
Title:

Description:

For Israeli and Goldenfeld

it's all about taking

one of two possible paths

and getting to the same place

One path says

Okay, look I have a finegrained description

of the system

I'm going to evolve it forward

with my finegrained rule and at the end

I'm going to simplify the answer
I'm gonna say look yes

I mean I kept track of all these details

but in fact

you know what?

I only need to know some of the output

I only need to know some of the final stages of system

So you can think of that as sort of walking along like this and then projecting up

Another way you can do it though

right, is you can say look you give you this finegrained description

but I know I don't really care so much about this I'm going to project that down.

Here's my finegrain beginning initial condition

I'm going to project that down to a simpler

description and then I'm going to use a new rule

that allows me to evolve
that simplified description forward.

So there are two paths
and if you've done it right,

they'll get you to the same place

evolve the finegrained
system forward and project or

project and evolve the
coarsegrained description forward.

A mathematician would say
that these two operations

right, the operation of evolving forward
and the operation of projecting up commute

you can do one or the other
in either order,

and you'll get the same answer

A then B projecting then evolving is the same as B then A

evolving and projecting.

So let's see how this plays out with a particular example.

Again, I'll just take one from their paper.

This is rule 105. Rule 105 is quite similar
to rule 150 which you've seen before.

It takes the XOR of the three pixels
above the pixel in question

and then inverts them.

So that's the only difference between
rule 105 and 150 is that final inversion.

Another way to think of that is

equivalently, the output is black when
there's an even number of black cells above.

alright

So now you know what we have to do, right?

The first thing is we're going to consider
not the final state of one pixel

but the final state of two pixels.

And we're going to ask what happens

not when you take one time step
but in fact when you take two time steps.

And that means that those two final pixels
will depend on a group of six pixels

two time steps previously.

And now we'll consider those pairs
of pixels to be the supercells.

So you have a big supercell here, which is two pixels
[takes] four possible states

and you have three supercells up here.

So that's our f hat

What we have to do now is
find a combination, a projection

p that takes that supercell
and summarizes it,

simplifies it.

It maps each of those four possible states down to one of two possible states

We have a projection p and we want to
find an evolution operator g

that allows us to evolve forward

those projected down superstates.

So the p is what takes you from the
finegrained descriptions up

to the coarsegrained description

and that g is what's going to take you

between to coarsegrained descriptions
at different times.

So you can either go

f, f, f, f, f, p or

p, g, g

right

For every two times you iterate f you're going
[of course] only to iterate g once

and this simple example

we'll just do the case where you skip one step
and so you have supercells

of size two

so

Fortunately it turns out

It is possible to find a p and g that enable that diagram to commute and here it is in the case of Rule 105

Right in this case the projection rule says look if the supercell has one cell

that's black and one cell that's white make that's all white if

Both cells are white or both cells or black

then make it black

It's sort of like an extra rule

itself in fact

Actually, it looks a little bit like an edge detector

it's as, look, if there's a difference within the supercell mark it one way

But if there's no difference in the supercells
that are homogeneous mark it the other way

If you use that projection operator then it turns out in fact that your g

Right which is now of course taking binary values right take a binary value because you projected the four possible states supercell down to a cell

It only has one of two possible states,
and that's what g operates on

that g evolution operator actually turns out to be rule 150

So what we've shown

is that it's possible to find a nontrivial coarsegraining
??? an interesting one ???

to find a nontrivial coarse graining and

an evolution operator that's still within the space of

cellular automata some of an evolution operator that enables that diagram to commute.

And so now just as we were able to talk about different kinds of

Markov chains coarsegraining into each other

we're now able to talk about how
rules coarsegrain into each other.

And in fact for a nontrivial projection operator Rule 105

coursegrains into rule 150.

Here's what it looks like.

In the top you can see the
finegrained level description

and the bottom you can see
the coarsegrained level description.

At the top there you can see
that we have the smaller pixels

and those smaller pixels are both

small in the x direction along this axis here

and smaller in the time direction

then in the coarsegrained case
the coarsegrained case of course

???mps pairs of states into one and then in fact the jumps are now larger

there are two time steps instead of one.

And by looking at the the comparison
between these two

you can sort of see what's going on, right.
First of all of course

the coarsegrained description is
capturing something interesting

about the finegrained description, right.

We still have this idea that
these triangles are sort of

These these little perturbations that begin

That we begin with that lead to these expanding ways

right we still get that kind of wave like texture.

These sort of propagating spaces that
have kind of internal structure

But you can also see that
we are missing things too, right?

So if you look at those two
triangles at the finegrained level

one of them is sort of darker than the other

But in fact when we coarsegrain the differences
between those two triangles goes away.

So somehow rule 150

when we evolve it forward is
operating our coarsegrained descriptions.

That's thrown out some interesting features of Rule 105

another obvious feature here that
distinguishes rule 105 from rule 150

is that we lose that kind of zebra
stripe pattern

and that's of course because if

A pair of squares is both white
or both black

the projection operator maps them both
into a square that's both black.

So we've lost some of the structure

both in the sort of places
where those opening

Propagating triangle waves have
sort of reached it

and also within the triangles themselves

This gives you a little bit of a better sense

now as we'll see it's not always the case.

That's the picture that you get

when you coarsegrain

looks

similar in some important respects
to the finegrained descriptions

Here it's a particularly elegant example
of how we we're able to capture something about the rule

But of course not everything

we can't really capture everything of course

because that projection operator

is a lossy compression,
it throws out information.

And for rule 105 it really matters

whether everything is white or everything is black

But in fact the rule 150 when
we do the projection

it masks both of those cases to the same state.

so

Israeli and Goldenfeld hacked and they
hacked and hacked and hacked

and they looked at all 256 rules.
And they tried to figure out

how or the extent to which one rule
could coursegrain into another.

So these arrows here
show you how it's possible

to find whether or not it's possible
to find a projection operator

and an evolution operator that
allows one rule set to map to another

upon that coarsegraining and
in fact they consider not only

supercells of size 2
but also size 3 and size 4 and

and computation of this starts getting really hard

because there's so many different kinds
of projection operators you can use

and there's so many different possible

evolutions that you can pick
that you start to run out of time

it gets exponentially hard to find a good projection operator,

that gets exponentially hard to
search the space.

There's only a partial map, but what
you can see here is for example

in the bottom the result that we just
talked through a little bit laboriously

which is the fact that it's possible
to find a projection operator

that takes you from state 105 or

from evolution rule 105 to evolution rule 150.

By the way one of the things you
can notice from this graph

is that it's clear that
Israeli and Goldenfeld

haven't actually found every possible
coarsegraining relationship.

And that's because there should be a
feature of this network

that doesn't actually happen.

And that's that if A coarsegrains into B

If A renormalizes into B,

if it's possible to find a projection
and evolution operator to take A and B

and B renormalizes into C

it's possible to find a projection that takes B into C

then it should also be possible

to renormalize A into C

of course now you're going
to be coursegraining twice

and it's harder of course for
Israeli and Goldenfeld to find those

but if you look at this chart

what you should see

is for example the fact that not only
does rule 23 coarsegrain to rule 128

and not only does rule 128
coarsegrain to rule 0

but it also should be the case
that it's possible for rule 23

to coarsegrain all the way down to rule 0

just by doing two projections
and zooming out even further.

That said Israeli and Goldenfeld did a pretty good job

looking at an enormous number of possible

relationships between all of these possible rulesets

And I find these diagrams quite compelling.

It tells you something really complicated,
really interesting about how

deterministic rules
and deterministic projections

map into each other other.

One of the things that you'll see
from that network

is that not only does rule 105
coarsegrains to rule 150

but in fact rule 50 coarsegrains into itself.

So the pretentious way to say this

this is a fixed point
of renormalization group.

With that projection operator

you actually take rule 150
into a zoomedout version of itself.

You sort of skip a step,
you project down the supercells

and you recover the same rules.

Now it's important to notice
there's a subtlety here, right.

It doesn't mean that
the image itself is selfsimilar

doesn't necessarily mean the rule 150

is kind of fractal in some interesting way.

Because the coarsegraining
may not be

the kind of coarsegraining
that just simply zooms out.

Consider for example the projection
we had going from rule 105 to 150.

Now wasn't a simple decimation in the way
that we did on the Alice picture for example

at the beginning of this renormalization module.

In that case, right, we're renormalizing Alice.
We took her picture,

we looked at little packages of cells,
and we just take one of the values to define

The value of that grid of
that larger grid cell

that's supercell in the Alice case.

But if you remember the
rule 105 to rule 150 projection

that worked and that case
was actually an edge detector.

If the cell was all white
or all black

it got mapped to something
that was all black.

So it doesn't necessarily mean
that if you kind of fuzz

rule 150 it still looks like rule 150.

It really depends upon the details
of that projection operator.

That said in fact you might
think of it another way.

Ihe rule 150 is a fixed point
of renormalization group

with potentially a much more interesting projection

than a simple decimation.