m FitnessLandscapes,
J. Theor. Biot. 165(1993)
373-388.