ryan odonnell carnegie mellon university karl wimmer cmu & duquesne amir shpilka technion rocco...
TRANSCRIPT
Ryan O’Donnell
Carnegie Mellon University
Karl Wimmer
CMU & Duquesne
Amir Shpilka
Technion
Rocco Servedio
Columbia
Parikshit Gopalan
UW & Microsoft SVC
joint with
Outline
1. Testing boolean functions overview
2. Statement & proof sketch of Result 2
3. Statements of other results
Property Testing Boolean Functions
• Query access to
• Say YES whp if f has property P
• Say NO whp if f is ϵ-far from all g w/ ppty P
• P “testable” if doable with # queries
depending only on ϵ
Property Testing
Characterizing testability fairly
well-understood for
graphs [AS05a,AS05b,AFNS06,AT08,
…]
& codes [KS07,KS08,…].
But wide open for boolean functions.
Testable boolean function properties
• Linearity (f(x) = λ•x) [BLR90]
• Degree k [AKKLR03]
• Dicatators [BGS95]
• Conjunctions, size-s mono. DNF [PRS01]
• s-juntas (≤ s relevant vbls) [FKKRS03,B09]
• size-s DNFs, DTs, BPs, formulas,
ckts, polynomials [DLMORSW07]
• Halfspaces [MORS09]
poly(s/ϵ) q’s
Characterization??
Fischer’s survey [Fis01]
suggests Fourier may be key.
This paper:
Having concise Fourier representation
is testable.
Fourier analysis of boolean functions
Fourier analysis of boolean functions
Fourier analysis of boolean functions
There are 2n linear functions,
Every uniquely expressible as
a real #
Fourier sparsity
Def: f is s-sparse
⇔ # of nonzero is ≤ s
Eg: Linear functions are 1-sparse.
k-juntas are 2k-sparse.
Result #2
Thm: “Is s-sparse?”
is ϵ-testable with poly(s/ϵ)
nonadaptive queries.
Proof ingredients
• Hashing Fourier coefficients
to affine subspaces [FGKP06]
• New structural facts about
s-sparse boolean functions
Physical space Fourier space
Physical space Fourier space
+ + − − −
− + − − +
− − + + +
− + + + −
− − − − +
0 0 0 0 0
0 0 0 ⅝ 0
0 –⅜ 0 0 0
0 0 0 0 0
0 0 0 ¾ 0
Fourier space
Hashing Fourier coefficients idea
s2
buckets
Hashing Fourier coefficients idea
• Birthday Paradox ⇒ s Fourier coeffs split
• Test that at most s buckets are nonzero?
s2
buckets
Hashing to affine subspaces
Pick α1, …, αd ∈ at random, d = 2 log s.
α1 • λ = 0
α2 • λ = 0
αd • λ = 0
λ :subpace of
codimension d
Hashing to affine subspaces
Pick α1, …, αd ∈ at random, d = 2 log s.
α1 • λ = b1
α2 • λ = b2
αd • λ = bd
λ :subpace of
codimension d
affine
b ∈
Hashing Fourier coefficients idea
Birthday Paradox? OK, by pairwise independence.
2d = s2
buckets(aff. subsps.)
indexed by F2
Projection functions
Projection functions
Projection functions
b
Physical to Fourier link
b
ex:
Physical to Fourier link
b
ex: Pbf(x) = avg. of ±f on 2d = s2 strings rel’d to x
Projection functions
• We can compute Pbf(x) exactly,
for any bucket b, with s2 queries
• Pbf(x) is always
The Test
1. Hash to a random 2d = s2 buckets b.
2. For each b,
3. ϵ’-test if Pbf(x) ≡ 0, for ϵ’ =
poly(ϵ/s)
4. Let B = {b : Pbf wasn’t ≡ 0}
5. Say NO if |B| > s
The Test
1. Hash to a random 2d = s2 buckets b.
2. For each b,
3. ϵ’-test if Pbf(x) ≡ 0, for ϵ’ = poly(ϵ/s)
4. Let B = {b : Pbf wasn’t ≡ 0}
5. Say NO if |B| > s
6. Say YES
The Test
1. Hash to a random 2d = s2 buckets b.
2. For each b,
3. ϵ’-test if Pbf(x) ≡ 0, for ϵ’ = poly(ϵ/s)
4. Let B = {b : Pbf wasn’t ≡ 0}
5. Say NO if |B| > s
6. Say YES poly(s/ϵ) nonadapt. queries
The Test
Recall that for b ∈ B we “should” have Pbf =
Analysis easier if you also do:
6. For each b ∈ B,
7. Test |Pbf| constant
8. Test sgn(Pbf) is linear, using [BLR90]
The Test
Recall that for b ∈ B we “should” have Pbf =
Analysis easier if you also do:
6. For each b ∈ B,
7. Test |Pbf| constant
8. Test sgn(Pbf) is linear, using [BLR90]
Key for analysis
Granularity Theorem:
If is s-sparse, then
for all λ.
(hence 0 or > 1/s)
Other results
Def: f is k-dimensional if { }
lies in a k-dimensional subspace.
⇔ f is a “k-junta of parities”
Result 1: “Is f k-dimensional?” is ϵ-testable
with 2O(k)/ϵ nonadaptive queries.
Other results
Result 3: Lower bounds.
Even for nonadaptive testers, ϵ = .49,
• queries needed for s-sparsity
• 2Ω(k) queries need for k-dimensionality
Improves on [AKKLR03, BFNR08].
Other results
Result 4: Exact, proper learning algorithm
for s-sparse functions under unif.
(Easy consequence of Granularity Theorem.)
Other results
Result 5: “Every subclass of k-dimensional
functions is testable with 2O(k)/ϵ
nonadaptive queries.”
(Uses “Testing by implicit learning” [DLMORSW07])
Open problems
• More on characterizing testability of
boolean function properties.
• Test functions with (cf [GS06])
• Characterize s-sparse boolean functions.
Are they disj. unions of poly(s) many
affine subspaces of codimension O(log s)?