49
Invariance principle on the slice Y. Filmus, E. Mossel, G. Kindler, K. Wimmer 30 May 2016

Invariance principle on the slice fileInvariance principle on the slice Y. Filmus, E. Mossel, G. Kindler, K. Wimmer 30 May 2016

  • Upload
    votu

  • View
    215

  • Download
    0

Embed Size (px)

Citation preview

Page 1: Invariance principle on the slice fileInvariance principle on the slice Y. Filmus, E. Mossel, G. Kindler, K. Wimmer 30 May 2016

Invariance principle on the slice

Y. Filmus, E. Mossel, G. Kindler, K. Wimmer

30 May 2016

Page 2: Invariance principle on the slice fileInvariance principle on the slice Y. Filmus, E. Mossel, G. Kindler, K. Wimmer 30 May 2016

1 The result

2 Fourier analysis on the slice

3 The proof

4 Summary

Page 3: Invariance principle on the slice fileInvariance principle on the slice Y. Filmus, E. Mossel, G. Kindler, K. Wimmer 30 May 2016

Constant-weight vectorsfool low-degree polynomials∗

Page 4: Invariance principle on the slice fileInvariance principle on the slice Y. Filmus, E. Mossel, G. Kindler, K. Wimmer 30 May 2016

In more detail...

Suppose P is a suitable low-degree polynomial.

If (X1, . . . ,Xn) ∼ {0, 1}n and (Y1, . . . ,Yn) ∼( [n]n/2

)then

P(X1, . . . ,Xn) ≈ P(Y1, . . . ,Yn).

Page 5: Invariance principle on the slice fileInvariance principle on the slice Y. Filmus, E. Mossel, G. Kindler, K. Wimmer 30 May 2016

In more detail...

Suppose P is a suitable low-degree polynomial.

If (X1, . . . ,Xn) ∼ {0, 1}n and (Y1, . . . ,Yn) ∼( [n]n/2

)then

P(X1, . . . ,Xn) ≈ P(Y1, . . . ,Yn).

Page 6: Invariance principle on the slice fileInvariance principle on the slice Y. Filmus, E. Mossel, G. Kindler, K. Wimmer 30 May 2016

Unsuitable polynomials

Goal

P(~X ) ≈ P( ~Y ) where ~X ∼ {0, 1}n, ~Y ∼( [n]n/2

)

Obstructions

x1 + · · ·+ xn − n/2

(x1 + · · ·+ xn − n/2)x1

x1(1− n/2) + (x2 + · · ·+ xn)x1

Page 7: Invariance principle on the slice fileInvariance principle on the slice Y. Filmus, E. Mossel, G. Kindler, K. Wimmer 30 May 2016

Unsuitable polynomials

Goal

P(~X ) ≈ P( ~Y ) where ~X ∼ {0, 1}n, ~Y ∼( [n]n/2

)Obstructions

x1 + · · ·+ xn − n/2

(x1 + · · ·+ xn − n/2)x1

x1(1− n/2) + (x2 + · · ·+ xn)x1

Page 8: Invariance principle on the slice fileInvariance principle on the slice Y. Filmus, E. Mossel, G. Kindler, K. Wimmer 30 May 2016

Unsuitable polynomials

Goal

P(~X ) ≈ P( ~Y ) where ~X ∼ {0, 1}n, ~Y ∼( [n]n/2

)Obstructions

x1 + · · ·+ xn − n/2

(x1 + · · ·+ xn − n/2)x1

x1(1− n/2) + (x2 + · · ·+ xn)x1

Page 9: Invariance principle on the slice fileInvariance principle on the slice Y. Filmus, E. Mossel, G. Kindler, K. Wimmer 30 May 2016

Unsuitable polynomials

Goal

P(~X ) ≈ P( ~Y ) where ~X ∼ {0, 1}n, ~Y ∼( [n]n/2

)Obstructions

x1 + · · ·+ xn − n/2

(x1 + · · ·+ xn − n/2)x1

x1(1− n/2) + (x2 + · · ·+ xn)x1

Page 10: Invariance principle on the slice fileInvariance principle on the slice Y. Filmus, E. Mossel, G. Kindler, K. Wimmer 30 May 2016

Suitable polynomials

Definition

A polynomial P is harmonic if

n∑i=1

∂P

∂xi= 0.

Examples

x1 − x2(x1 − x2)(x3 − x4)

Theorem (Dunkl)

Every function on( [n]n/2

)has unique representation as

harmonic multilinear polynomial of degree ≤ n/2.

Page 11: Invariance principle on the slice fileInvariance principle on the slice Y. Filmus, E. Mossel, G. Kindler, K. Wimmer 30 May 2016

Suitable polynomials

Definition

A polynomial P is harmonic if

n∑i=1

∂P

∂xi= 0.

Examples

x1 − x2

(x1 − x2)(x3 − x4)

Theorem (Dunkl)

Every function on( [n]n/2

)has unique representation as

harmonic multilinear polynomial of degree ≤ n/2.

Page 12: Invariance principle on the slice fileInvariance principle on the slice Y. Filmus, E. Mossel, G. Kindler, K. Wimmer 30 May 2016

Suitable polynomials

Definition

A polynomial P is harmonic if

n∑i=1

∂P

∂xi= 0.

Examples

x1 − x2(x1 − x2)(x3 − x4)

Theorem (Dunkl)

Every function on( [n]n/2

)has unique representation as

harmonic multilinear polynomial of degree ≤ n/2.

Page 13: Invariance principle on the slice fileInvariance principle on the slice Y. Filmus, E. Mossel, G. Kindler, K. Wimmer 30 May 2016

Suitable polynomials

Definition

A polynomial P is harmonic if

n∑i=1

∂P

∂xi= 0.

Examples

x1 − x2(x1 − x2)(x3 − x4)

Theorem (Dunkl)

Every function on( [n]n/2

)has unique representation as

harmonic multilinear polynomial of degree ≤ n/2.

Page 14: Invariance principle on the slice fileInvariance principle on the slice Y. Filmus, E. Mossel, G. Kindler, K. Wimmer 30 May 2016

Formal statement

For harmonic multilinear polynomial P and test function ϕ s.t.

degP = o(√n)

‖P‖ = 1

ϕ is Lipschitz

holds: ∣∣∣∣∣∣ E~X∼{0,1}n

[ϕ(P(~X ))]− E~Y∼( [n]

n/2)[ϕ(P( ~Y ))]

∣∣∣∣∣∣ = o(1).

Works also for other slices.

Page 15: Invariance principle on the slice fileInvariance principle on the slice Y. Filmus, E. Mossel, G. Kindler, K. Wimmer 30 May 2016

Formal statement

For harmonic multilinear polynomial P and test function ϕ s.t.

degP = o(√n)

‖P‖ = 1

ϕ is Lipschitz

holds: ∣∣∣∣∣∣ E~X∼{0,1}n

[ϕ(P(~X ))]− E~Y∼( [n]

n/2)[ϕ(P( ~Y ))]

∣∣∣∣∣∣ = o(1).

Works also for other slices.

Page 16: Invariance principle on the slice fileInvariance principle on the slice Y. Filmus, E. Mossel, G. Kindler, K. Wimmer 30 May 2016

Consequences

Results for {0, 1}n proved using classical invariance principlegeneralize (almost) immediately to the slice.

Some examples

Majority is Stablest for the slice.

Bourgain’s tail bound for the slice.

Kindler–Safra theorem for the slice.

Page 17: Invariance principle on the slice fileInvariance principle on the slice Y. Filmus, E. Mossel, G. Kindler, K. Wimmer 30 May 2016

Consequences

Results for {0, 1}n proved using classical invariance principlegeneralize (almost) immediately to the slice.

Some examples

Majority is Stablest for the slice.

Bourgain’s tail bound for the slice.

Kindler–Safra theorem for the slice.

Page 18: Invariance principle on the slice fileInvariance principle on the slice Y. Filmus, E. Mossel, G. Kindler, K. Wimmer 30 May 2016

Consequences

Results for {0, 1}n proved using classical invariance principlegeneralize (almost) immediately to the slice.

Some examples

Majority is Stablest for the slice.

Bourgain’s tail bound for the slice.

Kindler–Safra theorem for the slice.

Page 19: Invariance principle on the slice fileInvariance principle on the slice Y. Filmus, E. Mossel, G. Kindler, K. Wimmer 30 May 2016

Consequences

Results for {0, 1}n proved using classical invariance principlegeneralize (almost) immediately to the slice.

Some examples

Majority is Stablest for the slice.

Bourgain’s tail bound for the slice.

Kindler–Safra theorem for the slice.

Page 20: Invariance principle on the slice fileInvariance principle on the slice Y. Filmus, E. Mossel, G. Kindler, K. Wimmer 30 May 2016

1 The result

2 Fourier analysis on the slice

3 The proof

4 Summary

Page 21: Invariance principle on the slice fileInvariance principle on the slice Y. Filmus, E. Mossel, G. Kindler, K. Wimmer 30 May 2016

Coarse spectral decomposition

Theorem (Dunkl)

Every function P on( [n]n/2

)has unique representation as

harmonic multilinear polynomial of degree ≤ n/2.

Can decompose P into sum of homogeneous polynomials.

Theorem (Dunkl)

Homogeneous parts are orthogonal on all slices, and so on {0, 1}n.

Page 22: Invariance principle on the slice fileInvariance principle on the slice Y. Filmus, E. Mossel, G. Kindler, K. Wimmer 30 May 2016

Coarse spectral decomposition

Theorem (Dunkl)

Every function P on( [n]n/2

)has unique representation as

harmonic multilinear polynomial of degree ≤ n/2.

Can decompose P into sum of homogeneous polynomials.

Theorem (Dunkl)

Homogeneous parts are orthogonal on all slices, and so on {0, 1}n.

Page 23: Invariance principle on the slice fileInvariance principle on the slice Y. Filmus, E. Mossel, G. Kindler, K. Wimmer 30 May 2016

Coarse spectral decomposition

Theorem (Dunkl)

Every function P on( [n]n/2

)has unique representation as

harmonic multilinear polynomial of degree ≤ n/2.

Can decompose P into sum of homogeneous polynomials.

Theorem (Dunkl)

Homogeneous parts are orthogonal on all slices, and so on {0, 1}n.

Page 24: Invariance principle on the slice fileInvariance principle on the slice Y. Filmus, E. Mossel, G. Kindler, K. Wimmer 30 May 2016

L2 invariance

Theorem (F)

If a harmonic multilinear polynomial P has low degree then

L2 norm of P on( [n]n/2

)≈ L2 norm of P on {0, 1}n.

Prerequisite for full invariance principle.

Page 25: Invariance principle on the slice fileInvariance principle on the slice Y. Filmus, E. Mossel, G. Kindler, K. Wimmer 30 May 2016

L2 invariance

Theorem (F)

If a harmonic multilinear polynomial P has low degree then

L2 norm of P on( [n]n/2

)≈ L2 norm of P on {0, 1}n.

Prerequisite for full invariance principle.

Page 26: Invariance principle on the slice fileInvariance principle on the slice Y. Filmus, E. Mossel, G. Kindler, K. Wimmer 30 May 2016

Fourier basis for the slice

Srinivasan and F (independently) constructed an explicitorthogonal basis for the slice.

F used basis to reprove Wimmer’s junta theorem for the slice.

Basis isn’t needed to prove the invariance principle.

Page 27: Invariance principle on the slice fileInvariance principle on the slice Y. Filmus, E. Mossel, G. Kindler, K. Wimmer 30 May 2016

Fourier basis for the slice

Srinivasan and F (independently) constructed an explicitorthogonal basis for the slice.

F used basis to reprove Wimmer’s junta theorem for the slice.

Basis isn’t needed to prove the invariance principle.

Page 28: Invariance principle on the slice fileInvariance principle on the slice Y. Filmus, E. Mossel, G. Kindler, K. Wimmer 30 May 2016

Fourier basis for the slice

Srinivasan and F (independently) constructed an explicitorthogonal basis for the slice.

F used basis to reprove Wimmer’s junta theorem for the slice.

Basis isn’t needed to prove the invariance principle.

Page 29: Invariance principle on the slice fileInvariance principle on the slice Y. Filmus, E. Mossel, G. Kindler, K. Wimmer 30 May 2016

1 The result

2 Fourier analysis on the slice

3 The proof

4 Summary

Page 30: Invariance principle on the slice fileInvariance principle on the slice Y. Filmus, E. Mossel, G. Kindler, K. Wimmer 30 May 2016

Classical proof

Classical invariance principle (MOO)

Suitable low-degree polynomials behave similarlyon input {−1, 1}n and on input N(0, 1)n.

The proof

Replace one by one the {−1, 1} variables with N(0, 1) variablesand bound the error.

Our case

Strategy fails since coordinates of( [n]n/2

)are not independent!

Page 31: Invariance principle on the slice fileInvariance principle on the slice Y. Filmus, E. Mossel, G. Kindler, K. Wimmer 30 May 2016

Classical proof

Classical invariance principle (MOO)

Suitable low-degree polynomials behave similarlyon input {−1, 1}n and on input N(0, 1)n.

The proof

Replace one by one the {−1, 1} variables with N(0, 1) variablesand bound the error.

Our case

Strategy fails since coordinates of( [n]n/2

)are not independent!

Page 32: Invariance principle on the slice fileInvariance principle on the slice Y. Filmus, E. Mossel, G. Kindler, K. Wimmer 30 May 2016

Classical proof

Classical invariance principle (MOO)

Suitable low-degree polynomials behave similarlyon input {−1, 1}n and on input N(0, 1)n.

The proof

Replace one by one the {−1, 1} variables with N(0, 1) variablesand bound the error.

Our case

Strategy fails since coordinates of( [n]n/2

)are not independent!

Page 33: Invariance principle on the slice fileInvariance principle on the slice Y. Filmus, E. Mossel, G. Kindler, K. Wimmer 30 May 2016

Our proof(s)

Let P be a suitable low-degree polynomial on( [n]n/2

).

Dist. of P doesn’t change too much if we flip a coordinate.

Hence dist. of P on( [n]n/2

)is similar to its dist. on

( [n]n/2±t

).

and so similar to its distribution on( [n]n/2−t≤·≤n/2+t

).

Hence dist. of P on( [n]n/2

)is similar to its dist. on {0, 1}n.

Page 34: Invariance principle on the slice fileInvariance principle on the slice Y. Filmus, E. Mossel, G. Kindler, K. Wimmer 30 May 2016

Our proof(s)

Let P be a suitable low-degree polynomial on( [n]n/2

).

Dist. of P doesn’t change too much if we flip a coordinate.

Hence dist. of P on( [n]n/2

)is similar to its dist. on

( [n]n/2±t

).

and so similar to its distribution on( [n]n/2−t≤·≤n/2+t

).

Hence dist. of P on( [n]n/2

)is similar to its dist. on {0, 1}n.

Page 35: Invariance principle on the slice fileInvariance principle on the slice Y. Filmus, E. Mossel, G. Kindler, K. Wimmer 30 May 2016

Our proof(s)

Let P be a suitable low-degree polynomial on( [n]n/2

).

Dist. of P doesn’t change too much if we flip a coordinate.

Hence dist. of P on( [n]n/2

)is similar to its dist. on

( [n]n/2±t

).

and so similar to its distribution on( [n]n/2−t≤·≤n/2+t

).

Hence dist. of P on( [n]n/2

)is similar to its dist. on {0, 1}n.

Page 36: Invariance principle on the slice fileInvariance principle on the slice Y. Filmus, E. Mossel, G. Kindler, K. Wimmer 30 May 2016

Our proof(s)

Let P be a suitable low-degree polynomial on( [n]n/2

).

Dist. of P doesn’t change too much if we flip a coordinate.

Hence dist. of P on( [n]n/2

)is similar to its dist. on

( [n]n/2±t

).

and so similar to its distribution on( [n]n/2−t≤·≤n/2+t

).

Hence dist. of P on( [n]n/2

)is similar to its dist. on {0, 1}n.

Page 37: Invariance principle on the slice fileInvariance principle on the slice Y. Filmus, E. Mossel, G. Kindler, K. Wimmer 30 May 2016

1 The result

2 Fourier analysis on the slice

3 The proof

4 Summary

Page 38: Invariance principle on the slice fileInvariance principle on the slice Y. Filmus, E. Mossel, G. Kindler, K. Wimmer 30 May 2016

Recap

Constant-weight vectors foollow-degree harmonic multilinear polys

Page 39: Invariance principle on the slice fileInvariance principle on the slice Y. Filmus, E. Mossel, G. Kindler, K. Wimmer 30 May 2016

Open questions

Fourier analysis on the slice

Known: KKL (O’D–W), junta theorem (W), FKN (F).

Kindler–Safra theorem: sub-optimal version known (this work)

Linearity testing: some versions known (DDGKS, Tal)

Other domains

Association schemes (e.g. Grassmann)

Non-abelian groups (e.g. Sn)

Multislice

Applications

Suggestions welcome!

Page 40: Invariance principle on the slice fileInvariance principle on the slice Y. Filmus, E. Mossel, G. Kindler, K. Wimmer 30 May 2016

Open questions

Fourier analysis on the slice

Known: KKL (O’D–W), junta theorem (W), FKN (F).

Kindler–Safra theorem: sub-optimal version known (this work)

Linearity testing: some versions known (DDGKS, Tal)

Other domains

Association schemes (e.g. Grassmann)

Non-abelian groups (e.g. Sn)

Multislice

Applications

Suggestions welcome!

Page 41: Invariance principle on the slice fileInvariance principle on the slice Y. Filmus, E. Mossel, G. Kindler, K. Wimmer 30 May 2016

Open questions

Fourier analysis on the slice

Known: KKL (O’D–W), junta theorem (W), FKN (F).

Kindler–Safra theorem: sub-optimal version known (this work)

Linearity testing: some versions known (DDGKS, Tal)

Other domains

Association schemes (e.g. Grassmann)

Non-abelian groups (e.g. Sn)

Multislice

Applications

Suggestions welcome!

Page 42: Invariance principle on the slice fileInvariance principle on the slice Y. Filmus, E. Mossel, G. Kindler, K. Wimmer 30 May 2016

Open questions

Fourier analysis on the slice

Known: KKL (O’D–W), junta theorem (W), FKN (F).

Kindler–Safra theorem: sub-optimal version known (this work)

Linearity testing: some versions known (DDGKS, Tal)

Other domains

Association schemes (e.g. Grassmann)

Non-abelian groups (e.g. Sn)

Multislice

Applications

Suggestions welcome!

Page 43: Invariance principle on the slice fileInvariance principle on the slice Y. Filmus, E. Mossel, G. Kindler, K. Wimmer 30 May 2016

Open questions

Fourier analysis on the slice

Known: KKL (O’D–W), junta theorem (W), FKN (F).

Kindler–Safra theorem: sub-optimal version known (this work)

Linearity testing: some versions known (DDGKS, Tal)

Other domains

Association schemes (e.g. Grassmann)

Non-abelian groups (e.g. Sn)

Multislice

Applications

Suggestions welcome!

Page 44: Invariance principle on the slice fileInvariance principle on the slice Y. Filmus, E. Mossel, G. Kindler, K. Wimmer 30 May 2016

Open questions

Fourier analysis on the slice

Known: KKL (O’D–W), junta theorem (W), FKN (F).

Kindler–Safra theorem: sub-optimal version known (this work)

Linearity testing: some versions known (DDGKS, Tal)

Other domains

Association schemes (e.g. Grassmann)

Non-abelian groups (e.g. Sn)

Multislice

Applications

Suggestions welcome!

Page 45: Invariance principle on the slice fileInvariance principle on the slice Y. Filmus, E. Mossel, G. Kindler, K. Wimmer 30 May 2016

Open questions

Fourier analysis on the slice

Known: KKL (O’D–W), junta theorem (W), FKN (F).

Kindler–Safra theorem: sub-optimal version known (this work)

Linearity testing: some versions known (DDGKS, Tal)

Other domains

Association schemes (e.g. Grassmann)

Non-abelian groups (e.g. Sn)

Multislice

Applications

Suggestions welcome!

Page 46: Invariance principle on the slice fileInvariance principle on the slice Y. Filmus, E. Mossel, G. Kindler, K. Wimmer 30 May 2016

Thanks!

Page 47: Invariance principle on the slice fileInvariance principle on the slice Y. Filmus, E. Mossel, G. Kindler, K. Wimmer 30 May 2016

Fourier basis for the slice

Setup

A slice([n]k

).

Basis elements

For every ` ≤ min(k , n − k) and b1 < · · · < b` ≤ n:

χb1,...,b` =∑′

a1,...,a`

(xa1 − xb1) · · · (xa` − xb`),

where:

a1, b1, . . . , a`, b` all distinct.

a1 < b1, . . . , a` < b`.

Properties(n`

)−( n`−1)

basis elements at level `.

Basis is orthogonal.

Explicit formula for norm of χb1,...,b` .

Page 48: Invariance principle on the slice fileInvariance principle on the slice Y. Filmus, E. Mossel, G. Kindler, K. Wimmer 30 May 2016

Fourier basis for the slice

Setup

A slice([n]k

).

Basis elements

For every ` ≤ min(k , n − k) and b1 < · · · < b` ≤ n:

χb1,...,b` =∑′

a1,...,a`

(xa1 − xb1) · · · (xa` − xb`),

where:

a1, b1, . . . , a`, b` all distinct.

a1 < b1, . . . , a` < b`.

Only consider b1, . . . , b` for which sum is non-zero.

Properties(n`

)−( n`−1)

basis elements at level `.

Basis is orthogonal.

Explicit formula for norm of χb1,...,b` .

Page 49: Invariance principle on the slice fileInvariance principle on the slice Y. Filmus, E. Mossel, G. Kindler, K. Wimmer 30 May 2016

Fourier basis for the slice

Basis elements

For every ` ≤ min(k , n − k) and b1 < · · · < b` ≤ n:

χb1,...,b` =∑′

a1,...,a`

(xa1 − xb1) · · · (xa` − xb`),

where:

a1, b1, . . . , a`, b` all distinct.

a1 < b1, . . . , a` < b`.

Properties(n`

)−( n`−1)

basis elements at level `.

Basis is orthogonal.

Explicit formula for norm of χb1,...,b` .