# Is seeing believing? Cellular automata in theory and ... holroyd/talks/cumc.pdf¢ Is...

date post

25-Jun-2020Category

## Documents

view

0download

0

Embed Size (px)

### Transcript of Is seeing believing? Cellular automata in theory and ... holroyd/talks/cumc.pdf¢ Is...

Is seeing believing? Cellular automata in theory and experiment

Alexander E. Holroyd, University of British Columbia

CUMC 2008

Cellular automaton:

- regular lattice of cells - cell can be in finite number of possible states

(e.g. alive/dead, full/empty) - local rule for updating states

Idealized models of real-world systems

Easy to describe

(Sometimes) astonishing behaviour...

Mathematical analysis challenging and surprising...

Cells:

empty (water vapour)

full (ice)

Packard’s snowflake models (1984)

Triangular lattice

Start with one full cell

Update rule: full ! full empty ! full if has 1,4,5 or 6 full neighbours

Cells:

empty (water vapour)

full (ice)

Packard’s snowflake models (1984)

Triangular lattice

Start with one full cell

Update rule: full ! full empty ! full if has 1,4,5 or 6 full neighbours

Cells:

empty (water vapour)

full (ice)

Packard’s snowflake models (1984)

Triangular lattice

Start with one full cell

Update rule: full ! full empty ! full if has 1,4,5 or 6 full neighbours

Cells:

empty (water vapour)

full (ice)

Packard’s snowflake models (1984)

Triangular lattice

Start with one full cell

Update rule: full ! full empty ! full if has 1,4,5 or 6 full neighbours

More! (Mirek’s Cellebration) 1,3,5 or 6 1,5 or 6 etc.

(16 interesting rules)

mcell\snow1456.mcl

Source: Janko Gravner

“An elementary schoolchild could look at any of the gorgeous pictures of computer screens in Packard’s collection and instantly identify it as a snowflake.” – Steven Levy

“Simulation by computer may be the only way to predict how certain complicated systems evolve. [. . .] The only practical way to generate the [Packard snowflake] pattern is by computer simulation.” – Stephen Wolfram

Questions: behaviour as time ! 1 ? shape of outer boundary ? internal holes?

Let S = set of eventually full cells (in the infinite lattice)

Guess (from simulations): S1456 = S1346 = the entire lattice S1345, S156 have holes (etc.)

S1346 = entire lattice

S1345, S156 have holes...

Theorem (Gravner and Griffeath, 2006) S1456 has holes!! (but not within distance

109=10000000000 of the origin!)

but:

Theorem (Gravner and Griffeath) The density

exists for all the models, and

r13=r135=5/6, r134=r1345=21/22, r135=r1356=r1346=r13456=1,

r1=0.635§0.001, r14 ,r145 =0.969§0.001, r15=0.803§0.001, r16=0.740§0.001, r156=0.938§0.001,

0.995 < r146 < 1, 0.9999994 < r1456 < 1.

Cells x such that:

time when x becomes full = distance from O to x

Key tool in proof:

Bootstrap Percolation Model

square lattice (Z2)

Cells: full empty

Update rule: full ! full empty ! full if has ¸ 2 full neighbours

square lattice

Cells: full empty

Update rule: full ! full empty ! full if has ¸ 2 full neighbours

square lattice

Cells: full empty

Update rule: full ! full empty ! full if has ¸ 2 full neighbours

square lattice

Cells: full empty

Update rule: full ! full empty ! full if has ¸ 2 full neighbours

square lattice

Cells: full empty

Update rule: full ! full empty ! full if has ¸ 2 full neighbours

square lattice

Cells: full empty

Update rule: full ! full empty ! full if has ¸ 2 full neighbours

square lattice

Cells: full empty

Update rule: full ! full empty ! full if has ¸ 2 full neighbours

square lattice

Cells: full empty

Update rule: full ! full empty ! full if has ¸ 2 full neighbours

Random starting state: Fix 0 < p < 1. Start with each cell:

full with probability p empty with probability 1-p

independently for different cells.Simulations

mcell\bootstrap20.mcl

Guess: for some pcrit ¼ 0.04,

if p > pcrit, every cell eventually full if p < pcrit, not every cell eventually full

but

Theorem (Van Enter 1987) For any p > 0,

P(every cell eventually full) = 1.

Proof: One way to fill everything:

For p>0, (1-p)3+(1-p)5+(1-p)7+L < 1, so P(fill everything) > 0.

P(fill everything) ¸ P(this) = p5 [(1-(1-p)3)(1-(1-p)5)(1-(1-p)7)L]4

Theorem (Zero-One Law): For any translation-invariant event A on the space of p-coin flips on the lattice Zd,

P(A) = 0 or 1.

E.g. {the origin is initially full} not translation-invariant

{every cell is eventually full} is translation-invariant

not affected by translating all coins

Theorem (Zero-One Law): For any translation-invariant event A on the space of p-coin flips on the lattice Zd,

P(A) = 0 or 1.

E.g. {the origin is initially full} not translation-invariant

{every cell is eventually full} is translation-invariant

not affected by translating all coins

Theorem (Zero-One Law): For any translation-invariant event A on the space of p-coin flips on the lattice Zd,

P(A) = 0 or 1.

E.g. {the origin is initially full} not translation-invariant

{every cell is eventually full} is translation-invariant

not affected by translating all coins

P(A) = 0 or 1.

E.g. {the origin is initially full} not translation-invariant

{every cell is eventually full} is translation-invariant

So P(every cell eventually full) = 0 or 1 but P(every cell eventually full) > 0 (from before)

so P(every cell eventually full) = 1.

not affected by translating all coins

Proof of Zero-One Law

For any event A, any e > 0, can find an approximation Ae depending only on coins in a box of size n = n(e):

P(A D Ae) < e

so P(TnA D TnAe) < e.

symmetric differencetranslation by n

Independence: P(Ae Å TnAe) – P(Ae) P(TnAe) = 0 so |P(A Å TnA) – P(A) P(TnA)| < 4e.

But A translation-invariant, so TnA = A !

|P(A) - P(A)2| < 4e

P(A) - P(A)2 = 0

P(A) = 0 or 1.

Consider model on an L by L square. L=5

Going further:

Consider model on an L by L square. L=5

Theorem (Aizenman and Lebowitz, 1989) Let p!0 and L = ea/p.

If a > C then P(fill square) ! 1; if a < c then P(fill square) ! 0.

Going further:

Consider model on an L by L square. L=5

Theorem (Holroyd, 2003) Let p!0 and L = ea/p.

If a > l then P(fill square) ! 1; if a < l then P(fill square) ! 0,

where l = p2/18.

Simulation prediction (Adler, Stauffer, Aharony 1989): l = 0.245 § 0.015

but p2/18 = 0.548311... !

Going further:

1 / log L

p: P(fill)=1/2

Slope 0.245

L=28000L=1020

“crossover?”

Slope p2/18

P(this) = p5 [(1-(1-p)3)(1-(1-p)5)(1-(1-p)7)L]4

) p2/18

Theorem (Gravner, Holroyd, 2008) Let p!0 and L = ea/p.

If a(L) > l - c/plog L then P(fill square) ! 1; if a < l then P(fill square) ! 0,

where l = p2/18.

And further... Understanding the slow convergence:

1/p log 28000 = 0.31... 1/p log 1020 = 0.15...

Need L a L4 to halve the “error”!

Each cell of Z2 contains: North-facing car (") or East-facing car (!) or empty space (0).

At odd time steps, each " tries to move one unit North (succeeds if there is a 0 for it to move into).

At even time steps, each ! tries to move one unit East (succeeds if there is a 0 for it to move into).

Biham-Middleton-Levine traffic model (1992)

" !!

0

" !!

1

" !!

2

" !!

3

" !!

4

" !!

5

" !!

6

!!

7

!

8

Random initial configuration:

0 < p < 1

Each cell of Z2 contains: North-facing car (") with probability p/2 East-facing car (!) with probability p/2 empty space (0) with probability 1 – p

independently for different sites.

Simulation

mcell\b30.mcl

Conjecture. For some 0 < pJ < 1, p > pJ: every car eventually stuck p < pJ: no car eventually stuck

Conjecture. For 0 < pF < 1, p < pF: every car eventually free flowing p > pF: no car eventually free flowing

Question. pF = pJ ?

Intermediate behaviour on finite torus ? (D’Souza 2005)

Theorem (Angel, Holroyd, Martin 2005). For some p1 < 1,

if p > p1 then P(all

Recommended

*View more*