Markov Chains and Polynomial time Algorithms
Ravi Kannan*
Computer Science Department
Camegio-Mellon U1
versity
Pittsburgh, PA, 15213
Abstract
“This paper cuttings the use of rapidly mining
Markov Chains in randomised polynomial fine algo
fithmg to solve apprasimately certain counting prob
fel fl nt to cass = combinatorial prob
ls tke counting the number of perfect matchings in
cof gris fad grometric ony ie ecenputing the
‘olumes.of conven ss,
1 Introduction
We consider the problems of determining apprea
imately'the number of elements in certain sels, fox
ararple, the number ef perfect matchangsan a graphy
the nimber of eomnectod subgraphs of a given sraphy
te lather of lattice posts ana ven pelytope, of
the munher of m>
tnial time rsucite to another prcblem which we may
fall the random) sampling problem: oufput an
clament of the desea st gam s hat the pro
lily cf any particular elesnent beng outpu Bape
Draaiinately I Bver the carinality of the set [For the
hecio rest, see dexrurn, Valiant and Vawirani [51]
his ets is ars) urn samplings A ote
general (weighted) versen reqtines the probability
GFE Tengu tobe appreafaty pegerionl
FG) wae FO) 9 pot value funetion on the
fej we may call this Sampling acconding ta. In
the paper, we ako conser the weighted version,
‘Our purpose here ito desembe the moent sucess in
developing polynomial time algorithms for some sam
ing ‘problems. ‘sing the felling general heme (for
Smupheity,consider the uniform sup hngcaee): devise
a Maroy Chain wlioee states are the elements ofthe
fet and where steady state probabilities are the same
forall sas [his afin the ey pat Then soy
thal after a polynomial number of steps, the pretabib
ites ame “eooe” fo Ube steady stale probabilities [is
is Ge hard part), so that rutming the Marken Cheam
fbx this map stefs soles the sampling problem.
Let us Mlstrate vith a toy anne suppose we
ih (o sample fem the se ates points ian =
Himensjonal cube ead of whose sides goes from 1 io
arash tee dG lav ogee the obs
thetiod): Heres a Markov Chat whose stay sae
Panay sapped ty NSF-Grant OCROINDT
probabilities are (easily shown to be) 1d for each
Halotiemn He Goneat ities pont (ithe cabo
tliocse one ef the Sn coordinate dictions, each seth
pots entangle pt gle
por in that nection iff isin the cube, otherwise,
stay at the cument point. [This ceseribes one move
ofthe dn] Heal Le stn th ve ste the
Marlow Chain at any state after O(n logd) sexs,
Me pgs df each state is beter 1 via and
!YCP). This & thns an ewample of a rap moe
iy Matic Chain? where the numdsr of steps nose
so that the probabilities are close to the ste state
is small (Pelynomial in nae) wheres the number of
Bates iP
“The profes for which this approach has yielded
polynemial time algorithms, fall mio to. main eate-
dks Peer eae ec ft a
Ennples or geotnetne probes like our last oro cent
ples [The algorithms age in general fully polyacng
Tandotniged appraimation schemes (fpras) which take
ko ys ceed pet Veal en the dave
ror; in this paper, we do not rpention tise anymore
toroid olsctrtg the main points
‘The tate of convergence of general Markov Chins
vee tate f convergence of general Markey Chains
theca hy Sock a Sarat [ad Mon [7
Broker [19] introduced the Markov Chain appacach
to combinatorial algoituns. demu and Sinai [8]
Pioneral ty algertinie approach Ly shew lever
founcs en the ecetictance of een Manin Chains,
leading to 2 proof of fast comergence 10 the stay
Sate.
Inthe geometric setting, the conductance is relate
to a natural gecrnetric elation calla) boperimetry.
"The proof of ioperinciie inequalities farms the oa
tral part of the proof of fat ecayvergence of sect
Marko chains. ‘The initial motivation came from the
clissical problem of comping the volume cf con
si fe well civ wl to or than au
ple above), but te approach has since been applied to
Ether problems lie tultveriae statistical epling,
gs nels etain'Serorsie Option ol
“Techniques other than the Markov cans appncach
have also been weal to solve counting probes effe
cently, Ceeed form formulas as'one encounters in
Combinatorics the matritnee Theorem of Kirchott
that-ccunts among others the number of toes in
raph, the algorithm of Rarp an Lay to count the