Learning and Fourier Analysis

Transcription

1 Learning and Fourier Analysis Grigory Yaroslavtsev CIS 625: Computational Learning Theory

2 Fourier Analysis and Learning Powerful tool for PAC-style learning under uniform distribution over 0,1 n Sometimes requires queries of the form f(x) Works for learning many classes of functions, e.g: Monotone, DNF, decision trees, low-degree polynomials Small circuits, halfspaces, k-linear, juntas (depend on small # of variables) Submodular functions (analog of convex/concave) Can be extended to product distributions over 0,1 n, i.e. D = D 1 D 2 D n where means that draws are independent

3 Boolean Functions Book: Analysis of Boolean Functions, Ryan O Donnell f(x 1,, x n ): 0,1 n R Notation switch: f : 1,1 n R Example: f x 1,, x n = x 2 x 3 x n = Lin x 2, x 3, x n f x 1,, x n = x 2 x 3 x n

4 Fourier Expansion Example: min x 1, x 2 = x x x 1x 2 For S,n- let character χ S (x) = i S Thm: Every function f: 1,1 R can be uniquely represented as a multilinear polynomial f x 1,, x n = f S χ S (x) S,n- f S Fourier coefficient of f on S x i

5 Functions = Vectors, Inner Product Functions as vectors form a vector space: f: 1,1 n R f R 2n Inner product on functions = correlation : f, g = 2 n f x g(x) x 1,1 n = E x 1,1 n f x g x For f: 1,1 n 1,1 : f, g = Pr f x = g x Pr f x g x = 1 2dist f, g where dist f, g = Pr,f x g(x)-

6 Orthonormal Basis: Proof Thm: Characters form an orthonormal basis in R 2n χ S, χ T = 1, if S = T χ S, χ T = 0, if S T χ S x χ T x = i S x i j T x j = i SΔT x i 2 j S T x j = i SΔT x i = χ SΔT (x) E x χ S x χ T x = E x χ SΔT x = E x i SΔT x i = i SΔT E x,x i - Since E x x i = 0 we have: E x χ S x χ T x = 0 if SΔT E x χ S x χ T x = 1 if SΔT = This proves that χ s are an orthonormal basis

7 Fourier Coefficients Recall linearity of dot products: f + g, h = f, h + g, h αf, g = α f, g Lemma: f S = f, χ S f, χ S = f T χ T, χ S = f T χ T, χ S = = f T χ T, χ S = f(s) T,n- T,n- T,n-

8 Parseval s Theorem Parseval s Thm: For any f: 1,1 n R f, f = E x 1,1 n f 2 x = f S 2 If f: 1,1 n * 1,1+ then Example: f S 2 S,n- = 1 f = min x 1, x 2 = x x x 1x 2 S,n- S,2- f S 2 = = 1

9 Plancharel s Theorem Plancharel s Thm: For any f, g: 1,1 R f, g = E x 1,1 n f x g x = f S g(s) Proof: f, g = f S χ S, S n T n g T χ T = = f S g T χ S, χ T = f S g(s) S,n- S,T,n- S,n-

10 Basic Fourier Analysis Mean E x f x = f, 1 = f( ) For f: 1,1 n * 1,1+ we have E x f x = Pr f x = 1 Pr f x = 1 2 Variance Var f E x f x E x f x Var f = E x f x E x f x 2 = E x f 2 x E x f x 2 = f 2 (S) S,n- f 2 ( ) (Parseval s Thm) = f 2 (S) S n,s

11 Convolution Def.: For x, y 1,1 n define x y 1,1 n : x y i = x i y i Def.: For f, g: 1,1 n R their convolution f g: 1,1 n R: f g x E y 1,1 n,f y g(x y)- Properties: 1. f g = g f 2. f g h = f g h 3. For all S n : f g S = f S g S

12 Convolution: Proof of Property 3 Property 3: For all S n : f g S Proof: f g S = ॱx 1,1 n f g x χ S x = f S g S = ॱx E y,f y g(x y)-χ S x (def. convolution) = ॱy f y E x,g(x y)χ S x - = ॱy f y E x,g(x y)χ S x y χ S x y χ S x - = ॱy f y χ S y E x,g(x y)χ S x y - = ॱy f y χ S y g S = ॱy f y χ S y g S = f S g S

13 Approximate Linearity Def: f: 0,1 n 0,1 is linear if f x y = f x f y for all x, y 0,1 n S n such that f x = i S x i for all x Def: f: 0,1 n 0,1 is approximately linear if 1. f x y = f x f y for most x, y 0,1 n 2. S n such that f x = i S x i for most x Q: Does 1. imply 2.?

14 Property Testing [Goldreich, Goldwasser, Ron; Rubinfeld, Sudan] Randomized Algorithm Property Tester YES Accept with probability 2 3 YES Accept with probability 2 3 ε-close Don t care NO Reject with probability 2 3 NO Reject with probability 2 3 ε-close : ε fraction has to be changed to become YES

15 Linearity Testing f: 0,1 n *0,1+ P = class of linear functions dist f, P = min dist(f, g) g P ε-close: dist f, P ε Linearity Tester Linear Accept with probability = 1 ε-close Nonlinear Don t care Reject with probability 2 3

16 Linearity Testing [Blum, Luby, Rubinfeld] BLR Test (given access to f: 0,1 n *0,1+): Choose x 0,1 n and y 0,1 n independently Accept if f x f y = f(x y) Thm: If BLR Test accepts with prob. 1 ε then f is εclose to being linear Proof: Apply notation switch 0 1, 1 1 BLR accepts if f x f y = f x y f x f y f x y = 1 if f x f y = f x y f x f y f x y = 0 if f x f y f x y 2 2

17 Linearity Testing: Analysis 1 ε = Pr BLR Accepts f 1 = E x,y f x f y f x y 2 = E x,y f x f y f x y = E x,f x E y,f y f x y -- = E x f x f f x (def. of convolution) = = S n S,n- f S f 3 (S) f f (S) (Plancharel s thm)

18 Linearity Testing: Analysis Continued 1 2ε = S n f 3 S max S n *f S + f2 S S n = max*f S + S n Recall f S = f, χ S = 1 2dist(f, χ S ) Let S =argmax*f S +: 1 2ε 1 2dist(f, χ S ) S n dist f, χ S ε

19 Local Correction Learning linear functions takes n queries Lem: If f is ε-close to linear function χ S then for every x one can compute χ S (x) w.p. 1 2ε as: Pick y 0,1 n Output f y f(x y) Proof: Pr f y χ S y = Pr f x y χ S x y = ε By union bound: Pr f y = χ S y, f x y = χ S x y 1 2ε Then f y f x y = χ S y χ S x y = χ S (x)

20 Thanks!