Lecture 11: Fourier Basics for Boolean functions ...
Transcript of Lecture 11: Fourier Basics for Boolean functions ...
![Page 1: Lecture 11: Fourier Basics for Boolean functions ...](https://reader030.fdocuments.net/reader030/viewer/2022012707/61a83601b122ad6f47510c14/html5/thumbnails/1.jpg)
Lecture 11: Fourier Basics for Boolean functions.Linearity testing.
Lecturer: Ronitt RubinfeldSpring 2014
6.842: Randomness and Computation
![Page 2: Lecture 11: Fourier Basics for Boolean functions ...](https://reader030.fdocuments.net/reader030/viewer/2022012707/61a83601b122ad6f47510c14/html5/thumbnails/2.jpg)
Why all the fuss about Boolean functions?
§ Truth table of a function (complexity theory)
§ Concept to be learned (machine learning)
§ Subset of the Boolean cube (coding theory, combinatorics,…)
§ Etc.
![Page 3: Lecture 11: Fourier Basics for Boolean functions ...](https://reader030.fdocuments.net/reader030/viewer/2022012707/61a83601b122ad6f47510c14/html5/thumbnails/3.jpg)
Why Fourier/Harmonic Analysis?
§ Study “structural properties” of Boolean functions§ Low complexity
§ Depends on few inputs (dictator, junta)
§ “fair” (no variable has too much influence)
§ Homomorphism
§ Spread out/concentrated
![Page 4: Lecture 11: Fourier Basics for Boolean functions ...](https://reader030.fdocuments.net/reader030/viewer/2022012707/61a83601b122ad6f47510c14/html5/thumbnails/4.jpg)
The Boolean function
§ A “new”
repres
entation!
![Page 5: Lecture 11: Fourier Basics for Boolean functions ...](https://reader030.fdocuments.net/reader030/viewer/2022012707/61a83601b122ad6f47510c14/html5/thumbnails/5.jpg)
§
The slick (notational) trick:
![Page 6: Lecture 11: Fourier Basics for Boolean functions ...](https://reader030.fdocuments.net/reader030/viewer/2022012707/61a83601b122ad6f47510c14/html5/thumbnails/6.jpg)
The set of functions and inner product§
![Page 7: Lecture 11: Fourier Basics for Boolean functions ...](https://reader030.fdocuments.net/reader030/viewer/2022012707/61a83601b122ad6f47510c14/html5/thumbnails/7.jpg)
Which basis?§
![Page 8: Lecture 11: Fourier Basics for Boolean functions ...](https://reader030.fdocuments.net/reader030/viewer/2022012707/61a83601b122ad6f47510c14/html5/thumbnails/8.jpg)
A very useful basis:§
![Page 9: Lecture 11: Fourier Basics for Boolean functions ...](https://reader030.fdocuments.net/reader030/viewer/2022012707/61a83601b122ad6f47510c14/html5/thumbnails/9.jpg)
A useful property:
§
![Page 10: Lecture 11: Fourier Basics for Boolean functions ...](https://reader030.fdocuments.net/reader030/viewer/2022012707/61a83601b122ad6f47510c14/html5/thumbnails/10.jpg)
Inner product
§
correlatio
n –
not same a
s in
probability theory
![Page 11: Lecture 11: Fourier Basics for Boolean functions ...](https://reader030.fdocuments.net/reader030/viewer/2022012707/61a83601b122ad6f47510c14/html5/thumbnails/11.jpg)
Orthogonal:
§
![Page 12: Lecture 11: Fourier Basics for Boolean functions ...](https://reader030.fdocuments.net/reader030/viewer/2022012707/61a83601b122ad6f47510c14/html5/thumbnails/12.jpg)
So we have an orthonormal basis!
§
![Page 13: Lecture 11: Fourier Basics for Boolean functions ...](https://reader030.fdocuments.net/reader030/viewer/2022012707/61a83601b122ad6f47510c14/html5/thumbnails/13.jpg)
Some examples:
§
![Page 14: Lecture 11: Fourier Basics for Boolean functions ...](https://reader030.fdocuments.net/reader030/viewer/2022012707/61a83601b122ad6f47510c14/html5/thumbnails/14.jpg)
Fourier coefficients of parity functions:
§
![Page 15: Lecture 11: Fourier Basics for Boolean functions ...](https://reader030.fdocuments.net/reader030/viewer/2022012707/61a83601b122ad6f47510c14/html5/thumbnails/15.jpg)
Agreement with parity function vs. max Fourier coefficient
§
![Page 16: Lecture 11: Fourier Basics for Boolean functions ...](https://reader030.fdocuments.net/reader030/viewer/2022012707/61a83601b122ad6f47510c14/html5/thumbnails/16.jpg)
Distance between parity functions
§
![Page 17: Lecture 11: Fourier Basics for Boolean functions ...](https://reader030.fdocuments.net/reader030/viewer/2022012707/61a83601b122ad6f47510c14/html5/thumbnails/17.jpg)
Plancherel’s Theorem
§
![Page 18: Lecture 11: Fourier Basics for Boolean functions ...](https://reader030.fdocuments.net/reader030/viewer/2022012707/61a83601b122ad6f47510c14/html5/thumbnails/18.jpg)
Parseval’s Theorem
§
![Page 19: Lecture 11: Fourier Basics for Boolean functions ...](https://reader030.fdocuments.net/reader030/viewer/2022012707/61a83601b122ad6f47510c14/html5/thumbnails/19.jpg)
More useful facts:
§
![Page 20: Lecture 11: Fourier Basics for Boolean functions ...](https://reader030.fdocuments.net/reader030/viewer/2022012707/61a83601b122ad6f47510c14/html5/thumbnails/20.jpg)
Linearity (homomorphism) testing
∀ x,y f(x) + f(y) = f(x+y)
![Page 21: Lecture 11: Fourier Basics for Boolean functions ...](https://reader030.fdocuments.net/reader030/viewer/2022012707/61a83601b122ad6f47510c14/html5/thumbnails/21.jpg)
Linearity Property§ Want to quickly test if a function over a group is
linear , that is ∀ x,y f(x) + f(y) = f(x+y)
§ Useful for§ Checking correctness of programs computing matrix,
algebraic, trigonometric functions
§ Probabilistically Checkable Proofs Is the proof of the right format?
§ In these cases, enough for f to be close to homomorphism
![Page 22: Lecture 11: Fourier Basics for Boolean functions ...](https://reader030.fdocuments.net/reader030/viewer/2022012707/61a83601b122ad6f47510c14/html5/thumbnails/22.jpg)
What do we mean by ``close’’?
Definition: f, over domain of size N, is ε-close to linear if can change at most εN
values to turn it into one.Otherwise, ε-far.
![Page 23: Lecture 11: Fourier Basics for Boolean functions ...](https://reader030.fdocuments.net/reader030/viewer/2022012707/61a83601b122ad6f47510c14/html5/thumbnails/23.jpg)
What do we mean by ``quick’’?
§ query complexity measured in terms of domain size N
§ Our goal (if possible): § constant independent of N?
![Page 24: Lecture 11: Fourier Basics for Boolean functions ...](https://reader030.fdocuments.net/reader030/viewer/2022012707/61a83601b122ad6f47510c14/html5/thumbnails/24.jpg)
Linearity Testing
§ If f is linear (i.e., ∀ x,y f(x) + f(y) = f(x+y) ) then test should PASS with probability >2/3
§ If f is ε-far from linear then test should FAIL with probability >2/3
§ Note: If f not linear, but ε-close, then either output is ok
![Page 25: Lecture 11: Fourier Basics for Boolean functions ...](https://reader030.fdocuments.net/reader030/viewer/2022012707/61a83601b122ad6f47510c14/html5/thumbnails/25.jpg)
§
![Page 26: Lecture 11: Fourier Basics for Boolean functions ...](https://reader030.fdocuments.net/reader030/viewer/2022012707/61a83601b122ad6f47510c14/html5/thumbnails/26.jpg)
§
![Page 27: Lecture 11: Fourier Basics for Boolean functions ...](https://reader030.fdocuments.net/reader030/viewer/2022012707/61a83601b122ad6f47510c14/html5/thumbnails/27.jpg)
Proposed Tester:
§
![Page 28: Lecture 11: Fourier Basics for Boolean functions ...](https://reader030.fdocuments.net/reader030/viewer/2022012707/61a83601b122ad6f47510c14/html5/thumbnails/28.jpg)
Characterizing “close” to linear
§
![Page 29: Lecture 11: Fourier Basics for Boolean functions ...](https://reader030.fdocuments.net/reader030/viewer/2022012707/61a83601b122ad6f47510c14/html5/thumbnails/29.jpg)
Nontriviality [Coppersmith]:
§ f: Z3kà Z3k-1
§ f(3h+d)=h, for h < 3k , d ∈{-1,0,1}
§ f satisfies f(x)+f(y) ≠ f(x+y) for only 2/9 of choices of x,y (i.e. δf = 2/9)§ f is 2/3-far from a linear!
![Page 30: Lecture 11: Fourier Basics for Boolean functions ...](https://reader030.fdocuments.net/reader030/viewer/2022012707/61a83601b122ad6f47510c14/html5/thumbnails/30.jpg)
Our goal:
§
![Page 31: Lecture 11: Fourier Basics for Boolean functions ...](https://reader030.fdocuments.net/reader030/viewer/2022012707/61a83601b122ad6f47510c14/html5/thumbnails/31.jpg)
Lemma à Theorem§
![Page 32: Lecture 11: Fourier Basics for Boolean functions ...](https://reader030.fdocuments.net/reader030/viewer/2022012707/61a83601b122ad6f47510c14/html5/thumbnails/32.jpg)
Before the main lemma:
§
![Page 33: Lecture 11: Fourier Basics for Boolean functions ...](https://reader030.fdocuments.net/reader030/viewer/2022012707/61a83601b122ad6f47510c14/html5/thumbnails/33.jpg)
§
Focus here
What is this?
![Page 34: Lecture 11: Fourier Basics for Boolean functions ...](https://reader030.fdocuments.net/reader030/viewer/2022012707/61a83601b122ad6f47510c14/html5/thumbnails/34.jpg)
A final calculation:§
![Page 35: Lecture 11: Fourier Basics for Boolean functions ...](https://reader030.fdocuments.net/reader030/viewer/2022012707/61a83601b122ad6f47510c14/html5/thumbnails/35.jpg)
§
Focus here
0 otherwise
![Page 36: Lecture 11: Fourier Basics for Boolean functions ...](https://reader030.fdocuments.net/reader030/viewer/2022012707/61a83601b122ad6f47510c14/html5/thumbnails/36.jpg)
Linearity tests over other domains
§ Still constant, even for general nonabelian groups
§ Slightly weaker relationship between parameters
![Page 37: Lecture 11: Fourier Basics for Boolean functions ...](https://reader030.fdocuments.net/reader030/viewer/2022012707/61a83601b122ad6f47510c14/html5/thumbnails/37.jpg)
Self-correction
§ Given program P computing linear f that is correct on at least 7/8 of the inputs (BUT YOU DON’T KNOW WHICH ONES!)
§ Can you correctly compute f on each input?§ To compute f(x), can’t just call P on x…
![Page 38: Lecture 11: Fourier Basics for Boolean functions ...](https://reader030.fdocuments.net/reader030/viewer/2022012707/61a83601b122ad6f47510c14/html5/thumbnails/38.jpg)
Self-corrector:
§