Complexity Lower Bounds, P vs NP & Gowers Blog ๐ท ≠ ๐ต๐ท? Try to prove they are different by using circuit complexity. Define a function to be ๐: {0,1}๐ → {±1} To built a circuit we define the following: 1) "Basic Functions" – ๐๐ (๐ฅ1 , … , ๐ฅ๐ ) = (-1) 2) "Basic Operations": ๐ฅ๐ +1 a. ๐ → -๐ b. ๐, ๐ → ๐ ∨ ๐ c. ๐, ๐ → ๐ ∧ ๐ 3) Straight line composition: ๐1 , ๐2 , … , ๐๐ such that ๐๐ is either a basic function or obtained from ๐1 , ๐2 < ๐ by basic operations. The definition represents a DAG (Directional Acyclic Graph) Definition: a function ๐น has (circuit) complexity ๐ if ∃ a circuit ๐1 , … , ๐๐ as above. ๐⁄ ๐๐๐๐ฆ: All functions with a polynomial circuit complexity are our equivalent of ๐. ๐{{0,1}๐ → {±1}|๐ โ๐๐ ๐ ๐๐๐๐๐ข๐๐ก ๐๐๐๐๐๐๐ฅ๐๐ก๐ฆ < ๐log log ๐ } If we find โ ∈ ๐๐ ๐ . ๐ก. โ ∉ ๐⁄๐๐๐๐ฆ then ๐ ≠ ๐๐. For that purpose, lets define a complexity measuring function ๐ . Such ๐ must satisfy: 1) ๐ (๐) = 1 if ๐ is basic 2) ๐ (๐) = ๐ (-๐) 3) If ๐ (๐), ๐ (๐) are small then ๐(๐ ∨ ๐) is small 4) ๐ (๐) is large for some โ ∈ ๐๐ Attempts to find such ๐ฟ Idea 1 Take the fourier representation of ๐น = ∑๐⊂[๐] ๐(๐ )๐๐ ๐ฬ๐๐๐ฅ = max|๐(๐ )| ๐ Define ๐ (๐) = ๐ 1 ๐๐๐ฅ ๐: {0,1}๐ → {±1} ๐ ๐ ∈ โ2 Instead of the standard base, we use the Fourier base: ๐ฅ๐ ๐{๐} = (-1) . ∀๐ ⊆ {1, … , ๐}. ๐๐ = ∏ ๐(๐) ๐∈๐ Fourier base is: {๐๐ }๐ Problem 2 1 = ๐๐๐ท๐ = โ๐โ22 = ∑|๐ฬ(๐ )| ∀๐{0,1}2 → {๐ผ1 } ∃๐ ๐ . ๐ก. |๐ฬ(๐ )| > And the function ๐(๐ฅ) = (-1) ๐ฅ1 ๐ฅ2 +๐ฅ2 ๐ฅ3 +๐ฅ3 ๐ฅ4 ,…,๐ฅ๐ ๐ฅ1 1 2๐ also has ๐ฬ๐๐๐ฅ = 1 ๐ 22 Could there be a good measure ๐ฟ? ๐ (๐ ∧ ๐) ≤ ๐ (๐) + ๐ (๐) } - a formal complexity measure ๐ (๐ ∧ ๐) ≤ ๐ (๐) + ๐ (๐) Related to the formula size of ๐ – ๐(๐)? Formula trees with basic formulas on leaves and basic operations in internal nodes. Formula size is the number of leaves. Claim: For any formal complexity measure ๐ – ๐ (๐) ≤ ๐(๐) Proof: By induction ๐ (๐ ∨ ๐) ≤ ๐ (๐) + ๐ (๐) ≤ ๐(๐) + ๐(๐) Assume the smallest formula for f writes ๐ = ๐ ∨ โ, then ๐(๐) = ๐, ๐(โ) = ๐ → ๐(๐) = ๐ + ๐ . Note that the smallest formula might be even smaller! ๐ (๐) = ๐ (๐) + ๐ (โ) ≤ ๐(๐) + ๐(โ) ≤ ๐(๐) So far it seems as if there isn't a good ๐ … On one hand: ๐ (๐) = ๐ฬ 1 ๐๐๐ฅ is bad for easy functions. On the other hand, ๐ (๐) = ๐(๐) is bad because tautological! We haven't made any progress... Natural Proofs – Razborov & Rudich ๐ {0,1}๐ด={0,1} X (0,1,1,0, … ) ๐ฅ=โ |๐ด| ๐ ⊂ {0,1}๐ด ๐น ⊂ {{0,1}๐ด → {±1}} ๐ is ๐ pseudo-random w.r.t. ๐น if ∀๐น ∈ โฑ |๐[๐น(๐ฅ) = 1|๐ฅ ∈ ๐] − ๐[๐น(๐ฅ) = 1]| < ๐ Extreme opposite – 1๐ ∈ โฑ ๐ is pseudo random if every ๐น ∈ โฑ cannot distinguish ๐ฅ ≈ ๐ from ๐ฅ ≈ {0,1} ๐ด Main point: Random functions of lowe complexity look like random functions w.r.t. a poly time distinguisher. Let ๐ = all points in {0,1}๐ด with respect to polyline functions ๐ฟ = {๐น: {0,1}๐ด → {±1}|๐น ๐๐ ๐๐๐๐ฆ๐๐๐๐๐๐ ๐ก๐๐๐ (๐๐ ๐๐ก๐ ๐๐๐๐ข๐ก ๐๐๐๐๐กโ ๐ด)} Statement: ∀๐ > 0. ๐ is ๐-pseudorandom w.r.t. โฑ. A "natural proof" for ๐ ≠ ๐๐ would - Devise a "simplicity" probability S of Boolean functions, so that ๐ (๐) = 1 for all simple (poly-time complexity) functions ๐น ∈ ๐. - If ๐ itself is poly-time computable (in its input length - |๐ด| = 2๐ ) then since X is ๐pseudorandom for โฑ and ๐ ∈ โฑ, it follows that ๐(๐) = 1 for almost all ๐น ∈ {0,1}๐ด . This is bad because a random function ๐น ∈ {0,1}^๐ด shouldn't be simple! So either ๐ ∉ โฑ or ๐(๐) = 1 for almost all functions. A proof is "natural" if it defines a simplicity property ๐ such that: (1) All low complexity functions are simple (2) A random function is not simple (3) Whether or not a function is simple can be determined in poly-time (4) Some NP-function is not simple 1,2,3 cannot hold together! -------- end of lesson 1 Connection between P, NP and Circuits ๐ฟ ⊂ {0,1}∗ ๐ฟ๐ โ ๐ฟ ∩ {0,1}๐ ๐ฟ = {๐ฟ๐ }∞ ๐=1 The Clique language has a circuit complexity ≥ ๐(๐) ↔ ∞ {๐ } For any sequence of circuits ๐ ๐=1 solves Clique ∃๐0 ∀๐ > ๐0 ๐ (๐ถ๐ ) > ๐ (๐) The class ๐⁄๐๐๐๐ฆ ≡ all languages computable by poly circuits. The set of poly-time functions looks like the set of all functions Looks like = To a simple observer (another polynomial time algorithm). Exercise: Let ๐ be a formal complexity measure. Prove that if there exist โ: {0,1}๐ → {±1} 1 ๐ (โ) > 4 โ ๐ , then ๐๐๐๐[๐ (๐) > ๐ ] ≥ 4 ๐ random, ๐: {0,1}๐ → {±1} Things that are known ๏ท The discrete log function Let โค๐ be the cyclic group with N elements Let g be a generator ๐บ = {๐1 , ๐2 , ๐3 , … } โค๐∗ ๐๐ : ๐บ → ๐บ ๐(๐ฅ) = ๐ ๐ฅ ๐ -−1 ← discrete log function, is 1-1, believed hard to compute. ๐ CONJ – There exists some ๐ > 0 ๐ . ๐ก. the complexity of this problem is ≥ 2๐ . Goldreich-Levin “Hard core bit”: Any one way permutation → gives rise to a pseudorandom generator. A pseudo random generator is: {0,1}๐ → {0,1}๐+1 Such that you cannot tell the difference between the half that emerged from the domain and the half that didn’t (in the range). ๐๐ ๐บ ( โ ๐ฅ , โ ๐ ) = (๐(๐ฅ) ๐ , ∑ ๐ฅ๐ ๐๐ ๐๐๐ 2) โ ,โ โ ๐ ๐ ๐ 2 ๐๐๐ก๐ 2 ๐ 2 ๐๐๐ก๐ {0,1}๐ ๐๐๐ ๐๐๐๐ ๐๐๐ก 2๐ {0,1} 2 ๐๐ ๐บ: → Now they constructed a pseudo random function generator The took a seed (denoted ๐ฆ). y y(๐1 (๐ฆ)) y(๐0 (๐ฆ)) y(๐0 (๐0 (๐ฆ))) y(๐1 (๐0 (๐ฆ))) y(๐0 (๐1 ((๐ฆ))) y(๐1 (๐1 ((๐ฆ))) ๐น(๐ฆ, ๐ฅ) โ ๐๐๐ต (๐๐ฅ๐ โ … โ ๐๐ฅ2 ๐๐ฅ1 (๐ฆ)) Define ๐๐ฆ : {0,1}๐ → {0,1} ๐๐ฆ (๐ฅ) = ๐น(๐ฆ, ๐ฅ) Consider the distribution {๐๐ฆ }๐ฆ∈{0,1}๐ , ๐ > ๐๐ ๐ > 2 ๏ท For each y ๐๐ฆ is poly-time computable ๏ท This distribution is pseudo-random against polynomial time(in ๐๐ ) distinguishers On the other hand… ๐ , or any property being used in a lower bound proof, shouldn’t be too complex either! ๐ (๐) = 1 iff ๐ has low circuit complexity is not good (trivial). Note ๐ ∈ ๐๐! Take the basic functions: (- − 1) โฎ 100 ๐ ๐ฅ1 (- − 1) โฎ ๐ฅ2 (- − 1) โฎ ๐ฅ3 … (- − 1) โฎ ๐ฅ๐ A model for generating a random formula. I have ๐ โ ๐100 โ 2 functions. But this is false! Why? Because using AND or OR changes the distribution from Gowers Norms (๐ผ๐ ๐ต๐๐๐๐) Fix a finite set ๐ด and consider โ๐ด (the vector space of functions ๐: ๐ด → โ. A norm on this space is a function โ๐ฅโ: โ ๐ด → โ+ s.t. โ๐ผ๐โ = |๐ผ|โ๐โ ๐ ≠ 0 → โ๐โ ≠ 0 โ๐ + ๐โ ≤ โ๐โ + โ๐โ 1 2 3 4 ๐๐๐ 0 to … Example: โ๐โ = max|๐(๐ฅ)| (๐๐ ๐ฅ ∈ ๐ด) Definition (dual norm): 1 โ๐โ∗ = max{〈๐, ๐〉|โ๐โ ≤ 1} where 〈๐, ๐〉 = ∑๐ฅ∈๐ด ๐(๐ฅ)๐(๐ฅ) ← “Correlation” |๐ด| Example: For โ๐โ∞, the dual is: โ๐โ∗∞ = max{< ๐, ๐ > |โ๐โ∞ ≤ 1, ∀๐ฅ: |๐(๐ฅ)| ≤ 1} 1 1 max {∑ ๐(๐ฅ)๐(๐ฅ) ||๐(๐ฅ)| ≤ 1} = ∑ ๐(๐ฅ)๐๐๐๐(๐(๐ฅ)) = ∑|๐(๐ฅ)| = โ๐โ1 |๐ด| |๐ด| In general: If โโโis in P, it doesn’t mean that โโโ∗is in P. Another Example: Let๐ดฬ be an abelian group โ๐โ4 42 = ๐ธ [๐(๐ฅ)๐(๐ฅ + ๐)๐(๐ฅ + ๐)๐(๐ฅ + ๐ + ๐)] ๐ฅ,๐,๐ Is it in P? YES. 1 Is โ๐โ∗42 in P? Turns out that โ๐โ∗42 = โ๐ฬโ4 = (∑๐ ⊆[๐](๐ฬ(๐ฅ)4 ))4 Exercise: Prove that the ๐ข2 norm is a norm. Hint: Cauchy’s norm. However, can define ๐ข2 norm TODO: Did not have enough time to copy the formula: โ๐โ๐ข2 = ( ๐ธ (๐(๐ฅ)๐(๐ฅ + ๐)๐(๐ฅ + ๐)๐(๐ฅ + ๐ + ๐)๐(๐ฅ + ๐)๐(๐ฅ + ๐ + ๐)๐() ๐ฅ,๐,๐,๐ Do not know a poly-time algorithm for โ โ โ∗๐ข2 “Goal” for introducing these norms was to extend fourier analysis to “higher degree”. ๐: {0,1}๐ → {±1}. ๐ฬ(๐ ) = ๐๐๐๐๐๐๐๐ก๐๐๐ ๐๐ ๐ ๐ค๐๐กโ ๐กโ๐ ๐๐๐๐๐๐ ๐๐ข๐๐๐ก๐๐๐ (−1)∑๐∈๐ ๐๐ ๐ฅ๐ : {0,1}๐ → {±1} ๐ฅ2 (๐ฅ1 , … , ๐ฅ๐ ) = ∏๐∈๐(−1)๐ฅ๐ ←linear phase functions Consider degree d phase functions (−1)^๐(๐ฅฬ ) where deg ๐ ≤ ๐ Fix ๐: {0,1}๐ → {±1} ๐๐ฆ : {0,1}๐ ๐ฆ ∈ {0,1}๐ → {±1} defined by ๐๐ฆ (๐ฅ) = ๐(๐ฅ) โ ๐(๐ฅ + ๐ฆ) ๐(๐ฅ) = (−1)(๐(๐ฅ)) Say, for instance: ๐(๐ฅ) = (−1)๐ฅ1 +๐ฅ2 +๐ฅ3 โ (−1)๐ฅ1 +๐ฅ2 +๐ฅ3 +๐ฆ1 +๐ฆ2 +๐ฆ3 = (−1)๐ฆ1 +๐ฆ+๐ฆ3 =constant! Doesn’t depend on x. If ๐(๐ฅ) = (−1)๐(๐ฅ) for deg ๐ ≤ ๐ Then ๐๐ฆ (๐ฅ) = (−1)๐ ′ (๐ฅ) ๐คโ๐๐๐ deg ๐ ′ ≤ ๐ − 1 (−1)๐(๐ฅ) โ (−1)๐(๐ฅ+๐ฆ) Define ๐๐ฆ,๐ง ≡ (๐๐ฆ )๐ง Similarily: ๐๐ฆ1 …๐ฆ๐ = (… ((๐๐ฆ1 ) ) … ) ๐ฆ2 ---- end of lesson 2 ๐ฆ๐ ๐: {0,1}๐ → {±1} ๐ โ๐โ2๐ข๐ = ๐ธ ๐ฅ1 …๐ฅ๐ ∈{0,1}๐ ∏ ๐(๐ฅ0 + ๐1 ๐ฅ1 + โฏ + ๐๐ ๐ฅ๐ ) ๐1 …๐๐ ∈{0,1} ๐ฅ0 = ๐๐๐๐๐๐ ๐ ๐๐๐ (๐ฅ1 … ๐ฅ๐ ) (dimension k affine subspace) ๐ฅ0 , ๐ฅ0 + ๐ฅ1 , ๐ฅ0 + ๐ฅ2 , ๐ฅ0 + ๐ฅ1 + ๐ฅ2 Example: Suppose ๐ is a linear function ∃๐1 , … , ๐๐ ๐ . ๐ก. ๐(๐ฅ) = ∑ ๐๐ ๐ฅ๐ (๐๐๐2) Then for any choice of ๐ฅ0 , ๐ฅ1 , ๐ฅ2 ๐(๐ฅ0 ) + +๐(๐ฅ0 ) + ๐(๐ฅ1 ) + + โฏ = 0 Hence for any choise of ๐ฅ0 , ๐ฅ1 , ๐ฅ2 + + ๐(๐ฅ1 + ๐1 ๐ฅ1 + ๐2 ๐ฅ2 ) ≡ 0 ๐1 ,๐2 Similarly the expectancy is 0 as well. Linearity Testing (proven by Blum-Luby-Rubinfeld) ๐: {0,1}๐ → {±1} Question: Is ๐ a linear function? Definition 1: ∃๐1 , … , ๐๐ ๐ . ๐ก. ∀๐ฅ. ๐(๐ฅ) = ∑ ๐๐ ๐ฅ๐ Definition 2: ∀๐ฅ, ๐ฆ ๐(๐ฅ) + ๐(๐ฆ) = ๐(๐ฅ + ๐ฆ) Definition 2 implies definition 1 since: We can define ๐๐ = ๐(๐๐ ), then ๐(๐ฅ) = ๐(∑ ๐ฅ๐ ๐๐ ) ๐๐๐๐๐๐๐๐ก๐ฆ = ∑ ๐ฅ๐ โ ๐(๐๐ )… Testing Global object, e.g. ๐: {0,1}๐ → {0,1} Want to test if ๐ ∈ ๐ซ In our example – all linear functions. Only willing to invest limited resources, but willing to randomize. Question: Can we deduce global property by considering local behavior? If the answer is yes we say that this ๐ซ is testable. Non testable property: ๐ฅ1 … ๐ฅ๐ - Boolean variables. ~๐ฅ1 … ~๐ฅ๐ - their negations. Fom the list above, I select ๐ 3-CNF clauses indices at andom. Most of the time we will find clauses that don’t have any shared variables If ๐ > 50๐ than with high probability ๐ is unsatisfiable! PCP “theory” implies that every polynomialy varifiable property (e.g. formula satisfiability) can be cast (“encoded”) in testable form. Definition: ∀๐, ๐ {0,1}๐ → {0,1} Distance(๐, ๐) = ๐๐๐๐๐ [๐(๐ฅ) ≠ ๐(๐ฅ)] ๐ฅ∈{0,1} Distance(๐, ๐) = min ๐ท๐๐ ๐ก๐๐๐๐(๐, ๐) f∈S Theorem (BLR): Let ๐: {0,1}๐ → {0,1} If distance (๐, ๐ฟ๐ผ๐๐ธ๐ด๐ ) ≥ 0 Then ๐๐๐๐(๐(๐ฅ) + ๐(๐ฆ) + ๐(๐ฅ + ๐ฆ) ≠ 0) ≥ Ω(๐ฟ) ๐ฅ,๐ฆ Theorem (AKKLR): Let ๐: {0,1}๐ → {0,1} if ๐ท๐๐ ๐ก๐๐๐๐(๐, ๐ท๐ธ๐บ๐ ๐ธ๐ธ(๐)) ≥ δ Then ๐๐๐๐ (๐๐๐ (๐(๐ฅ0 + ๐1 ๐ฅ1 + โฏ + ๐๐+1 ๐ฅ๐+1 ))) ≥ Ω(๐ฟ โ 2−๐ ) ๐ฅ0 ,…,๐ฅ๐+1 So we select an affine space of ๐ + 1 points. We look at all of these points and check that they are not zero. ๐ is a degree k function if ๐(๐ฅ1 , … , ๐ฅ๐ ) = ∑ ๐๐ ∏ ๐ฅ๐ |๐|⊆๐ ๐∈๐ ๐ฅ, ๐ฆ ๐(๐ฅ1 , … , ๐ฅ๐ ) = ๐ฅ1 ๐ฅ2 … ๐ฅ๐ If ๐ has degree ๐ then ∀๐ฅ0 , … , ๐ฅ๐+1 ๐๐๐ (๐(๐ฅ0 + ๐1 ๐ฅ1 + โฏ + ๐๐+! ๐ฅ๐+1 )) Proof: Let ๐๐ฆ (๐ฅ) = ๐(๐ฅ + ๐ฆ) The function ๐ = ๐๐ฅ๐+1 ,๐ฅ๐ ,…,๐ฅ2 has degree 0 (it is constant) The above expression equals ๐(๐ฅ0 )๐๐๐ ๐(๐ฅ0 + ๐ฅ1 ) ≡ 0 Claim: Let ๐๐๐๐(๐, ๐ − 1) be the correlation of ๐ with degree ๐ − 1 polynomials. ๐๐๐๐ = (1 − ๐ฟ) − ๐ฟ = 1 − 2 โ ๐ท๐๐ ๐ก๐๐๐๐(๐, ๐ − 1 ๐ท๐ธ๐บ๐ธ๐ธ ๐น๐๐๐ถ๐๐ผ๐๐๐) ๐๐๐๐(๐, ๐ − 1) ≤ โ๐โ๐ข๐ Reed-Mulle (Low-Degee) Test Let ๐ be the closest degree ๐ polynomial to ๐. ๐ฟ โ ๐ท๐๐ ๐ก๐๐๐๐(๐, ๐) 1. ๐ฟ is tiny – ๐ฟ < ๐ฝ2−๐ If an affine ๐ + 1 space contains exactly one point ๐ฅ ๐ . ๐ก. ๐(๐ฅ) ≠ ๐(๐ฅ) then the test rejects. We will prove that this happens with constant probability. Assume ๐ฟ~2−๐ Choose a random affine subspace ๐ด by choosing ๐๐×๐ a random full rank matrix over ๐ฝ2 and a random ๐ ∈ ๐ฝ๐2 ๐ด = {๐๐ฅ = ๐๐ฅ + ๐|๐ฅ ∈ ๐ฝ๐2 } For each ๐ฅ ๐ธ๐ฅ − The event ๐(๐ฅ) ≠ ๐(๐ฅ) ๐น๐ฅ − ๐ธ๐ฅ and ∀ ๐ฆ ≠ ๐ฅ ๐(๐๐ฆ ) = ๐(๐๐ฆ ) ๐๐ฅ is distributed uniformly in ๐ฝ๐2 ๐๐ฆ is distributed uniformly on ๐ฝ๐2 \{๐๐ฅ } ∀๐ฅ ≠ ๐ฆ๐๐๐๐[๐ธ๐ฅ ] = ๐ฟ ๐๐๐๐[๐ธ๐ฅ ๐๐๐ ๐ธ๐ฆ ] ≤ ๐ฟ 2 ๐,๐ ๐๐๐๐[๐น๐ฅ ] ≥ ๐๐๐๐[๐ธ๐ฅ ] − ∑ ๐๐๐๐[๐ธ๐ฆ ∧ ๐ธ๐ฅ ] ≥ ๐ฟ − 2๐ โ ๐ฟ 2 ≈ ๐ฟ ๐ฅ≠๐ฆ ๐๐๐๐ (โ ๐น๐ฅ ) = ∑ ๐๐๐๐[๐น๐ฅ ] = 2๐ โ ๐ฟ = ๐ถ๐๐๐ ๐ก๐๐๐ก ๐,๐ 2. --- End of lesson 3 ๐ฅ ๐ฅ Last week we: ๏ท Defined the ๐ข๐ norm ⇔Degree ๐ − 1 test (“low degree test”) ๏ท Proved the following theorem: If โ๐โ๐ข๐ > 1 − ๐ฟ Then there ∃๐ of degree ๐ − 1 〈๐, ๐〉 = ๐ธ [๐(๐ฅ)๐(๐ฅ)] > 1 − ๐ฟ ๐ ๐ฅ Lemma 1: Let ๐: {0,1}๐ → {0,1} ๐ - some polynomial Denote ๐๐๐๐(๐, deg ๐) = max{〈๐, ๐〉|๐ = (−1) ๐(๐ฅ) , ๐ − deg ๐} So: ๐๐๐๐(๐, deg ๐) ≤ โ๐โ๐ข๐+1 Lemma 2: For every โ: {0,1}๐ → {0,1} โโโ๐ข๐ ≤ โโโ๐ข๐+1 Proof of 2: We shall use the fact that: ๐ธ[๐ 2 ] ≥ (๐ธ[๐])2 ๐+1 ๐+1 โโโ2๐ข๐+1 = ๐ธ ∏ ๐ฅ ๐ฆ1 ,…,๐ฆ๐ ๐ ,…,๐ ๐+1 ๐ฆ๐+1 1 โ (๐ฅ + ∑ ๐๐ ๐ฆ๐ ) ๐=1 ๐ = ๐ธ ๐ธ ๐ ∏ โ (๐ฅ + ∑ ๐๐ ๐ฆ๐ ) ∏ โ (๐ฅ + ๐ฆ๐+1 ∑ ๐๐ ๐ฆ๐ ) ๐ฆ1 ,…,๐ฆ๐ ๐ฅ ๐ฆ๐+1 ๐ ,…,๐ 1 ๐ ๐=1 ๐1 ,…,๐๐ ๐=1 ′ Now let’s fix: ๐ฅ ← ๐ฅ ๐ฆ ′ ← ๐ฅ + ๐ฆ๐+1 ๐ = ๐ธ ๐ธ ๐ ′ ′ ∏ โ (๐ฅ + ∑ ๐๐ ๐ฆ๐ ) ∏ โ (๐ฆ ∑ ๐๐ ๐ฆ๐ ) ๐ฆ1 ,…,๐ฆ๐ ๐ฅ ๐ฆ๐+1 ๐ ,…,๐ 1 ๐ ๐=1 ๐1 ,…,๐๐ ๐=1 2 ๐ = ๐ธ ๐ฆ1 ,…,๐ฆ๐ (๐ธ ∏ โ (๐ฅ ′ + ∑ ๐๐ ๐ฆ๐ )) ๐ฅ ๐1 ,…,๐๐ ๐=1 2 ๐ ≥( ๐ธ ๐ฅ ๐+1 ∏ โ (๐ฅ ′ + ∑ ๐๐ ๐ฆ๐ )) = โโโ2๐ข๐ ๐ฆ! ,…,๐ฆ๐ ๐ ,…,๐ 1 ๐ โ ๐=1 Proof of 1: For any โ: {0,1}๐ → {±1} 1 1) |๐ธ โ(๐ฅ)| = โโโ๐ข′ ๐ฅ ๐×3 (โโโ๐ข2 4 4 = (๐ธ |โฬ(๐ )| ) ) ๐ฅ 2) ∀๐ โโโ๐ข๐ ≤ โโโ๐ข๐+1 โ๐ โ ๐โ๐ข๐ 3) ∀๐. ∀๐: {0,1}๐ → {±1} degree ๐ polynomial โ ๐๐๐๐๐ก๐ค๐๐ ๐ ๐๐ข๐๐ก๐๐๐๐๐๐๐ก๐๐๐ = โ๐โ๐ข๐+1 Proof for 1: โโโ2๐ข1 = ๐ธ ∏ โ(๐ฅ + ๐๐ฆ1 ) = ๐ธ ∏ โ(๐ฅ + ๐๐ฆ1 ) ๐ฅ ๐ฅ ๐ฆ1 ๐=0,1 ′ ′ ๐ฆ1 ๐=0,1 Denote ๐ฅ = ๐ฅ, ๐ฆ = ๐ฅ + ๐ฆ = (๐ธ โ(๐ฅ)) (๐ธ โ(๐ฆ)) ๐ฅ ๐ฆ Proof for 3: ๐+1 โ๐ โ ๐+1 ๐โ2๐ข๐+1 = ๐ธ ๐ฅ ๐+1 ๐+1 ∏ ๐ฆ1 ,…,๐ฆ๐+1 ๐ ,…,๐ 1 ๐+1 ๐ (๐ฅ + ∑ ๐๐ ๐ฆ๐ ) ๐ (๐ฅ + ∑ ๐๐ ๐ฆ๐ ) = โ๐โ2๐ข๐+1 ๐=1 Because ∀๐ฅ, ๐ฆ1 , … , ๐ฆ๐+1 ∏๐1 ,…,๐๐ ๐ฅ + ๐=1 ∑๐+1 ๐=1 ๐๐ ๐ฆ๐ =1 Let ๐: {0,1}๐ → {±1} be the degree ๐ function closest to ๐ (i.e. attaining max correlation). Define โ(๐ฅ) = ๐(๐ฅ) โ ๐(๐ฅ) ๐๐๐๐(๐, ๐) = |๐ธ ๐(๐ฅ)๐(๐ฅ)| ๐๐ฆ ๐ ๐ก๐๐ 1 ๐ฅ - = ๐๐ฆ ๐ ๐ก๐๐ 2 โโโ๐ข๐ ≤ โโโ๐ข๐+1 ๐๐ฆ ๐ ๐ก๐๐ 3 = โ๐โ๐ข๐+1 โ considered โ๐โ๐ข3 as a “formal complexity measure” Consider dual norms: โ๐โ∗ = max{〈๐, ๐〉|โ๐โ ≤ 1} Motivation for dual: 1) “NP-ish” definition possibly circumvents RR 2) More robust TODO: Draw world Suppose we have two parts in our world. A and B And we have two functions ๐, ๐ such that ๐ is random on A and is 1 on B and ๐ is the exact opposite. โ๐โ๐ข3 =constant. โ๐โ๐ข3 =constant as well. โ =๐∨๐ โโโ๐ข3 = ๐ง๐๐๐! This is a problem! We just use an or and got such a dramatic difference โโโ∗ ≥ 〈โ, โ〉 = 1 〈โ, ๐ผโ〉 = ๐ผ 1 Can take ๐ผ = โโโ - very large! โ๐โ∗๐ข3 =? โ๐โ∗๐ข3 Need to find a “norming function” Use โ! โ โ โ=1 โโโ 1 โ 1 1 โ๐โ∗ ≥ 〈๐, 〉 = ๐ธ ๐(๐ฅ)โ(๐ฅ) = ๐๐๐๐(๐ด) โ ๐ธ๐(๐ฅ)๐(๐ฅ) + ๐๐๐๐[๐ด] โ ๐ธโ(๐ฅ) โ 1 = 2 โ โโโ โโโ ๐ฅ 1 โโโ =very large! Note: 〈๐, ๐〉 ≤ โ๐โ โ โ๐โ∗ Question 1: Given ๐: {0,1}๐ → {±1} Can we compute โ๐โ∗๐ข๐ in polytime? (an open question even for k=3) Question 2: Suppose you know that โ๐โ∗๐ข3 ≥ ๐ by [Samorodnitsky ‘07] ∃ deg 2 polynomial 2 that correlates with ๐. Can P be found in time poly(2๐ )? (search space is 2๐ ) Gappalan-Klivans-Zuckerman ’08: “list-decoding Reed-Muller Codes”. If ๐ is ๐-correlated with some degree-2 function, then the following is true: 1) Number of deg 2 polynomials correlation with ๐ ≤ 2๐๐ (๐) 2) Can find list above in time 2๐๐ (๐) Interpretation: if โ๐โ small then โ๐โ∗ is large Simplicity Property = {๐|โ๐โ∗ ๐๐ ๐ ๐๐๐๐} ⊆ {๐|โ๐โ ๐๐ ๐๐๐๐๐}- This is a property in ๐! If the first if ๐1 and the second is ๐2 , then the first one that contains all functions. Next idea: ๐ข๐ norm for super-constant ๐. Now the naïve algorithm for computing โ๐โ๐ข๐ Takes (2๐ )๐ −time. If ๐ = ๐(๐) this is not polytime ๐ = 2๐ , ๐^๐พ Still intent to use dual โ โ∗๐ข๐ norm (want robustness) Question 3: Is there an algorithm for โ โ๐ข๐ running in time better than ๐ ๐−1 โ log ๐.