the polynomial method strikes back: tight quantum query...
TRANSCRIPT
![Page 1: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/1.jpg)
The Polynomial Method Strikes Back: TightQuantum Query Bounds Via Dual Polynomials
Justin Thaler (Georgetown)
Joint work with:Mark Bun (Princeton)
Robin Kothari (MSR Redmond)
![Page 2: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/2.jpg)
Boolean Functions
Boolean function f : −1, 1n → −1, 1
ANDn(x) =
−1 (TRUE) if x = (−1)n
1 (FALSE) otherwise
![Page 3: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/3.jpg)
Approximate Degree
A real polynomial p ε-approximates f if
|p(x)− f(x)| < ε ∀x ∈ −1, 1n
degε(f) = minimum degree needed to ε-approximate f
deg(f) := deg1/3(f) is the approximate degree of f
![Page 4: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/4.jpg)
Example 1: The Approximate Degree of ANDn
![Page 5: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/5.jpg)
Example: What is the Approximate Degree of ANDn?
deg(ANDn) = Θ(√n).
Upper bound: Use Chebyshev Polynomials.
Markov’s Inequality: Let G(t) be a univariate polynomial s.t.deg(G) ≤ d and maxt∈[−1,1] |G(t)| ≤ 1. Then
maxt∈[−1,1]
|G′(t)| ≤ d2.
Chebyshev polynomials are the extremal case.
![Page 6: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/6.jpg)
Example: What is the Approximate Degree of ANDn?
deg(ANDn) = O(√n).
After shifting a scaling, can turn degree O(√n) Chebyshev
polynomial into a univariate polynomial Q(t) that looks like:
!"#$%&'()*+*&',*
Define n-variate polynomial p via p(x) = Q(∑n
i=1 xi/n).
Then |p(x)−ANDn(x)| ≤ 1/3 ∀x ∈ −1, 1n.
![Page 7: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/7.jpg)
Example: What is the Approximate Degree of ANDn?
[NS92] deg(ANDn) = Ω(√n).
Lower bound: Use symmetrization.
Suppose |p(x)−ANDn(x)| ≤ 1/3 ∀x ∈ −1, 1n.
There is a way to turn p into a univariate polynomial psym
that looks like this:
!"#$%&'()*+*&',*
Claim 1: deg(psym) ≤ deg(p).
Claim 2: Markov’s inequality =⇒ deg(psym) = Ω(n1/2).
![Page 8: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/8.jpg)
Why Care about Approximate Degree?
![Page 9: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/9.jpg)
Applications of deg Upper Bounds
Upper bounds on degε(f) yield efficient learning algorithms.
ε ≈ 1/3: Agnostic Learning [KKMS05]
ε ≈ 1− 2−nδ: Attribute-Efficient Learning [KS04, STT12]
ε→ 1 (i.e., threshold degree, deg±(f)): PAC learning [KS01]
Upper bounds on deg1/3(f) also:
Imply fast algorithms for differentially private data release[TUV12, CTUW14].Underly the best known lower bounds on formula complexityand graph complexity [Tal2014, 2016a, 2016b]
![Page 10: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/10.jpg)
Applications of deg Upper Bounds
Upper bounds on degε(f) yield efficient learning algorithms.
ε ≈ 1/3: Agnostic Learning [KKMS05]
ε ≈ 1− 2−nδ: Attribute-Efficient Learning [KS04, STT12]
ε→ 1 (i.e., threshold degree, deg±(f)): PAC learning [KS01]
Upper bounds on deg1/3(f) also:
Imply fast algorithms for differentially private data release[TUV12, CTUW14].
Underly the best known lower bounds on formula complexityand graph complexity [Tal2014, 2016a, 2016b]
![Page 11: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/11.jpg)
Applications of deg Upper Bounds
Upper bounds on degε(f) yield efficient learning algorithms.
ε ≈ 1/3: Agnostic Learning [KKMS05]
ε ≈ 1− 2−nδ: Attribute-Efficient Learning [KS04, STT12]
ε→ 1 (i.e., threshold degree, deg±(f)): PAC learning [KS01]
Upper bounds on deg1/3(f) also:
Imply fast algorithms for differentially private data release[TUV12, CTUW14].Underly the best known lower bounds on formula complexityand graph complexity [Tal2014, 2016a, 2016b]
![Page 12: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/12.jpg)
This Talk: Two Focuses Involving deg Lower Bounds
Focus 1: A nearly optimal bound on the approximate degreeof AC0, and its applications [BT17].
Focus 2: Proving tight quantum query lower bounds forspecific functions [BKT17].
![Page 13: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/13.jpg)
First Focus: Approximate Degree of AC0
Approximate degree is a key tool for understanding AC0.
At the heart of the best known bounds on the complexity ofAC0 under measures such as:
Multi-Party (Quantum) Communication ComplexityApproximate RankSign-rank ≈ Unbounded Error Communication (UPP)Discrepancy≈Margin complexityMajority-of-Threshold circuit sizeThreshold-of-Majority circuit sizeand more.
Problem 1: Is there a function on n variables thatis in AC0, and has approximate degree Ω(n)?
![Page 14: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/14.jpg)
First Focus: Approximate Degree of AC0
Approximate degree is a key tool for understanding AC0.
At the heart of the best known bounds on the complexity ofAC0 under measures such as:
Multi-Party (Quantum) Communication ComplexityApproximate RankSign-rank ≈ Unbounded Error Communication (UPP)Discrepancy≈Margin complexityMajority-of-Threshold circuit sizeThreshold-of-Majority circuit sizeand more.
Problem 1: Is there a function on n variables thatis in AC0, and has approximate degree Ω(n)?
![Page 15: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/15.jpg)
Approximate Degree of AC0: Details
Best known result: Ω(n2/3) for the Element Distinctnessfunction (Aaronson and Shi, 2004).
Our result: For any constant δ > 0, a function in AC0 withapproximate degree Ω(n1−δ).
More precisely, circuit depth is O(log(1/δ)).Lower bound also applies to DNFs of polylogarithmic width(and quasipolynomial size).
![Page 16: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/16.jpg)
Approximate Degree of AC0: Details
Best known result: Ω(n2/3) for the Element Distinctnessfunction (Aaronson and Shi, 2004).
Our result: For any constant δ > 0, a function in AC0 withapproximate degree Ω(n1−δ).
More precisely, circuit depth is O(log(1/δ)).
Lower bound also applies to DNFs of polylogarithmic width(and quasipolynomial size).
![Page 17: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/17.jpg)
Approximate Degree of AC0: Details
Best known result: Ω(n2/3) for the Element Distinctnessfunction (Aaronson and Shi, 2004).
Our result: For any constant δ > 0, a function in AC0 withapproximate degree Ω(n1−δ).
More precisely, circuit depth is O(log(1/δ)).Lower bound also applies to DNFs of polylogarithmic width(and quasipolynomial size).
![Page 18: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/18.jpg)
Applications
Nearly optimal Ω(n1−δ) lower bounds on quantumcommunication complexity of AC0.
Essentially optimal (quadratic) separation of certificatecomplexity and approximate degree.
Better secret sharing schemes with reconstruction in AC0.
![Page 19: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/19.jpg)
Second Focus: Quantum Query Complexity
In the quantum query model, a quantum algorithm is givenquery access to the bits of an input x.
Goal: compute some function f of x while minimizing thenumber of queried bits.
Most quantum algorithms were discovered in or can easily bedescribed in the query setting.
![Page 20: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/20.jpg)
Connecting deg and Quantum Query Complexity
Let A be a quantum algorithm making at most T queries.
[BBC+01] there is a polynomial p of degree 2T such that
p(x) = Pr[A(x) = 1].
So A computes f to error ε =⇒ 2p(x)− 1 approximates f toerror 2ε.So deg(f) is a lower bound on the quantum query complexityof f .
This is called the polynomial method in quantum querycomplexity.
![Page 21: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/21.jpg)
Our Results
Problem Prior Upper Bound Our Lower Bound Prior Lower Bound
k-distinctness O(n3/4−1/(2k+2−4)) Ω(n3/4−1/(2k)) Ω(n2/3)
Image Size Testing O(√n logn) Ω(
√n) Ω(n1/3)
k-junta Testing O(√k log k) Ω(
√k) Ω(k1/3)
SDU O(√n) Ω(
√n) Ω(n1/3)
Shannon Entropy O(√n) Ω(
√n) Ω(n1/3)
Our lower bounds on quantum query complexity and deg vs. prior work.
Problem Prior Upper Bound Our Upper and Lower Bounds Prior Lower Bound
Surjectivity O(n3/4) O(n3/4) and Ω(n3/4) Ω(n2/3)
Our bounds on the approximate degree of Surjectivity vs. prior work.
![Page 22: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/22.jpg)
Lower Bound Methods in Quantum Query Complexity
Since 2002, the positive-weights adversary method, and thenewer negative-weights adversary method have been tools ofchoice for proving quantum query lower bounds.
Negative-weights method can prove a tight lower bound forany function [Rei11, LMR+11].But is often challenging to apply to specific functions.
Quantum query bounds proved via approximate degree “lift”to communication lower bounds [She11].
Not known to hold for adversary methods.
![Page 23: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/23.jpg)
Ruminations on the Polynomial Method
Intuitively, how do we resolve questions that have resistedadversary methods?
A key fact exploited in our analysis is:
Fact (1)
Any polynomial p : −1, 1n → R satisfying the following conditionsrequires degree Ω(n1/4):
|p(x)−ORn(x)| ≤ 1/3 if |x| ≤ n1/4
|p(x)| ≤ exp(|x| · n−1/4) if |x| > n1/4.
Fact (1) is “non-quantum” because any quantum queryalgorithm always produces polynomials bounded in [0, 1].
Reasoning about such “non-quantum” polynomials seemsdifficult to capture by adversary methods.
![Page 24: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/24.jpg)
Prior Work: The Method of Dual Polynomials andthe AND-OR Tree
![Page 25: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/25.jpg)
Beyond Symmetrization
Symmetrization is “lossy”: in turning an n-variate poly p intoa univariate poly psym, we throw away information about p.
Challenge Problem: What is deg(AND-ORn)?
1/2
1/2
1/2
![Page 26: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/26.jpg)
History of the AND-OR Tree
Theorem
deg(AND-ORn) = Θ(n1/2).
Tight Upper Bound of O(n1/2)
[HMW03] via quantum algorithms[BNRdW07] different proof of O(n1/2 · log n) (via error reduction+composition)[She13] different proof of tight upper bound (via robustification)
Tight Lower Bound of Ω(n1/2)
[BT13] and [She13] via the method of dual polynomials
![Page 27: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/27.jpg)
History of the AND-OR Tree
Theorem
deg(AND-ORn) = Θ(n1/2).
Tight Upper Bound of O(n1/2)
[HMW03] via quantum algorithms[BNRdW07] different proof of O(n1/2 · log n) (via error reduction+composition)[She13] different proof of tight upper bound (via robustification)
Tight Lower Bound of Ω(n1/2)
[BT13] and [She13] via the method of dual polynomials
![Page 28: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/28.jpg)
History of the AND-OR Tree
Theorem
deg(AND-ORn) = Θ(n1/2).
Tight Upper Bound of O(n1/2)
[HMW03] via quantum algorithms[BNRdW07] different proof of O(n1/2 · log n) (via error reduction+composition)[She13] different proof of tight upper bound (via robustification)
Tight Lower Bound of Ω(n1/2)
[BT13] and [She13] via the method of dual polynomials
![Page 29: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/29.jpg)
Linear Programming Formulation of Approximate Degree
What is best error achievable by any degree d approximation of f?Primal LP (Linear in ε and coefficients of p):
minp,ε ε
s.t. |p(x)− f(x)| ≤ ε for all x ∈ −1, 1n
deg p ≤ d
Dual LP:
maxψ∑
x∈−1,1nψ(x)f(x)
s.t.∑
x∈−1,1n|ψ(x)| = 1
∑x∈−1,1n
ψ(x)q(x) = 0 whenever deg q ≤ d
![Page 30: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/30.jpg)
Dual Characterization of Approximate Degree
Theorem: degε(f) > d iff there exists a “dual polynomial”ψ : −1, 1n → R with
(1)∑
x∈−1,1nψ(x)f(x) > ε “high correlation with f”
(2)∑
x∈−1,1n|ψ(x)| = 1 “L1-norm 1”
(3)∑
x∈−1,1nψ(x)q(x) = 0, when deg q ≤ d “pure high degree d”
A lossless technique. Strong duality implies any approximatedegree lower bound can be witnessed by dual polynomial.
Example: 2−n · PARITYn witnesses the fact thatdegε(PARITYn) = n for any ε < 1.
![Page 31: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/31.jpg)
Dual Characterization of Approximate Degree
Theorem: degε(f) > d iff there exists a “dual polynomial”ψ : −1, 1n → R with
(1)∑
x∈−1,1nψ(x)f(x) > ε “high correlation with f”
(2)∑
x∈−1,1n|ψ(x)| = 1 “L1-norm 1”
(3)∑
x∈−1,1nψ(x)q(x) = 0, when deg q ≤ d “pure high degree d”
A lossless technique. Strong duality implies any approximatedegree lower bound can be witnessed by dual polynomial.
Example: 2−n · PARITYn witnesses the fact thatdegε(PARITYn) = n for any ε < 1.
![Page 32: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/32.jpg)
Goal: Construct an explicit dual polynomialψAND-OR for AND-OR
![Page 33: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/33.jpg)
Constructing a Dual Polynomial
By [NS92], there are dual polynomials
ψOUT for deg (ANDn1/2) = Ω(n1/4) and
ψIN for deg (ORn1/2) = Ω(n1/4)
Both [She13] and [BT13] combine ψOUT and ψIN to obtain adual polynomial ψAND-OR for AND-OR.
The combining method was proposed in earlier work by [SZ09,Lee09, She09].
![Page 34: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/34.jpg)
The Combining Method [SZ09, She09, Lee09]
ψAND-OR(x1, . . . , xn1/2) := C · ψOUT(. . . , sgn(ψIN(xi)), . . . )
n1/2∏i=1
|ψIN(xi)|
(C chosen to ensure ψAND-OR has L1-norm 1).
Must verify:
1 ψAND-OR has pure high degree ≥ n1/4 · n1/4 = n1/2.
2 ψAND-OR has high correlation with AND-OR.
1/2
1/2
1/2
![Page 35: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/35.jpg)
The Combining Method [SZ09, She09, Lee09]
ψAND-OR(x1, . . . , xn1/2) := C · ψOUT(. . . , sgn(ψIN(xi)), . . . )
n1/2∏i=1
|ψIN(xi)|
(C chosen to ensure ψAND-OR has L1-norm 1).
Must verify:
1 ψAND-OR has pure high degree ≥ n1/4 · n1/4 = n1/2.
2 ψAND-OR has high correlation with AND-OR.
1/2
1/2
1/2
![Page 36: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/36.jpg)
The Combining Method [SZ09, She09, Lee09]
ψAND-OR(x1, . . . , xn1/2) := C · ψOUT(. . . , sgn(ψIN(xi)), . . . )
n1/2∏i=1
|ψIN(xi)|
(C chosen to ensure ψAND-OR has L1-norm 1).
Must verify:
1 ψAND-OR has pure high degree ≥ n1/4 · n1/4 = n1/2.X[She09]
2 ψAND-OR has high correlation with AND-OR. [BT13, She13]
1/2
1/2
1/2
![Page 37: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/37.jpg)
Recent Progress on the Complexity of AC0:Applying the Method of Dual Polynomials to
Block-Composed Functions
![Page 38: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/38.jpg)
(Negative) One-Sided Approximate Degree
Negative one-sided approximate degree is an intermediatenotion between approximate degree and threshold degree.
A real polynomial p is a negative one-sided ε-approximationfor f if
|p(x)− 1| < ε ∀x ∈ f−1(1)
p(x) ≤ −1 ∀x ∈ f−1(−1)
odeg−,ε(f) = min degree of a negative one-sidedε-approximation for f .
Examples: odeg−,1/3(ANDn) = Θ(√n); odeg−,1/3(ORn) = 1.
![Page 39: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/39.jpg)
Recent Theorems
Theorem (BT13, She13)
Let f be a Boolean function with odeg−,1/2(f) ≥ d. LetF = ORt(f, . . . , f). Then deg1/2(F ) ≥ d ·
√t.
Theorem (BT14)
Let f be a Boolean function with odeg−,1/2(f) ≥ d. LetF = ORt(f, . . . , f). Then deg1−2−t(F ) ≥ d.
Theorem (She14)
Let f be a Boolean function with odeg−,1/2(f) ≥ d. LetF = ORt(f, . . . , f). Then deg±(F ) = Ω(mind, t).
Theorem (BCHTV16)
Let f be a Boolean function with deg1/2(f) ≥ d. LetF = GAPMAJt(f, . . . , f). Then deg±(F ) ≥ Ω(mind, t).
![Page 40: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/40.jpg)
Recent Theorems
Theorem (BT13, She13)
Let f be a Boolean function with odeg−,1/2(f) ≥ d. LetF = ORt(f, . . . , f). Then deg1/2(F ) ≥ d ·
√t.
Theorem (BT14)
Let f be a Boolean function with odeg−,1/2(f) ≥ d. LetF = ORt(f, . . . , f). Then deg1−2−t(F ) ≥ d.
Theorem (She14)
Let f be a Boolean function with odeg−,1/2(f) ≥ d. LetF = ORt(f, . . . , f). Then deg±(F ) = Ω(mind, t).
Theorem (BCHTV16)
Let f be a Boolean function with deg1/2(f) ≥ d. LetF = GAPMAJt(f, . . . , f). Then deg±(F ) ≥ Ω(mind, t).
![Page 41: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/41.jpg)
Recent Theorems
Theorem (BT13, She13)
Let f be a Boolean function with odeg−,1/2(f) ≥ d. LetF = ORt(f, . . . , f). Then deg1/2(F ) ≥ d ·
√t.
Theorem (BT14)
Let f be a Boolean function with odeg−,1/2(f) ≥ d. LetF = ORt(f, . . . , f). Then deg1−2−t(F ) ≥ d.
Theorem (She14)
Let f be a Boolean function with odeg−,1/2(f) ≥ d. LetF = ORt(f, . . . , f). Then deg±(F ) = Ω(mind, t).
Theorem (BCHTV16)
Let f be a Boolean function with deg1/2(f) ≥ d. LetF = GAPMAJt(f, . . . , f). Then deg±(F ) ≥ Ω(mind, t).
![Page 42: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/42.jpg)
Recent Theorems
Theorem (BT13, She13)
Let f be a Boolean function with odeg−,1/2(f) ≥ d. LetF = ORt(f, . . . , f). Then deg1/2(F ) ≥ d ·
√t.
Theorem (BT14)
Let f be a Boolean function with odeg−,1/2(f) ≥ d. LetF = ORt(f, . . . , f). Then deg1−2−t(F ) ≥ d.
Theorem (She14)
Let f be a Boolean function with odeg−,1/2(f) ≥ d. LetF = ORt(f, . . . , f). Then deg±(F ) = Ω(mind, t).
Theorem (BCHTV16)
Let f be a Boolean function with deg1/2(f) ≥ d. LetF = GAPMAJt(f, . . . , f). Then deg±(F ) ≥ Ω(mind, t).
![Page 43: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/43.jpg)
Reminder
Problem 1: Is there a function on n variables thatis in AC0, and has approximate degree Ω(n)?
![Page 44: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/44.jpg)
Our Techniques
![Page 45: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/45.jpg)
Hardness Amplification in AC0
Hardness Amplification in AC0
1–2-t 1–2-t Theorem Template: If f is “hard” to
approximate by low-degree polynomials, then F = g ¢ f is “even harder” to approximate by low-degree polynomials
g
f f
x1 xn
…
1–2-t “Block Composition Barrier”
Robust approximations, i.e., deg(g ¢ f ) ≤ O(deg(g) deg(f )) imply that block composition cannot increase approximate degree as a function of n
~ ~ ~
![Page 46: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/46.jpg)
Around the Block-Composition Barrier
Around the Block Composition Barrier
Prior work: • Hardness amplification “from the top” • Block composed functions
This work: • Hardness amplification “from the bottom”
• Non-block-composed functions
g
f f
x1 xn
…
f
g g…
![Page 47: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/47.jpg)
A General Hardness Amplification Result
Theorem (Strong Hardness Amplification Within AC0)
Let f : −1, 1n → −1, 1be computed by an AC0 circuit of depth k, and
deg(f) ≥ d.Then there exists an F on O(n log2 n) variables that
is computed by an AC0 circuit of depth k + 3, and
deg(F ) ≥ n1/2 · d1/2
Remarks:
E.g.: If f = AND, then deg(F ) ≥ n3/4.
Recursive application yields Ω(n1−δ) bound for AC0 function.
Analogous result holds for monotone DNF.
![Page 48: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/48.jpg)
A General Hardness Amplification Result
Theorem (Strong Hardness Amplification Within AC0)
Let f : −1, 1n → −1, 1be computed by an AC0 circuit of depth k, and
deg(f) ≥ d.Then there exists an F on O(n log2 n) variables that
is computed by an AC0 circuit of depth k + 3, and
deg(F ) ≥ n1/2 · d1/2
Remarks:
E.g.: If f = AND, then deg(F ) ≥ n3/4.
Recursive application yields Ω(n1−δ) bound for AC0 function.
Analogous result holds for monotone DNF.
![Page 49: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/49.jpg)
Idea of the Hardness Amplification Construction
![Page 50: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/50.jpg)
Idea of the Hardness-Amplifying Construction
Consider the function SURJECTIVITY : −1, 1n → −1, 1.Let n = N logR. SURJ interprets its input x as a list of Nnumbers (x1, . . . , xN ) from a range [R].SURJR,N (x) = −1 if and only if every element of the range[R] appears at least once in the list.
When we apply Main Theorem to f = ANDR, the “harder”function F is precisely SURJR,N .
We show that deg(SURJR,N ) = Θ(R1/4 ·N1/2).
If R = Θ(N), this is Θ(n3/4).
For convenience: let’s change the domain and range of allBoolean functions to 0, 1n and 0, 1.
![Page 51: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/51.jpg)
Idea of the Hardness-Amplifying Construction
Consider the function SURJECTIVITY : −1, 1n → −1, 1.Let n = N logR. SURJ interprets its input x as a list of Nnumbers (x1, . . . , xN ) from a range [R].SURJR,N (x) = −1 if and only if every element of the range[R] appears at least once in the list.
When we apply Main Theorem to f = ANDR, the “harder”function F is precisely SURJR,N .
We show that deg(SURJR,N ) = Θ(R1/4 ·N1/2).
If R = Θ(N), this is Θ(n3/4).
For convenience: let’s change the domain and range of allBoolean functions to 0, 1n and 0, 1.
![Page 52: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/52.jpg)
Resolving the Approximate Degree of SURJ
![Page 53: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/53.jpg)
The O(R1/4 ·N 1/2) Upper Bound For SURJ: First Try
Let’s start with how to achieve a (loose) bound of
deg(SURJR,N ) = O(R1/2 ·N1/2).
Let
yij =
1 if xj = i
0 otherwise
Then
SURJ(x)=ANDR(ORN (y1,1, . . . , y1,N ), . . . ,ORN (yR,1 . . . , yR,N )).
![Page 54: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/54.jpg)
The O(R1/4 ·N 1/2) Upper Bound For SURJ: First Try
Let’s start with how to achieve a (loose) bound of
deg(SURJR,N ) = O(R1/2 ·N1/2).
Let
yij =
1 if xj = i
0 otherwise
Then
SURJ(x)=ANDR(ORN (y1,1, . . . , y1,N ), . . . ,ORN (yR,1 . . . , yR,N )).
![Page 55: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/55.jpg)
SURJ Illustrated (R = 3, N = 6)
x1 x2 x3 x4 x5 x6
y11 y12 y13 y14 y15 y16 y21 y22 y23 y24 y25 y26 y31 y32 y33 y34 y35 y36
AND
OR OR OR
(Eachxjin[R])
![Page 56: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/56.jpg)
SURJ Illustrated (R = 3, N = 6)
0 1 0 1 0 0 1 0 1 0 0 0 0 0 0 0 1 1
AND
OR OR OR
2 1 2 1 3 3
![Page 57: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/57.jpg)
The Upper Bound For SURJ: First Try
Let’s start with how to achieve a (loose) bound of
deg(SURJR,N ) = O(R1/2 ·N1/2).
Let
yij =
1 if xj = i
0 otherwise
Then
SURJ(x)=ANDR(ORN (y1,1, . . . , y1,N ), . . . ,ORN (yR,1 . . . , yR,N )).
Let p be a degree O(R1/2 ·N1/2) polynomial approximatingANDR(ORN , . . . ,ORN ).Then p(y1,1, . . . , y1,N , . . . , yR,1, . . . , yR,N ) approximatesSURJ, with degree O(deg(p) · logR) = O(R1/2 ·N1/2 · logR).
![Page 58: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/58.jpg)
The Upper Bound For SURJ: First Try
Let’s start with how to achieve a (loose) bound of
deg(SURJR,N ) = O(R1/2 ·N1/2).
Let
yij =
1 if xj = i
0 otherwise
Then
SURJ(x)=ANDR(ORN (y1,1, . . . , y1,N ), . . . ,ORN (yR,1 . . . , yR,N )).
Let p be a degree O(R1/2 ·N1/2) polynomial approximatingANDR(ORN , . . . ,ORN ).Then p(y1,1, . . . , y1,N , . . . , yR,1, . . . , yR,N ) approximatesSURJ, with degree O(deg(p) · logR) = O(R1/2 ·N1/2 · logR).
![Page 59: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/59.jpg)
The Upper Bound For SURJ: Second Try
Fix R = N/2. We’ll show deg(SURJR,N ) = O(R1/4 ·N1/2).
We’ll want to think of polynomials as computing theprobability that a query algorithm outputs 1.
E.g., we can think of our “first try” as composing an queryalgorithm for computing ANDR with R copies of a queryalgorithm computing ORN .
We’ll approximate SURJ via a “two-stage” construction.
![Page 60: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/60.jpg)
Stage 1
Consider a query algorithm that samples O(n3/4) inputs.
Any range item appearing in the sample definitely hasfrequency at least 1, so we can just “remove it fromconsideration.”
Stage 2 just needs to determine whether all range items notappearing in the sample have frequency at least 1.
Let SURJunsamp be the function we need to compute in Stage2.
![Page 61: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/61.jpg)
Stage 1 Illustrated (R = 3, N = 6)
2 1 2 1 3 3
0 1 0 1 0 0 1 0 1 0 0 0 0 0 0 0 1 1
AND
OR OR OR
2 2
Sampleofsizen3/4
![Page 62: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/62.jpg)
Stage 1 Illustrated (R = 3, N = 6)
2 1 2 1 3 3
0 1 0 1 0 0 1 0 1 0 0 0 0 0 0 0 1 1
AND
OR OR
2 2
Sampleofsizen3/4
![Page 63: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/63.jpg)
Stage 2
Key observation: any range item with frequency larger thanT = n1/2 will appear in the sample at least once, withprobability 1− exp(−n1/4).
i.e., if a range item doesn’t appear in the sample, we are reallyconfident that it does not have a very high frequency.
So Stage 2 only needs an approximation p to SURJunsamp thatis accurate under the assumption that no range item hasfrequency higher than T .
If p is fed an input in which some range item has frequencyhigher than T , then p is allowed to be exponentially large onthat input.Specifically, if b unsampled range items have frequency largerthan T , then it is okay for |p(x)| to be as large as exp(n1/4 · b).
![Page 64: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/64.jpg)
Stage 2
Key observation: any range item with frequency larger thanT = n1/2 will appear in the sample at least once, withprobability 1− exp(−n1/4).
i.e., if a range item doesn’t appear in the sample, we are reallyconfident that it does not have a very high frequency.
So Stage 2 only needs an approximation p to SURJunsamp thatis accurate under the assumption that no range item hasfrequency higher than T .
If p is fed an input in which some range item has frequencyhigher than T , then p is allowed to be exponentially large onthat input.Specifically, if b unsampled range items have frequency largerthan T , then it is okay for |p(x)| to be as large as exp(n1/4 · b).
![Page 65: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/65.jpg)
Stage 2
Key observation: any range item with frequency larger thanT = n1/2 will appear in the sample at least once, withprobability 1− exp(−n1/4).
i.e., if a range item doesn’t appear in the sample, we are reallyconfident that it does not have a very high frequency.
So Stage 2 only needs an approximation p to SURJunsamp thatis accurate under the assumption that no range item hasfrequency higher than T .
If p is fed an input in which some range item has frequencyhigher than T , then p is allowed to be exponentially large onthat input.
Specifically, if b unsampled range items have frequency largerthan T , then it is okay for |p(x)| to be as large as exp(n1/4 · b).
![Page 66: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/66.jpg)
Stage 2
Key observation: any range item with frequency larger thanT = n1/2 will appear in the sample at least once, withprobability 1− exp(−n1/4).
i.e., if a range item doesn’t appear in the sample, we are reallyconfident that it does not have a very high frequency.
So Stage 2 only needs an approximation p to SURJunsamp thatis accurate under the assumption that no range item hasfrequency higher than T .
If p is fed an input in which some range item has frequencyhigher than T , then p is allowed to be exponentially large onthat input.Specifically, if b unsampled range items have frequency largerthan T , then it is okay for |p(x)| to be as large as exp(n1/4 · b).
![Page 67: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/67.jpg)
Stage 2 Illustrated (R = 3, N = 6)
2 1 2 1 3 3
0 1 0 1 0 0 1 0 1 0 0 0 0 0 0 0 1 1
AND
OR OR
2 2
Sampleofsizen3/4
OnlyapproximatetheremainingORgatesoninputsofHammingweightatmostn1/2.
![Page 68: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/68.jpg)
Stage 2 Details
Lemma (Chebyshev polynomials)
There is a polynomial q of degree O(n1/4) such that
|q(x)−ORn(x)| 1/n for all |x| ≤ n1/2.
|q(x)| ≤ exp(O(n1/4)
)otherwise.
Theorem
For x = (x1, . . . , xR), let b(x1, . . . , xR)=#i : |xi| > n1/2. Thereis a polynomial q of degree O(R1/2 ·N1/4) such that:
|q(x)−ANDR ORN (x)| ≤ 1/3 if b(x) = 0.
|p(x)| ≤ exp(O(b(x) · n1/4)
)otherwise.
Proof.
Let h approximate ANDR, and let p = h q.
![Page 69: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/69.jpg)
Lower Bound Analysis for SURJ
![Page 70: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/70.jpg)
Lower Bound Analysis for SURJ
Recall: to approximate SURJR,N , it is sufficient toapproximate the block-composed functionANDR(ORN , . . . ,ORN ) on N ·R bits, on inputs ofHamming weight exactly N .
![Page 71: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/71.jpg)
SURJ Illustrated (R = 3, N = 6)
x1 x2 x3 x4 x5 x6
y11 y12 y13 y14 y15 y16 y21 y22 y23 y24 y25 y26 y31 y32 y33 y34 y35 y36
AND
OR OR OR
(Eachxjin[R])
![Page 72: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/72.jpg)
Lower Bound Analysis for SURJ
Recall: to approximate SURJR,N , it is sufficient toapproximate the block-composed functionANDR(ORN , . . . ,ORN ) on N ·R bits, on inputs ofHamming weight exactly N .
Step 1: Show the converse.i.e., to approximate SURJ(x), it is necessary to approximateANDR(ORN , . . . ,ORN ), under the promise that the inputhas Hamming weight at most∗ N .
Follows from a symmetrization argument (Ambainis 2003).∗To get “at most N” rather than “equal to N”, we need tointroduce a dummy range item that is ignored by the function.
![Page 73: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/73.jpg)
Lower Bound Analysis for SURJ
Recall: to approximate SURJR,N , it is sufficient toapproximate the block-composed functionANDR(ORN , . . . ,ORN ) on N ·R bits, on inputs ofHamming weight exactly N .
Step 1: Show the converse.
i.e., to approximate SURJ(x), it is necessary to approximateANDR(ORN , . . . ,ORN ), under the promise that the inputhas Hamming weight at most∗ N .
Follows from a symmetrization argument (Ambainis 2003).∗To get “at most N” rather than “equal to N”, we need tointroduce a dummy range item that is ignored by the function.
![Page 74: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/74.jpg)
Lower Bound Analysis for SURJ
Recall: to approximate SURJR,N , it is sufficient toapproximate the block-composed functionANDR(ORN , . . . ,ORN ) on N ·R bits, on inputs ofHamming weight exactly N .
Step 1: Show the converse.i.e., to approximate SURJ(x), it is necessary to approximateANDR(ORN , . . . ,ORN ), under the promise that the inputhas Hamming weight at most∗ N .
Follows from a symmetrization argument (Ambainis 2003).∗To get “at most N” rather than “equal to N”, we need tointroduce a dummy range item that is ignored by the function.
![Page 75: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/75.jpg)
Lower Bound Analysis for SURJ
Recall: to approximate SURJR,N , it is sufficient toapproximate the block-composed functionANDR(ORN , . . . ,ORN ) on N ·R bits, on inputs ofHamming weight exactly N .
Step 1: Show the converse.i.e., to approximate SURJ(x), it is necessary to approximateANDR(ORN , . . . ,ORN ), under the promise that the inputhas Hamming weight at most∗ N .
Follows from a symmetrization argument (Ambainis 2003).∗To get “at most N” rather than “equal to N”, we need tointroduce a dummy range item that is ignored by the function.
![Page 76: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/76.jpg)
SURJ Illustrated (R = 3, N = 6)
x1 x2 x3 x4 x5 x6
y11 y12 y13 y14 y15 y16 y21 y22 y23 y24 y25 y26 y31 y32 y33 y34 y35 y36
AND
OR OR OR
(Eachxjin[R])
![Page 77: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/77.jpg)
Lower Bound Analysis for SURJ
Let n = N logR.
Recall: to approximate SURJ : −1, 1n → −1, 1, it issufficient to approximate the block-composed functionANDR(ORN , . . . ,ORN ) on N ·R bits, on inputs ofHamming weight exactly N .
Step 1: Show the converse.To approximate SURJ(x), it is necessary to approximateANDR(ORN , . . . ,ORN ), under the promise that the inputhas Hamming weight at most∗ N .
Follows from a symmetrization argument (Ambainis 2003).∗To get “at most N” rather than “equal to N”, we need tointroduce a dummy range item that is ignored by the function.
Step 2: Prove that for some N = O(R), this promise problemrequires degree & Ω(R3/4).
Builds on the “dual combining technique” used earlier toanalyze AND-ORn (with no promise).
![Page 78: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/78.jpg)
Lower Bound Analysis for SURJ
Let n = N logR.
Recall: to approximate SURJ : −1, 1n → −1, 1, it issufficient to approximate the block-composed functionANDR(ORN , . . . ,ORN ) on N ·R bits, on inputs ofHamming weight exactly N .
Step 1: Show the converse.To approximate SURJ(x), it is necessary to approximateANDR(ORN , . . . ,ORN ), under the promise that the inputhas Hamming weight at most∗ N .
Follows from a symmetrization argument (Ambainis 2003).∗To get “at most N” rather than “equal to N”, we need tointroduce a dummy range item that is ignored by the function.
Step 2: Prove that for some N = O(R), this promise problemrequires degree & Ω(R3/4).
Builds on the “dual combining technique” used earlier toanalyze AND-ORn (with no promise).
![Page 79: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/79.jpg)
Lower Bound Analysis for SURJ
Let n = N logR.
Recall: to approximate SURJ : −1, 1n → −1, 1, it issufficient to approximate the block-composed functionANDR(ORN , . . . ,ORN ) on N ·R bits, on inputs ofHamming weight exactly N .
Step 1: Show the converse.To approximate SURJ(x), it is necessary to approximateANDR(ORN , . . . ,ORN ), under the promise that the inputhas Hamming weight at most∗ N .
Follows from a symmetrization argument (Ambainis 2003).∗To get “at most N” rather than “equal to N”, we need tointroduce a dummy range item that is ignored by the function.
Step 2: Prove that for some N = O(R), this promise problemrequires degree & Ω(R3/4).
Builds on the “dual combining technique” used earlier toanalyze AND-ORn (with no promise).
![Page 80: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/80.jpg)
Overview of Step 2
Prove That For Some N = O(R), Approximating ANDR ORN
Under the Promise That The Input Has Hamming Weight AtMost N Requires Degree & R3/4.
![Page 81: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/81.jpg)
Attempt 1
For some N = O(R), want a dual witness forANDR(ORN , . . . ,ORN ) that only places mass on inputsof Hamming weight at most N .
Attempt 1: Use the dual witness for ANDR(ORN , . . . ,ORN )from prior work [She09, Lee09, BT13, She13].
ψAND-OR(y1, . . . , yR) := C · ψAND(. . . , sgn(ψOR(yj)), . . . )
R∏j=1
|ψOR(yj)|
(C chosen to ensure ψAND-OR has L1-norm 1).
Must verify:
1 ψAND-OR has pure high degree≥R1/2 ·N1/2 =Ω(N).
2 ψAND-OR well-correlated with AND-OR.
3 ψAND-OR places mass only on inputs of Hamming weight ≤ N .
ANDR
ORN ORN
y yR
![Page 82: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/82.jpg)
Attempt 1
For some N = O(R), want a dual witness forANDR(ORN , . . . ,ORN ) that only places mass on inputsof Hamming weight at most N .
Attempt 1: Use the dual witness for ANDR(ORN , . . . ,ORN )from prior work [She09, Lee09, BT13, She13].
ψAND-OR(y1, . . . , yR) := C · ψAND(. . . , sgn(ψOR(yj)), . . . )
R∏j=1
|ψOR(yj)|
(C chosen to ensure ψAND-OR has L1-norm 1).
Must verify:
1 ψAND-OR has pure high degree≥R1/2 ·N1/2 =Ω(N).
2 ψAND-OR well-correlated with AND-OR.
3 ψAND-OR places mass only on inputs of Hamming weight ≤ N .
ANDR
ORN ORN
y yR
![Page 83: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/83.jpg)
Attempt 1
For some N = O(R), want a dual witness forANDR(ORN , . . . ,ORN ) that only places mass on inputsof Hamming weight at most N .
Attempt 1: Use the dual witness for ANDR(ORN , . . . ,ORN )from prior work [She09, Lee09, BT13, She13].
ψAND-OR(y1, . . . , yR) := C · ψAND(. . . , sgn(ψOR(yj)), . . . )
R∏j=1
|ψOR(yj)|
(C chosen to ensure ψAND-OR has L1-norm 1).
Must verify:
1 ψAND-OR has pure high degree≥R1/2 ·N1/2 =Ω(N).
2 ψAND-OR well-correlated with AND-OR.
3 ψAND-OR places mass only on inputs of Hamming weight ≤ N .
ANDR
ORN ORN
y yR
![Page 84: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/84.jpg)
Attempt 1
For some N = O(R), want a dual witness forANDR(ORN , . . . ,ORN ) that only places mass on inputsof Hamming weight at most N .
Attempt 1: Use the dual witness for ANDR(ORN , . . . ,ORN )from prior work [She09, Lee09, BT13, She13].
ψAND-OR(y1, . . . , yR) := C · ψAND(. . . , sgn(ψOR(yj)), . . . )
R∏j=1
|ψOR(yj)|
(C chosen to ensure ψAND-OR has L1-norm 1).
Must verify:
1 ψAND-OR has pure high degree≥R1/2 ·N1/2 =Ω(N).X[She09]
2 ψAND-OR well-correlated with AND-OR.X[BT13, She13]
3 ψAND-OR places mass only on inputs of Hamming weight ≤ N .X
ANDR
ORN ORN
y yR
![Page 85: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/85.jpg)
Patching Attempt 1
Goal: Fix Property 3 without destroying Properties 1 or 2.
Fact (cf. Razborov and Sherstov 2008): Suppose∑|y|>N
|ψAND-OR(y)| R−D.
Then we can “post-process” ψAND-OR to “zero out” any massit places it inputs of Hamming weight larger than N .While ensuring that the resulting dual witness still has purehigh degree minD,PHD(ψAND-OR).
![Page 86: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/86.jpg)
Patching Attempt 1
Goal: Fix Property 3 without destroying Properties 1 or 2.
Fact (cf. Razborov and Sherstov 2008): Suppose∑|y|>N
|ψAND-OR(y)| R−D.
Then we can “post-process” ψAND-OR to “zero out” any massit places it inputs of Hamming weight larger than N .While ensuring that the resulting dual witness still has purehigh degree minD,PHD(ψAND-OR).
![Page 87: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/87.jpg)
Patching Attempt 1
New Goal: Show that, for D ≈ R3/4,∑|y|>N
|ψAND-OR(y)| R−D. (1)
Recall:
ψAND-OR(y1, . . . , yR) := C · ψAND(. . . , sgn(ψOR(yj)), . . . )
R∏j=1
|ψOR(yj)|
A dual witness ψOR for OR can be made “weakly” biasedtoward low Hamming weight inputs.
Specifically, can ensure:PHD(ψOR) ≥ n1/4.
For all t,∑|yi|=t |ψOR(yi)| ≤ t−2 · exp(−t/n1/4). (2)
|ψAND-OR(y1,. . ., yR)| resembles product distribution:∏Rj=1|ψOR(yj)|
So it is exponentially more biased toward low Hamming weightinputs than ψOR itself.
Intuition: By (2): the mass that∏Rj=1|ψOR(yj)| places on inputs of
Hamming weight > N is dominated by inputs with |yi| = N1/4 forat least N3/4 values of i.Also by (2), each |yi| = N1/4 contributes a factor of 1/poly(N).So total mass on these inputs is exp(−Ω(N3/4)).
![Page 88: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/88.jpg)
Patching Attempt 1
New Goal: Show that, for D ≈ R3/4,∑|y|>N
|ψAND-OR(y)| R−D. (1)
Recall:
ψAND-OR(y1, . . . , yR) := C · ψAND(. . . , sgn(ψOR(yj)), . . . )
R∏j=1
|ψOR(yj)|
A dual witness ψOR for OR can be made “weakly” biasedtoward low Hamming weight inputs.
Specifically, can ensure:PHD(ψOR) ≥ n1/4.
For all t,∑|yi|=t |ψOR(yi)| ≤ t−2 · exp(−t/n1/4). (2)
|ψAND-OR(y1,. . ., yR)| resembles product distribution:∏Rj=1|ψOR(yj)|
So it is exponentially more biased toward low Hamming weightinputs than ψOR itself.
Intuition: By (2): the mass that∏Rj=1|ψOR(yj)| places on inputs of
Hamming weight > N is dominated by inputs with |yi| = N1/4 forat least N3/4 values of i.Also by (2), each |yi| = N1/4 contributes a factor of 1/poly(N).So total mass on these inputs is exp(−Ω(N3/4)).
![Page 89: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/89.jpg)
Patching Attempt 1
New Goal: Show that, for D ≈ R3/4,∑|y|>N
|ψAND-OR(y)| R−D. (1)
Recall:
ψAND-OR(y1, . . . , yR) := C · ψAND(. . . , sgn(ψOR(yj)), . . . )
R∏j=1
|ψOR(yj)|
A dual witness ψOR for OR can be made “weakly” biasedtoward low Hamming weight inputs.
Specifically, can ensure:PHD(ψOR) ≥ n1/4.
For all t,∑|yi|=t |ψOR(yi)| ≤ t−2 · exp(−t/n1/4). (2)
|ψAND-OR(y1,. . ., yR)| resembles product distribution:∏Rj=1|ψOR(yj)|
So it is exponentially more biased toward low Hamming weightinputs than ψOR itself.
Intuition: By (2): the mass that∏Rj=1|ψOR(yj)| places on inputs of
Hamming weight > N is dominated by inputs with |yi| = N1/4 forat least N3/4 values of i.Also by (2), each |yi| = N1/4 contributes a factor of 1/poly(N).So total mass on these inputs is exp(−Ω(N3/4)).
![Page 90: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/90.jpg)
Patching Attempt 1
New Goal: Show that, for D ≈ R3/4,∑|y|>N
|ψAND-OR(y)| R−D. (1)
Recall:
ψAND-OR(y1, . . . , yR) := C · ψAND(. . . , sgn(ψOR(yj)), . . . )
R∏j=1
|ψOR(yj)|
A dual witness ψOR for OR can be made “weakly” biasedtoward low Hamming weight inputs.
Specifically, can ensure:PHD(ψOR) ≥ n1/4.
For all t,∑|yi|=t |ψOR(yi)| ≤ t−2 · exp(−t/n1/4). (2)
|ψAND-OR(y1,. . ., yR)| resembles product distribution:∏Rj=1|ψOR(yj)|
So it is exponentially more biased toward low Hamming weightinputs than ψOR itself.
Intuition: By (2): the mass that∏Rj=1|ψOR(yj)| places on inputs of
Hamming weight > N is dominated by inputs with |yi| = N1/4 forat least N3/4 values of i.Also by (2), each |yi| = N1/4 contributes a factor of 1/poly(N).
So total mass on these inputs is exp(−Ω(N3/4)).
![Page 91: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/91.jpg)
Patching Attempt 1
New Goal: Show that, for D ≈ R3/4,∑|y|>N
|ψAND-OR(y)| R−D. (1)
Recall:
ψAND-OR(y1, . . . , yR) := C · ψAND(. . . , sgn(ψOR(yj)), . . . )
R∏j=1
|ψOR(yj)|
A dual witness ψOR for OR can be made “weakly” biasedtoward low Hamming weight inputs.
Specifically, can ensure:PHD(ψOR) ≥ n1/4.
For all t,∑|yi|=t |ψOR(yi)| ≤ t−2 · exp(−t/n1/4). (2)
|ψAND-OR(y1,. . ., yR)| resembles product distribution:∏Rj=1|ψOR(yj)|
So it is exponentially more biased toward low Hamming weightinputs than ψOR itself.
Intuition: By (2): the mass that∏Rj=1|ψOR(yj)| places on inputs of
Hamming weight > N is dominated by inputs with |yi| = N1/4 forat least N3/4 values of i.
Also by (2), each |yi| = N1/4 contributes a factor of 1/poly(N).
So total mass on these inputs is exp(−Ω(N3/4)).
![Page 92: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/92.jpg)
General Hardness Amplification Within AC0
![Page 93: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/93.jpg)
General Hardness Amplification Within AC0
Recall: When we apply our hardness amplification tof = ANDR, the “harder” function F is precisely SURJ.
For a general function f , what is the “harder” function F?
![Page 94: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/94.jpg)
First Attempt: Amplifying Hardness off:−1, 1R→−1, 1 (R=3,N=6)
2 1 2 1 3 3
0 1 0 1 0 0 1 0 1 0 0 0 0 0 0 0 1 1
f
OR OR OR
![Page 95: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/95.jpg)
Hardness-Amplifying Construction: Second Attempt
First attempt at handling general f fails when f = OR.
F (x) = ORR(ORN (y1,1, . . . , y1,N ), . . . ,ORN (yR,1 . . . , yR,N ))has (exact) degree 0.
Let R′ = R logR. For f : −1, 1R → −1, 1, the real∗
definition of F is:
F (x)=(fANDlogR)(ORN (y1,1,. . . ,y1,N ),. . . ,ORN (yR′,1,. . ., yR′,N ))
∗This is still a slight simplification. Also need to introduce a dummy
range item that is ignored by F .
![Page 96: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/96.jpg)
Hardness-Amplifying Construction: Second Attempt
First attempt at handling general f fails when f = OR.
F (x) = ORR(ORN (y1,1, . . . , y1,N ), . . . ,ORN (yR,1 . . . , yR,N ))has (exact) degree 0.
Let R′ = R logR. For f : −1, 1R → −1, 1, the real∗
definition of F is:
F (x)=(fANDlogR)(ORN (y1,1,. . . ,y1,N ),. . . ,ORN (yR′,1,. . ., yR′,N ))
∗This is still a slight simplification. Also need to introduce a dummy
range item that is ignored by F .
![Page 97: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/97.jpg)
Future Directions
Resolve the quantum query complexity of k-distinctness,counting triangles, graph collision, etc.
Prove an Ω(nk/(k+1)) lower bound on approximate degree ofthe k-sum function?
Its quantum query complexity is known to be Θ(nk/(k+1)).
An Ω(n) lower bound on the approximate degree of AC0?
A sublinear upper bound for DNFs of polynomial size? Oreven polynomial size AC0 circuits?
Either result would yield new circuit lower bounds (namely, forAC0 MOD2 circuits).
Extend our bounds on degε(f) from ε = 1/3 to ε much closerto 1.
We believe our techniques can extend to give a Ω(n1−δ) lowerbound on the threshold degree of AC0.
![Page 98: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/98.jpg)
Thank you!
![Page 99: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/99.jpg)
Analysis of the Dual Witness for the AND-OR Tree
1/2
1/2
1/2
![Page 100: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/100.jpg)
The Combining Method [SZ09, She09, Lee09]
ψAND-OR(x1, . . . , xn1/2) := C · ψOUT(. . . , sgn(ψIN(xi)), . . . )
n1/2∏i=1
|ψIN(xi)|
(C chosen to ensure ψAND-OR has L1-norm 1).
Must verify:
1 ψAND-OR has pure high degree ≥ n1/4 · n1/4 = n1/2.
2 ψAND-OR has high correlation with AND-OR.
1/2
1/2
1/2
![Page 101: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/101.jpg)
The Combining Method [SZ09, She09, Lee09]
ψAND-OR(x1, . . . , xn1/2) := C · ψOUT(. . . , sgn(ψIN(xi)), . . . )
n1/2∏i=1
|ψIN(xi)|
(C chosen to ensure ψAND-OR has L1-norm 1).
Must verify:
1 ψAND-OR has pure high degree ≥ n1/4 · n1/4 = n1/2.
2 ψAND-OR has high correlation with AND-OR.
1/2
1/2
1/2
![Page 102: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/102.jpg)
The Combining Method [SZ09, She09, Lee09]
ψAND-OR(x1, . . . , xn1/2) := C · ψOUT(. . . , sgn(ψIN(xi)), . . . )
n1/2∏i=1
|ψIN(xi)|
(C chosen to ensure ψAND-OR has L1-norm 1).
Must verify:
1 ψAND-OR has pure high degree ≥ n1/4 · n1/4 = n1/2.X[She09]
2 ψAND-OR has high correlation with AND-OR. [BT13, She13]
1/2
1/2
1/2
![Page 103: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/103.jpg)
Pure High Degree Analysis [She09]
ψAND-OR(x1, . . . , xn1/2) := C · ψOUT(. . . , sgn(ψIN(xi)), . . . )
n1/2∏i=1
|ψIN(xi)|
Intuition: Consider ψOUT(y1, y2, y3) = y1y2. ThenψAND-OR(x1, x2, x3) equals:
C · sgn(ψIN(x1)) · sgn(ψIN(x2)) ·3∏i=1
|ψIN(xi)|
= ψIN(x1) · ψIN(x2) · |ψIN(x3)|
Each term of ψAND-OR is the product of PHD(ψOUT)polynomials over disjoint variable sets, each of pure highdegree at least PHD(ψIN).
So ψAND-OR has pure high degree at leastPHD(ψOUT) · PHD(ψIN).
![Page 104: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/104.jpg)
Pure High Degree Analysis [She09]
ψAND-OR(x1, . . . , xn1/2) := C · ψOUT(. . . , sgn(ψIN(xi)), . . . )
n1/2∏i=1
|ψIN(xi)|
Intuition: Consider ψOUT(y1, y2, y3) = y1y2. ThenψAND-OR(x1, x2, x3) equals:
C · sgn(ψIN(x1)) · sgn(ψIN(x2)) ·3∏i=1
|ψIN(xi)|
= ψIN(x1) · ψIN(x2) · |ψIN(x3)|
Each term of ψAND-OR is the product of PHD(ψOUT)polynomials over disjoint variable sets, each of pure highdegree at least PHD(ψIN).
So ψAND-OR has pure high degree at leastPHD(ψOUT) · PHD(ψIN).
![Page 105: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/105.jpg)
(Sub)Goal: Show ψAND-OR has high correlation withAND-OR
![Page 106: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/106.jpg)
Correlation Analysis
ψAND-OR(x1, . . . , xn1/2) := C · ψOUT(. . . , sgn(ψIN(xi)), . . . )
n1/2∏i=1
|ψIN(xi)|
Idea: Show∑x∈−1,1n
ψAND-OR(x) ·AND-ORn(x) ≈∑
y∈−1,1n1/2ψOUT(y) ·ANDn1/2(y).
Intuition: We are feeding sgn(ψIN(xi)) into ψOUT.
ψIN is correlated with ORn1/2 , so sgn(ψIN(xi)) is a “decentpredictor” of ORn1/2 .
But there are errors. Need to show errors don’t “build up”.
![Page 107: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/107.jpg)
Correlation Analysis
Goal: Show∑x∈−1,1n
ψAND-OR(x) ·AND-ORn(x) ≈∑
y∈−1,1n1/2ψOUT(y) ·ANDn1/2(y).
Case 1: Consider any y = (sgnψIN(x1), . . . , sgnψIN(xn1/2)) 6=All-True.
There is some coordinate of y that equals FALSE. Only needto “trust” this coordinate to force AND-ORn to evaluate toFALSE on (x1, . . . , xn1/2). So errors do not build up!
![Page 108: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/108.jpg)
Correlation Analysis
Case 2: Consider y = All-True.
ANDn1/2(y)=AND-ORn(x1, . . . , xn1/2) only if allcoordinates of y are “error-free”.
Fortunately, ψIN has a special one-sided error property:
If sgn(ψIN(xi)) = −1, then ORn1/2(xi) is guaranteed toequal -1.
![Page 109: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/109.jpg)
Summary of Correlation Analysis
Two Cases.
In first case (feeding at least one FALSE into ψOUT), errorsdid not build up, because we only needed to “trust” theFALSE value.
In second case (all values fed into ψOUT are TRUE), weneeded to trust all values. But we could do this because ψIN
had one-sided error.
![Page 110: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/110.jpg)
One-Sided Approximate Degree
A real polynomial p is a one-sided ε-approximation for f if
|p(x)− 1| < ε ∀x ∈ f−1(−1)
p(x) ≥ 1 ∀x ∈ f−1(1)
odeg−,ε(f) = min degree of a one-sided ε-approximation for f .
odeg−(f) := odeg−,1/3(f) is the one-sided approximate degreeof f .
![Page 111: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/111.jpg)
Dual Formulation of odeg−
Primal LP (Linear in ε and coefficients of p):
minp,ε ε
s.t. |p(x)− 1| ≤ ε for all x ∈ f−1(−1)
p(x) ≥ 1 for all x ∈ f−1(1)
deg p ≤ d
Dual LP:
maxψ∑
x∈−1,1nψ(x)f(x)
s.t.∑
x∈−1,1n|ψ(x)| = 1
∑x∈−1,1n
ψ(x)q(x) = 0 whenever deg q ≤ d
ψ(x) ≥ 0 ∀x ∈ f−1(1)
![Page 112: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/112.jpg)
Proof that odeg−(ANDn) = Ω(√n)
We argued that the symmetrization of any 1/3-approximation toANDn had to look like this:
!"#$%&'()*+*&',*
![Page 113: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/113.jpg)
Andris Ambainis, Aleksandrs Belovs, Oded Regev, and Ronaldde Wolf.Efficient quantum algorithms for (gapped) group testing andjunta testing.In Proceedings of the Twenty-Seventh Annual ACM-SIAMSymposium on Discrete Algorithms, pages 903–922. Societyfor Industrial and Applied Mathematics, 2016.
Scott Aaronson and Yaoyun Shi.Quantum lower bounds for the collision and the elementdistinctness problems.J. ACM, 51(4):595–605, 2004.
Robert Beals, Harry Buhrman, Richard Cleve, Michele Mosca,and Ronald de Wolf.Quantum lower bounds by polynomials.J. ACM, 48(4):778–797, 2001.
![Page 114: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/114.jpg)
Aleksandrs Belovs.Learning-graph-based quantum algorithm for k-distinctness.In Foundations of Computer Science (FOCS), 2012 IEEE 53rdAnnual Symposium on, pages 207–216. IEEE, 2012.
Sergey Bravyi, Aram Wettroth Harrow, and AvinatanHassidim.Quantum algorithms for testing properties of distributions.IEEE Trans. Information Theory, 57(6):3971–3981, 2011.
Harry Buhrman, Ilan Newman, Hein Rohrig, and Ronaldde Wolf.Robust polynomials and quantum algorithms.Theory Comput. Syst., 40(4):379–395, 2007.
Troy Lee, Rajat Mittal, Ben W. Reichardt, Robert Spalek, andMario Szegedy.
![Page 115: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/115.jpg)
Quantum query complexity of state conversion.In Proceedings of the 52nd Symposium on Foundations ofComputer Science (FOCS 2011), pages 344–353, 2011.
Tongyang Li and Xiaodi Wu.Quantum query complexity of entropy estimation.arXiv preprint arXiv:1710.06025, 2017.
Ben W Reichardt.Reflections for quantum query algorithms.In Proceedings of the twenty-second annual ACM-SIAMsymposium on Discrete Algorithms, pages 560–569. Society forIndustrial and Applied Mathematics, 2011.
Alexander A. Sherstov.The pattern matrix method.SIAM J. Comput., 40(6):1969–2000, 2011.Preliminary version in STOC 2008.
![Page 116: The Polynomial Method Strikes Back: Tight Quantum Query ...people.cs.georgetown.edu/jthaler/kdistslides.pdf · Lower Bound Methods in Quantum Query Complexity Since 2002, the positive-weights](https://reader034.vdocument.in/reader034/viewer/2022051912/60037caa77b4577e6807af32/html5/thumbnails/116.jpg)
Alexander A. Sherstov.Approximating the AND-OR Tree.Theory of Computing, 9(20):653–663, 2013.
Alexander A. Sherstov, 2017.Personal communication.