Slide 1 DIMACS 20 th Birthday Celebration, 20 November 2009 The Combinatori al Side of Statistical Physics Peter Winkler, Dartmouth
Dec 19, 2015
Slide 1
DIMACS 20th Birthday Celebration, 20 November 2009
The Combinatori
al Side of Statistical Physics
Peter Winkler, Dartmouth
Slide 2
Phase I: The Parties Meet
Slide 3
We begin with a checkerboard on which checkers
are placed uniformly at
random subject to the condition that
no two are orthogonally
adjacent.
Combinatorics---
or Statistical Physics?
Slide 4
To see better what’s going on, we color the even occupied squares blue and the odd ones red.
Notice the tendency to cluster…
Discrete hard-core
Slide 5
Actually, that was just one
corner of this picture
(generated by me and Peter
Shor using “coupling from
the past.”)
The big picture
Now, let’s raise the stakes by
rewarding larger independent sets with a
factor for each extra occupied
site.
Slide 6
The plot thickens
This is what it looks like
when we set = 3.787.
Slide 7
Take-over
At = 3.792, one of the
colors “breaks
symmetry” and takes over the picture.
We suddenly get “ordered
phase,” “long-range
correlation” and “slow mixing.”
Slide 8
Statistical physics Combinatorics
hard-core modelrandom independent
sets
monomer-dimer random matchings
branched polymers random lattice trees
Potts model random colorings
linear polymers self-avoiding random walks
percolation random subgraphs
Slide 9
CS theory’s favorite hard-constraint model
Physics techniques (e.g., “cavity method”) have helped to make major progress in
understandingsatisfiability. [Mezard, Parisi and Virasoro ’85]
(x y z)
(x y u)
(x z v)
(w z t)
(y z w)
x
y
z
w
Slide 10
Which graphs cause a phase transition?
[Brightwell & W. ’99]
On the Bethe lattice, where
things are nice:
Slide 11
Phase II: DIMACS Makes a Match
Slide 12
In 3 dimensions, the “critical activity” for the discrete hard-core model drops from about 3.8 to about 2.2. What happens when the dimension gets very high?
Combinatorialists settle a controversy
Even a certain well-known married couple at Microsoft Research couldn’t
agree.
Along came [David Galvin and Jeff Kahn ’04] (with ideas from Sapozhenko) to show the critical activity goes to zero.
Slide 13
Phase III: MSR Leads the Charge
Slide 14
Take your favorite graph G, and let its vertices (or its edges) live or die at random. What
happens?
Percolation
For example: edges of a large empty graph arecreated independently with probability p. When
do you get a giant connected component?
Physicists call this game percolation and usuallyplay it on a grid, asking: when is there an infinite
connected component?The physicists’ scaling methods are quite powerful. E.g., [Borgs, Chayes, Kesten &
Spencer ’01]find the scaling window and critical exponent
for the Erdos-Renyi giant component.
Slide 15
Colour the points of a Poisson process
green (with probability p ) or
red.
Voronoi percolation
Now draw in the Voronoi cells; do the green cells
percolate?
[Bollobas and Riordan ’07] proved that the critical probability is ½. Read their new book on
percolation!
Slide 16
Coordinate Percolation
independent percolation:
a vertex lives or dies based on an independent event
associated with the vertex.
?
coordinate percolation:
a vertex lives or dies based on independent events
associated with the vertex’s coordinates.
?
?
Motivation: water seeping through a porous
material.Motivation: scheduling!
Slide 17
This type of dependent
percolation came up in the study of a self-
stabilizing token management
protocol.
Coordinate Percolation
Here, each row and each column has been randomly
assigned a number from {1,2,3,4}.
A site is killed if it gets the same
number from both coordinates.
Slide 18
An easy variant of coordinate percolation
Random reals (say, uniform in [0,1]) are
assigned to the coordinates;
each grid point inherits the sum of its coordinate
reals;
and any vertex whose sum exceeds some
threshold t is deleted. (t=.75 in figure.)
0.1 .4 .30.1
.4
.1
.7
.3
.1 .4 .1.5.2
.5 .8 .4 .7 .4
.2 .5 .1 .4 .1
.8 1.1 .7 1.0 .7
.4 .7 .3 .6 .3
t
Let be the probability of escape from (0,0) when the
threshold is t.
tHmmm… does it matter if we’re allowed to move left or down, as
well?
Slide 19
Theta-functions, Independent vs. Coordinate Percolation
independent
p
1
1p
pc
Unknown: behavior of theta just above the
critical point, e.g.: what is the critical exponent?
Known: a precise closed expression for
the probability of percolation!
2
t
1
1 2tcoordinate
Slide 20
Back to the continuum?
Sometimes, paradoxically, you get better combinatoricsby not moving to the grid. Example: branched polymers.
Physicists have been studying these on the grid. But…
Slide 21
Branched polymers in dimensions 2 and 3
[Brydges and Imbrie ’03], using equivariant cohomology,proved a deep connection between branched polymers in dimension D+2 and the hard-core model in dimension D.
They get an exact formula for the volume of the spaceof branched polymers in dimensions 2 and 3.
Appearing in the work are random permutations,Cayley’s Theorem, Euler numbers, Tutte polynomial . . .
[Kenyon and W. ’09] use elementary calculus andcombinatorics to duplicate and extend some of
these results, i.e. showing that branched polymersof n balls in 3-space have diameter ~ n . 1/2
Slide 22
From this work we also get a method
for generating perfectly random
polymers.
Generating
randompolymers
Slide 23
Phase IV: DIMACS and StatisticalPhysics Face a Brilliant Future
Slide 24
“Boundary Influence”
Riordan
DIMACS DIMACS DIMACS DIMACS DIMACS
DIMACS DIMACS DIMACS DIMACS DIMACS
DIM
AC
S
DIM
AC
S
DIM
AC
S
DIM
AC
S D
IMA
CS
D
IMA
CS
Bollobas
LebowitzBowen
Chayes
BorgsLyons
RadinSteif
Van den Berg
SpencerVigoda
KestenSidoravicius
Martinelli
Schramm
Lovasz
Kannan
PeresSinclair
Jerrum
DyerRandall
HolroydBrightwell
Kenyon
MezardTetali
Montenari
Haggstrom
ProppPropp
SorkinGalvin
Kahn Gacs
Reimer
Slide 25
DIMACS
Happy 20th Birthday!!