Slides - Events @ CSA Dept., IISc Bangalore
Download
Report
Transcript Slides - Events @ CSA Dept., IISc Bangalore
Algebraic Property Testing:
A Unified Perspective
Arnab Bhattacharyya
Indian Institute of Science
Property Testing
Distinguish between
and
and
Property P
π-far from property P
Testing and Learning
ο
Proper learning (with membership queries) is as hard
as testing, for any property
ο
For many natural properties, testing is much easier
than learning
ο
Learning always requires time at least as large as size
of hypothesis but testing can be in constant time!
A brief history
ο
Initially appeared as a tool for program checking [BlumLuby-Rubinfeld β90]
ο
[Babai-Fortnow-Lund β90, Rubinfeld-Sudan β96]: application
to PCPs and low-degree testing
ο
[Goldreich-Goldwasser-Ron β98] considered property testing
as full-fledged computational problem in its own right
ο
Now, many other variants and connections known (e.g.,
implicit learning, active testing, coding theory,
inapproximability, β¦)
Testable Properties
Property P is testable if there exist functions π and πΏ and a
randomized algorithm A such that, given an input object
and a parameter π:
β’
A makes π(π) queries into input object
β’
A rejects with prob. β₯ πΏ(π) if input is π-far from P and
with prob. 0 if π = 0
πΏ
Positive at
π>0
Zero at π = 0
Ο΅
Testing all-orangeness
P = {all-orange rectangle}
β’ π=1
β’ Query hits non-orange region with
probability β₯ π
Testing linearity
P = {linear functions πΏ: {π, π}π β {π, π} }
000β¦
0
000β¦
1
001β¦
1
001β¦
1
010β¦
1
010β¦
0
100β¦
0
100β¦
1
101β¦
1
111β¦
0
111β¦
0
Functions of the form
101β¦
1
π³ π , β¦ , ππ = πππ + πππ + π110β¦
ππ + β―0+ πππ
110β¦ π 1
over characteristic011β¦
2
011β¦
0
1
β’ π=3
β’ π π₯ + π π¦ β π π₯ + π¦ with probability Ξ©(π). [BlumLuby-Rubinfeld]
Testing low-degree
P = {polynomials π: πππ β ππ of degree β€ π} }
Functions of the form
π· ππ , β¦ , ππ =
π
π
π
πππ β―ππ β
πππ πππ β― πππ
Check that function restricted
to a random O(1)-dim
ππ +β―+ππ β€π
subspace is poly of degree β€ π
over characteristic p
Test detects a violation with probability Ξ©(π) [KaufmanRon, Haramaty-Shpilka-Sudan]
Affine-invariant properties
Subject of this talk: βalgebraic propertiesβ of
multivariate functions over finite fields.
Definition (affine-invariant property):
If a function f on domain πππ satisfies an
affine-invariant property P, then π β π΄ must
also satisfy P for every affine map π΄: πππ β
πππ
[Kaufman-Sudan]
Why?
Introspective:
ο A principled understanding of testability (~
VC dimension theory for PAC learning)
ο What kinds of tools are needed to prove
testability? Limitations?
Extrospective:
ο Search for locally testable codes (applications
to inapproximability)
Testing Fourier Sparsity
P = {functions π: 0,1 π β {0,1} have at
most k non-zero Fourier coefficients}
[Gopalan-OβDonnell-Servedio-Shpilka-Wimmer]
showed P is testable (though only with πΏ 0 > 0).
Testing Decompositions
ο
P = {functions π: πππ β ππ that are a product of two
quadratics}
ο
P = {functions π: πππ β ππ that are the square of a
quartic}
ο
P = {polynomials π: πππ β ππ of the form ππ + ππ
where π, π, π, π are all cubics}
Testability of all such properties open previously!
Main Result
An exact characterization of testability for affineinvariant properties:
Theorem: An affine-invariant property is
testable* if and only if it is locally
characterized
[Joint work with Fischer, H&P Hatami, Lovett β13]
Locally Characterized
Definition (locally characterized):
A property P is locally characterized if there
always exists a constant-sized witness to
non-membership in P.
Examples:
β’ Linearity: If π isnβt linear, then there exist π₯, π¦ such that
π π₯ + π π¦ β π(π₯ + π¦).
β’ Low degree: If deg π > π, then there exists a point at
which the (π + 1)-th order derivative is nonzero.
Necessity of locality
Theorem: An affine-invariant property is
testable if and only if it is locally
characterized.
Proof of : The set of queries by the tester that
make it reject is itself a π-sized witness to nonmembership in the property.
Proof of
: Rest of this talkβ¦
Degree-structural properties
ο
Are properties like βFourier sparsityβ, βproduct
of two quadraticsβ, βsquare of a quarticβ, βsum
of two products of quadraticsβ, βsplitting into d
linear formsβ locally characterized?
Theorem: Any property described as the
property of decomposing into a known
structure of low-degree polynomials is
locally characterized.
Subsequent work
ο
Degree-structural properties are not only
testable but reconstructible!
β¦ Given degree-π poly π = π1 π2 + π3 π4 where
π1 , β¦ , π4 are degree-(π/2) polys, we can
reconstruct π1 , β¦ , π4 by making π(ππ ) queries to
π. [B. β14]
ο
Characterization of two-sided testability
[Yoshida β14]
Some drawbacksβ¦
Unfolded proof uses
non-constructive
results. We get no
bound on the locality
and query
complexity for even
simple properties
like βproduct of two
quadraticsβ!
A Running Example: RBG
π₯+π¦
π₯+π¦+π§
Witness to non-membership:
π₯
π₯+π§
RBG square
A function π: ππ β {π
ππ, π΅ππ’π, πΊππππ} satisfies the RBG
property if there are no π₯, π¦, π§ β ππ such that π π₯ =
π π₯ + π¦ = π
ππ, π π₯ + π§ = π΅ππ’π, and π π₯ + π¦ + π§ =
πΊππππ.
The RBG claim
Claim: The RBG property is testable with 4
queries.
Suffices to show that if π: ππ β
{π
ππ, π΅ππ’π, πΊππππ} is far from
RBG, then a random tuple
(π₯, π₯ + π¦, π₯ + π§, π₯ + π¦ + π§) is an
RBG square with constant
probability.
π₯+π¦
π₯
π₯+π¦+π§
π₯+π§
Dreams of a proof
(Image © Kozmic Konstructions)
The dreamiest situation
Suppose non-RBG function π has the form:
π(π₯) = Ξ π₯1 , π₯2 , β¦ , π₯π
(0,4)
β’ ππ cells of equal size
(0,3)
(0,2)
(0,1)
(0,0)
β’ Must exist at least one
RBG square
β’ Probability of selecting
an RBG square is πβ3π .
A slight nudge
Same analysis works if non-RBG function π
has the form:
π(π₯) = Ξ β1 π₯ , β2 π₯ , β¦ , βπ (π₯)
where β1 , β2 , β¦ , βπ are linearly independent
linear forms.
Also works if βlinearly
independentβ replaced
by βrandomβ
More rocking of the bed
Suppose non-RBG function π has the form:
π(π₯) = Ξ π1 π₯ , π2 π₯ , β¦ , ππ (π₯)
where π1 , β¦ , ππ are random bounded degree nonlinear polynomials.
Partitioning by random polys
(0,4)
(0,3)
Joint distribution of
π1 , β¦ , ππ
close to uniform
distribution
(0,2)
(0,1)
(0,0)
ππ cells of roughly
equal size
Partitioning by random polys
(0,4)
(0,3)
(0,2)
Joint distribution of
(π1 (π₯), . . , ππ (π₯), π1 (π₯ +
π¦), . . . , ππ (π₯ + π¦), π1 (π₯ +
π§), . . . , ππ (π₯ + π§), π1 (π₯ +
π¦ + π§), . . . , ππ (π₯ + π¦ +
π§))close to uniform
(0,1)
(0,0)
Probability of selecting an
RBG square is β πβ3π .
Returning to intermediate dream
Suppose non-RBG function π has the form:
π(π₯) = Ξ π1 π₯ , π2 π₯ , β¦ , ππ (π₯)
where π1 , β¦ , ππ are arbitrary low degree polynomials.
Instead of insisting π1 , β¦ , ππ be truly random, can
we weaken the requirement?
Returning to intermediate dream
Suppose non-RBG function π has the form:
π(π₯) = Ξ π1 π₯ , π2 π₯ , β¦ , ππ (π₯)
where π1 , β¦ , ππ are arbitrary low degree polynomials.
How to ensure (π1 , β¦ , ππ ) distributed close to
uniform? How to even ensure π1 distributed close
to uniform?
High rank polynomials
Theorem [Green-Tao, Kaufmann-Lovett]: A
polynomial π: πΉ π β πΉ is distributed close to
uniform if Q is of high rank.
Definition: A polynomial π: πΉ π β πΉ is of rank > π if
there are no polynomials π1 , β¦ , ππ of degrees less than
deg(Q) such that
π = Ξ π1 , β¦ , ππ
for some function Ξ.
High rank polynomial collection
Theorem [Green-Tao, Kaufmann-Lovett]:
Polynomials π1 , β¦ , ππ : πΉ π β πΉ jointly distributed
close to uniform if every nontrivial linear
combination of π1 , β¦ , ππ is high rank.
Polynomial Regularity Lemma
A straightforward inductive argument shows that,
given
π(π₯) = Ξ π1 π₯ , π2 π₯ , β¦ , ππ (π₯)
we can find a high rank collection of polynomials
π1 , β¦ , ππ β² such that:
π(π₯) = Ξβ² π1 π₯ , π2 π₯ , β¦ , ππ β² (π₯)
π1 , β¦ , ππ β² form cells of roughly equal size
Equidistribution of squares?
Recall we also wanted equidistribution of squares. Is
high rank sufficient for this purpose?
Wanted Lemma: If π1 , β¦ , ππ β² of sufficiently high
rank, then:
π1 π₯ , . . , ππ β² π₯ ,
π1 π₯ + π¦ , . . . , ππ β² π₯ + π¦ ,
π1 π₯ + π§ , . . . , ππ β² π₯ + π§ ,
π1 π₯ + π¦ + π§ , . . . , ππ β² π₯ + π¦ + π§
distributed close to uniform.
A bit of a nightmareβ¦
Wanted lemma not necessarily true!
Suppose π1 is of degree 1. Then:
π1 π₯ β π1 π₯ + π¦ β π1 π₯ + π§ + π1 π₯ + π¦ + π§ = 0
identically.
Lesson: some correlations may be present because
of degree, no matter how large the rank is
Corrected equidistribution
Lemma: Given polynomials π1 , β¦ , ππ of high rank,
either thereβs some linear identity among
ππ π₯ , ππ π₯ + π¦ , ππ π₯ + π§ , ππ (π₯ + π¦ + π§) for some i,
or the tuple
π1 π₯ , . . , ππ π₯ ,
π1 π₯ + π¦ , . . . , ππ π₯ + π¦ ,
π1 π₯ + π§ , . . . , ππ π₯ + π§ ,
π1 π₯ + π¦ + π§ , . . . , ππ π₯ + π¦ + π§
is close to uniformly distributed.
Proof uses βGowers
inverse theoremβ
[Tao-Ziegler]
Applying corrected equidistribution
π
π
2
2
π
4
π
π
π
π
π
Suppose π1 satisfies: π1 π₯ β
π1 π₯ + π¦ β π1 π₯ + π§ + π1 (π₯ +
Almost awake to realityβ¦
Now, letβs remove all restrictions on RBG-far
function π: ππ β {π
ππ, π΅ππ’π, πΊππππ}.
Can we βapproximateβ f by some function g of
the form:
π(π₯) = Ξ π1 π₯ , π2 π₯ , β¦ , ππ (π₯)
where π1 , β¦ , ππ are low degree polynomials?
Regularity Lemma
Given any function π: ππ β {0,1} and integer π, there exist
constantly many polynomials π1 , β¦ , ππ : ππ β π of degrees at
most π, a function Ξ: ππ β [0,1] and a function β: ππ β [β1,1]
such that:
β’ π π₯ = Ξ π1 π₯ , β¦ , ππ π₯
+β π₯
β’ Collection π1 , β¦ , ππ has high rank
Uses βGowers
inverse theoremβ
[Tao-Ziegler]
β’ Ξ(π1 , β¦ , ππ ) is the average value of π on the cell indexed
(π1 , β¦ , ππ )
β’ (π + 1)-th order Gowers norm of h is small
Victory in dreamland!
ο
Letting π π : πΉ π β {0,1} be indicator function of color c.
Then, rejection probability is:
πΈπ₯,π¦,π§ [π πππ π₯ π πππ π₯ + π¦ π πππ’π π₯ + π§ π πππππ π₯ + π¦ + π§ ]
ο
We apply the regularity lemma. The term with low
Gowers norm contributes negligibly to expectation.
ο
A combinatorial argument finishes the claim.
Waking up to hard reality
ο
βInverse conjecture for the Gowers norm is falseβ [LovettMeshulam-Samorodnitsky]
ο
Turns out that in the conjecture, one needs to modify the
definition of polynomials to include functions that are not πvalued.
ο
Need to reprove regularity lemma and equidistribution
claims for such βnon-classical polynomialsβ (technical core of
our work)
ο
Also, several lies encountered in this dream sequence
addressed
Non-classical polynomials
Definition (non-classical polynomial):
A function π: πππ β β/β€ is of degree β€ π if
Ξβ1 Ξβ2 β― Ξβπ+1 π β‘ 0
for any β1 , β2 , β¦ , βπ+1 β πππ , where:
Ξβ π π₯ = π π₯ + β β π(π₯)
β’ If range of π is
1 2
πβ1
0, , , β¦ ,
π π
π
same as classical
, non-classical
Non-classical polynomials
Definition (non-classical polynomial):
A function π: πππ β β/β€ is of degree β€ π if
Ξβ1 Ξβ2 β― Ξβπ+1 π β‘ 0
for any β1 , β2 , β¦ , βπ+1 β πππ , where:
Ξβ π π₯ = π π₯ + β β π(π₯)
1 1 3
4 2 4
β’ Function π: πΉ2 β 0, , ,
defined as
1
π 0 = 0, π 1 =
4
is non-classical poly of degree 2.
Gowers norm and inverse theorem
ο
For functions π: πΉππ β β, the Gowers norm of order d measures
the expected multiplicative derivative of π at a random point
in π random directions.
ο
Fact: If Gowers norm of f of order d is 1 and f takes values
inside the unit disk, then f is the exponential of a non-classical
polynomial of degree β€ π β 1.
ο
Gowers Inverse Theorem: If Gowers norm of f of order d is
πΏ > 0 and f takes values inside the unit disk, then f is
correlated with a non-classical polynomial of degree β€ π β 1.
Open Questions
ο
Proof for degree-structural property also uses the same
core technical ideas, so no good locality bounds even for
βsimpleβ properties like product of two quadratics.
ο
Give non-trivial lower bound for testing any natural
affine-invariant property. No superpolynomial lower
bounds known in 1/π.
ο
Find other uses of equidistribution of high rank
polynomials.
THANK YOU!