![Page 1: Kernel-Size Lower Bounds: The Evidence from …people.csail.mit.edu/andyd/Kernel_LBs_2.pdfAndrew Drucker Kernel-Size Lower Bounds Summary Fortnow-Santhanam technique applies to randomized](https://reader034.vdocuments.net/reader034/viewer/2022042922/5f6d79cbd438df27a00711ca/html5/thumbnails/1.jpg)
Kernel-Size Lower Bounds:The Evidence from Complexity Theory
Andrew Drucker
IAS
Worker 2013, Warsaw
Andrew Drucker Kernel-Size Lower Bounds
![Page 2: Kernel-Size Lower Bounds: The Evidence from …people.csail.mit.edu/andyd/Kernel_LBs_2.pdfAndrew Drucker Kernel-Size Lower Bounds Summary Fortnow-Santhanam technique applies to randomized](https://reader034.vdocuments.net/reader034/viewer/2022042922/5f6d79cbd438df27a00711ca/html5/thumbnails/2.jpg)
Part 2/3
Andrew Drucker Kernel-Size Lower Bounds
![Page 3: Kernel-Size Lower Bounds: The Evidence from …people.csail.mit.edu/andyd/Kernel_LBs_2.pdfAndrew Drucker Kernel-Size Lower Bounds Summary Fortnow-Santhanam technique applies to randomized](https://reader034.vdocuments.net/reader034/viewer/2022042922/5f6d79cbd438df27a00711ca/html5/thumbnails/3.jpg)
Note
These slides are a slightly revised version of a 3-part tutorial givenat the 2013 Workshop on Kernelization (“Worker”) at theUniversity of Warsaw. Thanks to the organizers for the opportunityto present!
Preparation of this teaching material was supported by the National ScienceFoundation under agreements Princeton University Prime Award No.CCF-0832797 and Sub-contract No. 00001583. Any opinions, findings andconclusions or recommendations expressed in this material are those of theauthor(s) and do not necessarily reflect the views of the National ScienceFoundation.
Andrew Drucker Kernel-Size Lower Bounds
![Page 4: Kernel-Size Lower Bounds: The Evidence from …people.csail.mit.edu/andyd/Kernel_LBs_2.pdfAndrew Drucker Kernel-Size Lower Bounds Summary Fortnow-Santhanam technique applies to randomized](https://reader034.vdocuments.net/reader034/viewer/2022042922/5f6d79cbd438df27a00711ca/html5/thumbnails/4.jpg)
Outline
1 Introduction
2 OR/AND-conjectures and their use
3 Evidence for the conjectures
Andrew Drucker Kernel-Size Lower Bounds
![Page 5: Kernel-Size Lower Bounds: The Evidence from …people.csail.mit.edu/andyd/Kernel_LBs_2.pdfAndrew Drucker Kernel-Size Lower Bounds Summary Fortnow-Santhanam technique applies to randomized](https://reader034.vdocuments.net/reader034/viewer/2022042922/5f6d79cbd438df27a00711ca/html5/thumbnails/5.jpg)
Outline
1 Introduction
2 OR/AND-conjectures and their use
3 Evidence for the conjectures
Andrew Drucker Kernel-Size Lower Bounds
![Page 6: Kernel-Size Lower Bounds: The Evidence from …people.csail.mit.edu/andyd/Kernel_LBs_2.pdfAndrew Drucker Kernel-Size Lower Bounds Summary Fortnow-Santhanam technique applies to randomized](https://reader034.vdocuments.net/reader034/viewer/2022042922/5f6d79cbd438df27a00711ca/html5/thumbnails/6.jpg)
To be proved
Evidence for the OR, AND conjectures:
Theorem
Assume NP * coNP/poly. If L is NP-complete, t(k) ≤ poly(k),
1 [Fortnow-Santhanam’08] No deterministic poly-time reductionR from OR=(L)t(·) to any problem can have output size
|R(x)| ≤ O(t log t) .
2 [D.’12] No probabilistic poly-time reduction R from
OR=(L)t(·) , AND=(L)t(·)
to any problem, with Pr[success] ≥ .99], can achieve
|R(x)| ≤ t .
Andrew Drucker Kernel-Size Lower Bounds
![Page 7: Kernel-Size Lower Bounds: The Evidence from …people.csail.mit.edu/andyd/Kernel_LBs_2.pdfAndrew Drucker Kernel-Size Lower Bounds Summary Fortnow-Santhanam technique applies to randomized](https://reader034.vdocuments.net/reader034/viewer/2022042922/5f6d79cbd438df27a00711ca/html5/thumbnails/7.jpg)
Background
Let’s back up and discuss:
What does NP * coNP/poly mean?
Why believe it?
Andrew Drucker Kernel-Size Lower Bounds
![Page 8: Kernel-Size Lower Bounds: The Evidence from …people.csail.mit.edu/andyd/Kernel_LBs_2.pdfAndrew Drucker Kernel-Size Lower Bounds Summary Fortnow-Santhanam technique applies to randomized](https://reader034.vdocuments.net/reader034/viewer/2022042922/5f6d79cbd438df27a00711ca/html5/thumbnails/8.jpg)
Background: circuits
We use ordinary model of Boolean circuits: ∧,∨,¬ gates,bounded fanin.
Say that decision problem L has poly-size circuits, and writeL ∈ P/poly, if
∃ { Cn : {0, 1}n → {0, 1} }n>0 :
size(Cn) ≤ poly(n), Cn(x) ≡ L(x) .
Non-uniform complexity class: def’n of Cn may dependuncomputably on n!
Example: if L ⊆ 1∗, then L ∈ P/poly. Also, BPP ⊂ P/poly.
Andrew Drucker Kernel-Size Lower Bounds
![Page 9: Kernel-Size Lower Bounds: The Evidence from …people.csail.mit.edu/andyd/Kernel_LBs_2.pdfAndrew Drucker Kernel-Size Lower Bounds Summary Fortnow-Santhanam technique applies to randomized](https://reader034.vdocuments.net/reader034/viewer/2022042922/5f6d79cbd438df27a00711ca/html5/thumbnails/9.jpg)
Background: nondeterminism
Recall: decision problem L is in NP if:
∃ poly-time algorithm A(x , y) on n + poly(n) input bits :
x ∈ L ⇐⇒ ∃ y : A(x , y) = 1 .
“Non-uniform NP”
Andrew Drucker Kernel-Size Lower Bounds
![Page 10: Kernel-Size Lower Bounds: The Evidence from …people.csail.mit.edu/andyd/Kernel_LBs_2.pdfAndrew Drucker Kernel-Size Lower Bounds Summary Fortnow-Santhanam technique applies to randomized](https://reader034.vdocuments.net/reader034/viewer/2022042922/5f6d79cbd438df27a00711ca/html5/thumbnails/10.jpg)
Background: nondeterminism
Say that decision problem L is in NP/poly if:
∃ poly-sized ckts {Cn(x , y)}n on n + poly(n) input bits :
x ∈ Ln ⇐⇒ ∃ y : Cn(x , y) = 1 .
“Non-uniform NP”
Andrew Drucker Kernel-Size Lower Bounds
![Page 11: Kernel-Size Lower Bounds: The Evidence from …people.csail.mit.edu/andyd/Kernel_LBs_2.pdfAndrew Drucker Kernel-Size Lower Bounds Summary Fortnow-Santhanam technique applies to randomized](https://reader034.vdocuments.net/reader034/viewer/2022042922/5f6d79cbd438df27a00711ca/html5/thumbnails/11.jpg)
Background: nondeterminism
Recall that coNP = {L : L ∈ NP}.
Complete problem: UNSAT = {ψ : ψ is unsatisfiable}.
coNP/poly = {L : L ∈ NP/poly}.
Andrew Drucker Kernel-Size Lower Bounds
![Page 12: Kernel-Size Lower Bounds: The Evidence from …people.csail.mit.edu/andyd/Kernel_LBs_2.pdfAndrew Drucker Kernel-Size Lower Bounds Summary Fortnow-Santhanam technique applies to randomized](https://reader034.vdocuments.net/reader034/viewer/2022042922/5f6d79cbd438df27a00711ca/html5/thumbnails/12.jpg)
Uniform and nonuniform complexity
Connect questions about non-uniform computation to uniformquestions?
Yes!
Need a broader view of nondeterminism...
Andrew Drucker Kernel-Size Lower Bounds
![Page 13: Kernel-Size Lower Bounds: The Evidence from …people.csail.mit.edu/andyd/Kernel_LBs_2.pdfAndrew Drucker Kernel-Size Lower Bounds Summary Fortnow-Santhanam technique applies to randomized](https://reader034.vdocuments.net/reader034/viewer/2022042922/5f6d79cbd438df27a00711ca/html5/thumbnails/13.jpg)
Games and computation
Given a circuit C (y1, y2, . . . , yk) with k input blocks, consider2-player game where Player 1 wants C → 1, P0 wants C → 0.
Take turns setting y1, . . . , yk .
Andrew Drucker Kernel-Size Lower Bounds
![Page 14: Kernel-Size Lower Bounds: The Evidence from …people.csail.mit.edu/andyd/Kernel_LBs_2.pdfAndrew Drucker Kernel-Size Lower Bounds Summary Fortnow-Santhanam technique applies to randomized](https://reader034.vdocuments.net/reader034/viewer/2022042922/5f6d79cbd438df27a00711ca/html5/thumbnails/14.jpg)
Games and computation
Andrew Drucker Kernel-Size Lower Bounds
![Page 15: Kernel-Size Lower Bounds: The Evidence from …people.csail.mit.edu/andyd/Kernel_LBs_2.pdfAndrew Drucker Kernel-Size Lower Bounds Summary Fortnow-Santhanam technique applies to randomized](https://reader034.vdocuments.net/reader034/viewer/2022042922/5f6d79cbd438df27a00711ca/html5/thumbnails/15.jpg)
Games and computation
Andrew Drucker Kernel-Size Lower Bounds
![Page 16: Kernel-Size Lower Bounds: The Evidence from …people.csail.mit.edu/andyd/Kernel_LBs_2.pdfAndrew Drucker Kernel-Size Lower Bounds Summary Fortnow-Santhanam technique applies to randomized](https://reader034.vdocuments.net/reader034/viewer/2022042922/5f6d79cbd438df27a00711ca/html5/thumbnails/16.jpg)
Games and computation
Andrew Drucker Kernel-Size Lower Bounds
![Page 17: Kernel-Size Lower Bounds: The Evidence from …people.csail.mit.edu/andyd/Kernel_LBs_2.pdfAndrew Drucker Kernel-Size Lower Bounds Summary Fortnow-Santhanam technique applies to randomized](https://reader034.vdocuments.net/reader034/viewer/2022042922/5f6d79cbd438df27a00711ca/html5/thumbnails/17.jpg)
Games and computation
Andrew Drucker Kernel-Size Lower Bounds
![Page 18: Kernel-Size Lower Bounds: The Evidence from …people.csail.mit.edu/andyd/Kernel_LBs_2.pdfAndrew Drucker Kernel-Size Lower Bounds Summary Fortnow-Santhanam technique applies to randomized](https://reader034.vdocuments.net/reader034/viewer/2022042922/5f6d79cbd438df27a00711ca/html5/thumbnails/18.jpg)
Games and computation
Andrew Drucker Kernel-Size Lower Bounds
![Page 19: Kernel-Size Lower Bounds: The Evidence from …people.csail.mit.edu/andyd/Kernel_LBs_2.pdfAndrew Drucker Kernel-Size Lower Bounds Summary Fortnow-Santhanam technique applies to randomized](https://reader034.vdocuments.net/reader034/viewer/2022042922/5f6d79cbd438df27a00711ca/html5/thumbnails/19.jpg)
Games and computation
Define d-ROUND GAME (∃) as:
Input: a d-block circuit C (y1, . . . , yd).
Decide: on 2-player game where P1 goes first, can P1 force awin? (C = 1)
Define complexity classΣpd
as set of languages poly-time (Karp)-reducible to d-ROUNDGAME (∃).
“d th level of Polynomial Hierarchy”
Andrew Drucker Kernel-Size Lower Bounds
![Page 20: Kernel-Size Lower Bounds: The Evidence from …people.csail.mit.edu/andyd/Kernel_LBs_2.pdfAndrew Drucker Kernel-Size Lower Bounds Summary Fortnow-Santhanam technique applies to randomized](https://reader034.vdocuments.net/reader034/viewer/2022042922/5f6d79cbd438df27a00711ca/html5/thumbnails/20.jpg)
Games and computation
Facts: NP = Σp1 (“solitaire”); Σp
d ⊆ Σpd+1.
Common conjecture: for all d > 0, Σpd 6= Σp
d+1.
Otherwise we could efficiently reduce a (d + 1)-round game toan equivalent d-round one, and
how the heck do you do that??
Andrew Drucker Kernel-Size Lower Bounds
![Page 21: Kernel-Size Lower Bounds: The Evidence from …people.csail.mit.edu/andyd/Kernel_LBs_2.pdfAndrew Drucker Kernel-Size Lower Bounds Summary Fortnow-Santhanam technique applies to randomized](https://reader034.vdocuments.net/reader034/viewer/2022042922/5f6d79cbd438df27a00711ca/html5/thumbnails/21.jpg)
Games and computation
Games allow us to connect uniform and non-uniformcomplexity questions:
Theorem (Karp-Lipton ’82)
Suppose NP is in P/poly.Then, for all d > 2,
Σpd = Σp
2 .
Andrew Drucker Kernel-Size Lower Bounds
![Page 22: Kernel-Size Lower Bounds: The Evidence from …people.csail.mit.edu/andyd/Kernel_LBs_2.pdfAndrew Drucker Kernel-Size Lower Bounds Summary Fortnow-Santhanam technique applies to randomized](https://reader034.vdocuments.net/reader034/viewer/2022042922/5f6d79cbd438df27a00711ca/html5/thumbnails/22.jpg)
Games and computation
Games allow us to connect uniform and non-uniformcomplexity questions:
Theorem (Yap ’83)
Suppose NP is in coNP/poly.Then, for all d > 3,
Σpd = Σp
3 .
Andrew Drucker Kernel-Size Lower Bounds
![Page 23: Kernel-Size Lower Bounds: The Evidence from …people.csail.mit.edu/andyd/Kernel_LBs_2.pdfAndrew Drucker Kernel-Size Lower Bounds Summary Fortnow-Santhanam technique applies to randomized](https://reader034.vdocuments.net/reader034/viewer/2022042922/5f6d79cbd438df27a00711ca/html5/thumbnails/23.jpg)
Games and computation
So: the assumption
NP * coNP/poly
can be based on an (easy-to-state, likely) assumption:
“One cannot efficiently reduce a 100-round gameto an equivalent 3-round game!”
Andrew Drucker Kernel-Size Lower Bounds
![Page 24: Kernel-Size Lower Bounds: The Evidence from …people.csail.mit.edu/andyd/Kernel_LBs_2.pdfAndrew Drucker Kernel-Size Lower Bounds Summary Fortnow-Santhanam technique applies to randomized](https://reader034.vdocuments.net/reader034/viewer/2022042922/5f6d79cbd438df27a00711ca/html5/thumbnails/24.jpg)
The minimax theorem
An extremely useful tool.
Many applications in complexity theory, beginning with[Yao’77].
Gives alternate (but similar) proof of [Fortnow-Santhanam’08]result;
seems crucial for best results in [D.’12].
Andrew Drucker Kernel-Size Lower Bounds
![Page 25: Kernel-Size Lower Bounds: The Evidence from …people.csail.mit.edu/andyd/Kernel_LBs_2.pdfAndrew Drucker Kernel-Size Lower Bounds Summary Fortnow-Santhanam technique applies to randomized](https://reader034.vdocuments.net/reader034/viewer/2022042922/5f6d79cbd438df27a00711ca/html5/thumbnails/25.jpg)
The minimax theorem
Setting: a 2-player, simultaneous-move, zero-sum game.
Players 1, 2 have finite sets X ,Y . (“possible moves”)
“Payoff function” Val(x , y) : X × Y → [0, 1].
Val(x , y) defines “payoff from P1 to P2,” given moves(x , y).
(P1 trying to minimize Val(x , y), P2 trying to maximize)
Andrew Drucker Kernel-Size Lower Bounds
![Page 26: Kernel-Size Lower Bounds: The Evidence from …people.csail.mit.edu/andyd/Kernel_LBs_2.pdfAndrew Drucker Kernel-Size Lower Bounds Summary Fortnow-Santhanam technique applies to randomized](https://reader034.vdocuments.net/reader034/viewer/2022042922/5f6d79cbd438df27a00711ca/html5/thumbnails/26.jpg)
The minimax theorem
Mixed strategy for P1: A distribution DX over X .
(Mixed strategies can be useful...)
Minimax thm says: for P1 to do well against all P2strategies...
it’s enough if P1 can do well against any fixed mixed strategy.
Andrew Drucker Kernel-Size Lower Bounds
![Page 27: Kernel-Size Lower Bounds: The Evidence from …people.csail.mit.edu/andyd/Kernel_LBs_2.pdfAndrew Drucker Kernel-Size Lower Bounds Summary Fortnow-Santhanam technique applies to randomized](https://reader034.vdocuments.net/reader034/viewer/2022042922/5f6d79cbd438df27a00711ca/html5/thumbnails/27.jpg)
The minimax theorem
Theorem (Minimax—Von Neumann)
Suppose that for every mixed strategy DY for P2, there is a P1move x ∈ X such that
Ey∼DY[Val(x , y)] ≤ α .
Then, there is a P1 mixed strategy D∗X such that, for all P2 movesy,
Ex∼D∗X
[Val(x, y)] ≤ α .
Follows from LP duality theorem.
Andrew Drucker Kernel-Size Lower Bounds
![Page 28: Kernel-Size Lower Bounds: The Evidence from …people.csail.mit.edu/andyd/Kernel_LBs_2.pdfAndrew Drucker Kernel-Size Lower Bounds Summary Fortnow-Santhanam technique applies to randomized](https://reader034.vdocuments.net/reader034/viewer/2022042922/5f6d79cbd438df27a00711ca/html5/thumbnails/28.jpg)
Back to business
Time to apply these tools.
Let’s restate the [Fortnow-Santhanam’08] result.
Will switch from k’s to n’s...
Andrew Drucker Kernel-Size Lower Bounds
![Page 29: Kernel-Size Lower Bounds: The Evidence from …people.csail.mit.edu/andyd/Kernel_LBs_2.pdfAndrew Drucker Kernel-Size Lower Bounds Summary Fortnow-Santhanam technique applies to randomized](https://reader034.vdocuments.net/reader034/viewer/2022042922/5f6d79cbd438df27a00711ca/html5/thumbnails/29.jpg)
Back to business
Theorem (FS’08, restated)
Let L be an NP-complete language, L′ another language, andt(n) ≤ poly(n).Suppose there is a poly-time reduction
R(x) = R(x1, . . . , x t(n))
taking t(n) inputs of length n, and producing output such that
R(x) ∈ L′ ⇐⇒∨j
[x j ∈ L] .
Suppose too we have the output-size bound
|R(x)| ≤ O(t(n) log t(n)) .
Then, NP ⊆ coNP/poly.
Andrew Drucker Kernel-Size Lower Bounds
![Page 30: Kernel-Size Lower Bounds: The Evidence from …people.csail.mit.edu/andyd/Kernel_LBs_2.pdfAndrew Drucker Kernel-Size Lower Bounds Summary Fortnow-Santhanam technique applies to randomized](https://reader034.vdocuments.net/reader034/viewer/2022042922/5f6d79cbd438df27a00711ca/html5/thumbnails/30.jpg)
Simplifying
To ease discussion:
Assume L′ = L;
Fix t(n) = n10;
Assume ∣∣∣R (x1, . . . , xn10)∣∣∣ ≡ n3 .
(No more ideas needed for general case!)
Andrew Drucker Kernel-Size Lower Bounds
![Page 31: Kernel-Size Lower Bounds: The Evidence from …people.csail.mit.edu/andyd/Kernel_LBs_2.pdfAndrew Drucker Kernel-Size Lower Bounds Summary Fortnow-Santhanam technique applies to randomized](https://reader034.vdocuments.net/reader034/viewer/2022042922/5f6d79cbd438df27a00711ca/html5/thumbnails/31.jpg)
Proof strategy
Recall L is NP-complete. To prove theorem, enough to showthat
L ∈ coNP/poly , i.e., L ∈ NP/poly .
Thus, want to use R to build a non-uniform proof systemwitnessing membership in L.
Andrew Drucker Kernel-Size Lower Bounds
![Page 32: Kernel-Size Lower Bounds: The Evidence from …people.csail.mit.edu/andyd/Kernel_LBs_2.pdfAndrew Drucker Kernel-Size Lower Bounds Summary Fortnow-Santhanam technique applies to randomized](https://reader034.vdocuments.net/reader034/viewer/2022042922/5f6d79cbd438df27a00711ca/html5/thumbnails/32.jpg)
Shadows
For x ∈ {0, 1}n, say that x = (x1, . . . , xn10
) contains x if
x occurs as one of the x j ’s.
Define the shadow of x ∈ {0, 1}n by
shadow(x) := {z = R(x) : x contains x} ⊆ {0, 1}n3 .
Andrew Drucker Kernel-Size Lower Bounds
![Page 33: Kernel-Size Lower Bounds: The Evidence from …people.csail.mit.edu/andyd/Kernel_LBs_2.pdfAndrew Drucker Kernel-Size Lower Bounds Summary Fortnow-Santhanam technique applies to randomized](https://reader034.vdocuments.net/reader034/viewer/2022042922/5f6d79cbd438df27a00711ca/html5/thumbnails/33.jpg)
Shadows
Andrew Drucker Kernel-Size Lower Bounds
![Page 34: Kernel-Size Lower Bounds: The Evidence from …people.csail.mit.edu/andyd/Kernel_LBs_2.pdfAndrew Drucker Kernel-Size Lower Bounds Summary Fortnow-Santhanam technique applies to randomized](https://reader034.vdocuments.net/reader034/viewer/2022042922/5f6d79cbd438df27a00711ca/html5/thumbnails/34.jpg)
Shadows
Andrew Drucker Kernel-Size Lower Bounds
![Page 35: Kernel-Size Lower Bounds: The Evidence from …people.csail.mit.edu/andyd/Kernel_LBs_2.pdfAndrew Drucker Kernel-Size Lower Bounds Summary Fortnow-Santhanam technique applies to randomized](https://reader034.vdocuments.net/reader034/viewer/2022042922/5f6d79cbd438df27a00711ca/html5/thumbnails/35.jpg)
Shadows
Andrew Drucker Kernel-Size Lower Bounds
![Page 36: Kernel-Size Lower Bounds: The Evidence from …people.csail.mit.edu/andyd/Kernel_LBs_2.pdfAndrew Drucker Kernel-Size Lower Bounds Summary Fortnow-Santhanam technique applies to randomized](https://reader034.vdocuments.net/reader034/viewer/2022042922/5f6d79cbd438df27a00711ca/html5/thumbnails/36.jpg)
Shadows
Andrew Drucker Kernel-Size Lower Bounds
![Page 37: Kernel-Size Lower Bounds: The Evidence from …people.csail.mit.edu/andyd/Kernel_LBs_2.pdfAndrew Drucker Kernel-Size Lower Bounds Summary Fortnow-Santhanam technique applies to randomized](https://reader034.vdocuments.net/reader034/viewer/2022042922/5f6d79cbd438df27a00711ca/html5/thumbnails/37.jpg)
Shadows
Andrew Drucker Kernel-Size Lower Bounds
![Page 38: Kernel-Size Lower Bounds: The Evidence from …people.csail.mit.edu/andyd/Kernel_LBs_2.pdfAndrew Drucker Kernel-Size Lower Bounds Summary Fortnow-Santhanam technique applies to randomized](https://reader034.vdocuments.net/reader034/viewer/2022042922/5f6d79cbd438df27a00711ca/html5/thumbnails/38.jpg)
Shadows
Fact: if some z /∈ L is in the shadow of x , then x /∈ L.
(by OR-property of R...)
This is our basic form of evidence for membership in L!
(z will be non-uniform advice...)
Andrew Drucker Kernel-Size Lower Bounds
![Page 39: Kernel-Size Lower Bounds: The Evidence from …people.csail.mit.edu/andyd/Kernel_LBs_2.pdfAndrew Drucker Kernel-Size Lower Bounds Summary Fortnow-Santhanam technique applies to randomized](https://reader034.vdocuments.net/reader034/viewer/2022042922/5f6d79cbd438df27a00711ca/html5/thumbnails/39.jpg)
The main claim
Claim (FS ’08)
There exists a set Z ⊆ Ln3 , with
|Z | ≤ poly(n) ,
such that for every x ∈ Ln ,
shadow(x) ∩ Z 6= ∅ .
Intuition: the massive compression by R =⇒ some z is the imageof many sequences x , hence is in many shadows. Can collect these“popular” z ’s to hit all shadows (of Ln).
Claim easily implies L ∈ NP/poly... take Z as non-uniform advice.
Andrew Drucker Kernel-Size Lower Bounds
![Page 40: Kernel-Size Lower Bounds: The Evidence from …people.csail.mit.edu/andyd/Kernel_LBs_2.pdfAndrew Drucker Kernel-Size Lower Bounds Summary Fortnow-Santhanam technique applies to randomized](https://reader034.vdocuments.net/reader034/viewer/2022042922/5f6d79cbd438df27a00711ca/html5/thumbnails/40.jpg)
Shadows
To prove x ∈ L...
nondeterministically choose x ⊃ x and z ∈ Z ,and check:
Conclusion: x ∈ L.
Andrew Drucker Kernel-Size Lower Bounds
![Page 41: Kernel-Size Lower Bounds: The Evidence from …people.csail.mit.edu/andyd/Kernel_LBs_2.pdfAndrew Drucker Kernel-Size Lower Bounds Summary Fortnow-Santhanam technique applies to randomized](https://reader034.vdocuments.net/reader034/viewer/2022042922/5f6d79cbd438df27a00711ca/html5/thumbnails/41.jpg)
Shadows
To prove x ∈ L... nondeterministically choose x ⊃ x and z ∈ Z ,and check:
Conclusion: x ∈ L.
Andrew Drucker Kernel-Size Lower Bounds
![Page 42: Kernel-Size Lower Bounds: The Evidence from …people.csail.mit.edu/andyd/Kernel_LBs_2.pdfAndrew Drucker Kernel-Size Lower Bounds Summary Fortnow-Santhanam technique applies to randomized](https://reader034.vdocuments.net/reader034/viewer/2022042922/5f6d79cbd438df27a00711ca/html5/thumbnails/42.jpg)
Shadows
To prove x ∈ L... nondeterministically choose x ⊃ x and z ∈ Z ,and check:
Conclusion: x ∈ L.
Andrew Drucker Kernel-Size Lower Bounds
![Page 43: Kernel-Size Lower Bounds: The Evidence from …people.csail.mit.edu/andyd/Kernel_LBs_2.pdfAndrew Drucker Kernel-Size Lower Bounds Summary Fortnow-Santhanam technique applies to randomized](https://reader034.vdocuments.net/reader034/viewer/2022042922/5f6d79cbd438df27a00711ca/html5/thumbnails/43.jpg)
The main claim
Claim (FS ’08)
There exists a set Z ⊆ Ln3 , with
|Z | ≤ poly(n) ,
such that for every x ∈ Ln ,
shadow(x) ∩ Z 6= ∅ .
Andrew Drucker Kernel-Size Lower Bounds
![Page 44: Kernel-Size Lower Bounds: The Evidence from …people.csail.mit.edu/andyd/Kernel_LBs_2.pdfAndrew Drucker Kernel-Size Lower Bounds Summary Fortnow-Santhanam technique applies to randomized](https://reader034.vdocuments.net/reader034/viewer/2022042922/5f6d79cbd438df27a00711ca/html5/thumbnails/44.jpg)
Proof by game
To prove Claim, consider the following simul-move game betweenP1 (“Maker”) and P2 (“Breaker”):
Game
P1: chooses z ∈ Ln3 ;
P2: chooses x ∈ Ln;
Payoff to P2: Val(x , z) = 1 if z /∈ shadow(x), otherwise 0.
Lemma
There is a P1 strategy (dist’n D∗ over Ln3) such that for any x,
Ez∼D∗ [Val(z, x)] ≤ o(1) .
Our Claim follows easily, with |Z | = O(n). (Repeated sampling!)Andrew Drucker Kernel-Size Lower Bounds
![Page 45: Kernel-Size Lower Bounds: The Evidence from …people.csail.mit.edu/andyd/Kernel_LBs_2.pdfAndrew Drucker Kernel-Size Lower Bounds Summary Fortnow-Santhanam technique applies to randomized](https://reader034.vdocuments.net/reader034/viewer/2022042922/5f6d79cbd438df27a00711ca/html5/thumbnails/45.jpg)
Proof by game
Lemma
There is a P1 strategy (dist’n D∗ over Ln3) such that for any x,
Ez∼D∗ [Val(z, x)] ≤ o(1) .
By minimax theorem, it’s enough to beat any fixed P2 mixedstrategy
Dn (dist’n over Ln ) .
Idea: use P1 strategy induced by outputs of R on inputs fromDn...
Andrew Drucker Kernel-Size Lower Bounds
![Page 46: Kernel-Size Lower Bounds: The Evidence from …people.csail.mit.edu/andyd/Kernel_LBs_2.pdfAndrew Drucker Kernel-Size Lower Bounds Summary Fortnow-Santhanam technique applies to randomized](https://reader034.vdocuments.net/reader034/viewer/2022042922/5f6d79cbd438df27a00711ca/html5/thumbnails/46.jpg)
Proof by game
Say that z ∈ Ln3 is bad, if
Prx∼Dn
[z ∈ shadow(x)] ≤ 1− 1/n .
We’ve beaten strategy Dn if some z is not bad.
Andrew Drucker Kernel-Size Lower Bounds
![Page 47: Kernel-Size Lower Bounds: The Evidence from …people.csail.mit.edu/andyd/Kernel_LBs_2.pdfAndrew Drucker Kernel-Size Lower Bounds Summary Fortnow-Santhanam technique applies to randomized](https://reader034.vdocuments.net/reader034/viewer/2022042922/5f6d79cbd438df27a00711ca/html5/thumbnails/47.jpg)
Proof by game
Let D⊗tn denote t ind. copies of Dn.
Define dist’n R by
R = R(D⊗n10n
).
We claim that
Prz∼R
[z is bad ] = o(1) .
Andrew Drucker Kernel-Size Lower Bounds
![Page 48: Kernel-Size Lower Bounds: The Evidence from …people.csail.mit.edu/andyd/Kernel_LBs_2.pdfAndrew Drucker Kernel-Size Lower Bounds Summary Fortnow-Santhanam technique applies to randomized](https://reader034.vdocuments.net/reader034/viewer/2022042922/5f6d79cbd438df27a00711ca/html5/thumbnails/48.jpg)
Proof by game
Let x = (x1, . . . , xn10
) ∼ D⊗n10n , and
z = R(x) .
Consider any bad z . For [z = z ], we must have
xj ∈ shadow(z) ∀j ,
with happens with probability
≤ (1− 1/n)n10< 2−n
9.
Union bound over all bad z completes proof:
Pr [z is bad] ≤ 2n3
2n9= o(1) .
Andrew Drucker Kernel-Size Lower Bounds
![Page 49: Kernel-Size Lower Bounds: The Evidence from …people.csail.mit.edu/andyd/Kernel_LBs_2.pdfAndrew Drucker Kernel-Size Lower Bounds Summary Fortnow-Santhanam technique applies to randomized](https://reader034.vdocuments.net/reader034/viewer/2022042922/5f6d79cbd438df27a00711ca/html5/thumbnails/49.jpg)
Summary
IfR(x) : {0, 1}n×n10 → {0, 1}n3
is a compressive mapping with the “OR-property” for L, thenthere are z ∈ Ln3 lying in the “shadow” of many x ∈ Ln.
We collect a small set Z of these non-uniformly, use it toprove membership in Ln.
Note: nondeterminism still required to verify membership inLn: we have to guess extensions x → x which map to z!
Minimax theorem made our job easier.
Andrew Drucker Kernel-Size Lower Bounds
![Page 50: Kernel-Size Lower Bounds: The Evidence from …people.csail.mit.edu/andyd/Kernel_LBs_2.pdfAndrew Drucker Kernel-Size Lower Bounds Summary Fortnow-Santhanam technique applies to randomized](https://reader034.vdocuments.net/reader034/viewer/2022042922/5f6d79cbd438df27a00711ca/html5/thumbnails/50.jpg)
Summary
Fortnow-Santhanam technique applies to randomizedalgorithms avoiding false negatives. Also toco-nondeterministic alg’s (observed in [DvM ’10]).
[FS ’08] also used their tools to rule out “succinct PCPs” forNP...
Left open: evidence again two-sided error OR-compression;any strong evidence against AND-compression.
poly(k)-kernelizability of problems like k-Treewidth left open...
Andrew Drucker Kernel-Size Lower Bounds
![Page 51: Kernel-Size Lower Bounds: The Evidence from …people.csail.mit.edu/andyd/Kernel_LBs_2.pdfAndrew Drucker Kernel-Size Lower Bounds Summary Fortnow-Santhanam technique applies to randomized](https://reader034.vdocuments.net/reader034/viewer/2022042922/5f6d79cbd438df27a00711ca/html5/thumbnails/51.jpg)
Side note: a mystery
Fortnow-Santhanam prove we cannot efficiently compress
OR=(L)t(n)n instances to size O(t(n) log t(n)).
Input size is t(n) · n.
There is still a big gap here; consequences of compression tosize O(t(n) ·
√n)? If, e.g., L = SAT?
Same issue with [D’12] bounds...
Andrew Drucker Kernel-Size Lower Bounds