Omer AngelAlexander Holroyd Dan RomikBalint Virag math.PR/0609538 Thanks to: Nathanael Berestycki, Alex Gamburd, Alan Hammond, Pawel Hitczenko, Martin

Omer Angel Alexander Holroyd

Dan Romik Balint Virag

math.PR/0609538

Thanks to: Nathanael Berestycki,Alex Gamburd, Alan Hammond,Pawel Hitczenko, Martin Kassabov,Rick Kenyon, Scott Sheffield, David Wilson, Doron Zeilberger

Random Sorting Networks

Adv. in Math. 2007

1234

1234

4321

1234

2134

4321

1234

2134

2314

4321

1234

2134

2314

2341

4321

1234

2134

2314

2341

3241

4321

1234

2134

2314

2341

3241

3421

4321

1234

2134

2314

2341

3241

3421

4321

1234

2134

2314

2341

3241

3421

4321

To get from 1n to n1requires

N:=

nearest-neighbour swaps

1234

2134

2314

2341

3241

3421

4321

any route from 1n to n1in exactly

N:=

nearest-neighbour swaps

A Sorting Network =

E.g. n=4:

Theorem (Stanley 1984). # of n-particle sorting networks =

Uniform Sorting Network (USN):choose an n-particle sorting networkuniformly at random.

E.g. n=3:

1234

2134

2314

2341

3241

3421

4321

1234

2134

2314

2341

3241

3421

4321

swaplocations

1234

2134

2314

2341

3241

3421

4321

swaplocations

particletrajectory

1234

2134

2314

2341

3241

3421

4321

swaplocations

particletrajectory

2341permutation matrix

(at half-time)

9 efficient simulation algorithm for USN...

1 2 3 4

Swap locations, n=100

Swap locations, n=2000

1234

2134

2314

2341

3241

3421

4321

swaplocations

s1

=1

s2

=2

s3

=3

s4

=1

s5

=2

s6

=1

Theorem: For USN:

1. Sequence of swap locations (s1,...,sN) is stationary

2. Scaled first swap location

3. Scaled swap process

as n!1

as n!1

8 n

(Note: not true for all sorting networks, e.g. bubble sort)

In progress: Process of first k swapsin positions cn...cn+k

as n!1not depending on c2(0,1)

! random limit

Proof of stationarity:

1234

2134

2314

2341

3241

3421

4321


2134

2314

2341

3241

3421

4321


1234

1324

1342

3142

3412

4312


1234

1324

1342

3142

3412

4312


1234

1324

1342

3142

3412

4312

4321


1234

1324

1342

3142

3412

4312

4321

(s1,...,sN) (s2,...,sN,n-s1) is a bijectionfrom {sorting networks} to itself.

So for USN:

Selected trajectories, n=2000

Scaled trajectory of particle i:Ti:[0,1]![-1,1]

i

0 1-1

1

Conjecture: trajectories ! random Sine curves:

as n!1

Theorem: scaled trajectories have subsequentiallimits which are a.s. Hölder(½).

as n!1

(random)

Half-time permutation matrix, n=2000

animation

Conjecture: scaled permutation matrix at time N/2

Archimedesmeasure

projection of surfacemeasure on S2 ½ R3 to R2

(unique circularly symmetric measurewith uniform linear projections;

on x2+y2<1 )

scaled permutation matrix at time tN

Arch.meas.

Theorem: scaled permutation matrix at time tNis supported within a certain octagon w.h.p.

as n!1

(1-½p3-)n

Some Proofs:

Key tools:

1. Bijection (Edelman-Greene 1987){sorting networks} $ {standard staircase

Young tableaux}

2. New result forprofile of random staircase Young tableau(deduced from Pittel-Romik 2006)

Staircase Young diagram:

n-1

N cells

(E.g. n=5)

Standard staircase Young tableau:

1 2

5 6

84

3

9

7 10

Fill with 1,,N so each row/col increasing

Edelman-Greene algorithm:

1 2

5 6

84

3

9

7 10

1. Remove largest entry


1 2

5 6

84

3

9

7

1. Remove largest entry


1 2

5 6

84

3

9

7

2. Replace with larger of neighbours " Ã


1 2

5 6

84

3

9

7

2. Replace with larger of neighbours " Ã


1 2

5 6

84

3

9

7

2. Replace with larger of neighbours " Ã...repeat


1

2

5 6

84

3

9

7

2. Replace with larger of neighbours " Ã...repeat


1

2

5 6

84

3

9

7

3. Add 0 in top corner

0


2

3

6 7

95

4

10

8

4. Increment

1


2

3

6 7

95

4 8

5. Repeat everything...

1


2

3

6 7

95

4

8


1


2

3

6 7

95

4

8


1


2

3

6 7

95

4

8


1


2

3

6 7

95

4

8


1

0


3

4

7 8

106

5

9


2

1


3

4

7 8

6

5

9


2

1


4

5

8 9

7

6

10


3

21


4

5

8 9

7

6


3

21



etc

Edelman-Greene Theorem:

After N steps,

45o


After N steps,get swap processof a sorting network!

1 1 1 1 4 4 4 4 5 5 52 2 4 4 1 1 5 5 4 4 43 4 2 2 2 5 1 2 2 2 34 3 3 5 5 2 2 1 1 3 25 5 5 3 3 3 3 3 3 1 1


After N steps,get swap processof a sorting network

And this is a bijection!

And can explicitly describe inverse!

Theorem (Pittel-Romik): For a uniform randomn x n square tableau, 9 limiting shape with contours:

Corollary (AHRV): For uniform random staircase tableau, limiting shape is half of this.

(Proof uses Greene-Nijenhuis-Wilf Hook Walk)

Proof of LLN (swap process ) semic. x Leb.)

Swaps in space-time window [an,bn]x[0,N]come from entries >(1-)N in tableau:

an

bn

entries exiting in [an,bn]

µ

µ

# ¼ area under contour ¼ semicircle

Proof of octagon and Holder bounds

Inverse Edelman-Greene bijection(¼ RSK algorithm) )

# entries <k in 1st row

¸ longest % subseq. of swaps by time k

¸ furthest any particle moves up by time k

So can bound this usinglimit shape.

Why do we believe the conjectures?

The permutahedron: embedding of Cayley graph (Sn, n.n. swaps) in Rn :

-1=(-1(1),...,-1(n))2Rn

n=4:

embedsin an(n-2)-sphere

n=5

Theorem: If a non-random sorting networklies close to some great circle on the permutahedron, then:

1. Trajectories ¼ Sine curves

2. Half-time permutation ¼ Archimedesmeasure

3. Swap process ¼ semicircle x Lebesgue

Theorem: If a non-random sorting networklies close to some great circle on the permutahedron, then:

1. Trajectories ¼ Sine curves

2. Half-time permutation ¼ Archimedesmeasure

3. Swap process ¼ semicircle x Lebesgue

o(n) in | |1Simulations suggestO(pn) for USN!

Proof of Theorem:

close to great circle ) ¼ Sine trajectories (up to a time change), ¼ rotating disc

calculation ) semicircle law

swap rate uniform ) rotation uniform

) no time change

projections uniform ) ¼ Archimedes

Stretchable sorting network:

1

2

34

(, rotating disc)

USN not stretchable w.h.p. as n!1

Main conj. ) - USN “¼ stretchable” w.h.p. - sub-network of k random

items is stretchable w.h.p.

12

3

4

N.B. Not every sorting network lies close toa great circle! E.g. typical network through

n/2...1n...

n/2+1

1...

n/2n/2+1

...n

n...

n/2+1n/2...1

USN

USN

(But this permutation is very unlikely).

Uniform swap model...

Documents

Omer AngelAlexander Holroyd Dan RomikBalint Virag math.PR/0609538 Thanks to: Nathanael Berestycki, Alex Gamburd, Alan Hammond, Pawel Hitczenko, Martin