9.1. Low-degree polynomials are reasonable
The Hypercontractivity Theorem. Let
f:\{-1,1\}^n\to\mathbb R and 1\le p\le q\le\infty. If
0\le\rho\le\sqrt{\frac{p-1}{q-1}},
then
\|T_\rho f\|_q\le\|f\|_p.
The book defers the proof to Chapter 10; this node supplies no assumption.
Bonami Lemma. If f:\{-1,1\}^n\to\mathbb R has degree at most k,
then
\|f\|_4\le(\sqrt3)^k\|f\|_2,
\qquad
\mathbb E[f^4]\le9^k\mathbb E[f^2]^2.
Lean code for Theorem9.1.2●1 theorem
Associated Lean declarations
-
FABL.bonamiLemma_fourthMoment[complete]
-
FABL.bonamiLemma_fourthMoment[complete]
-
theoremdefined in FABL/Chapter09/Reasonableness.leancomplete
theorem FABL.bonamiLemma_fourthMoment {n k : ℕ} (f : FABL.SignCube n → ℝ) (hdegree : FABL.fourierDegree f ≤ k) : (Finset.univ.expect fun x => f x ^ 4) ≤ 9 ^ k * (Finset.univ.expect fun x => f x ^ 2) ^ 2
theorem FABL.bonamiLemma_fourthMoment {n k : ℕ} (f : FABL.SignCube n → ℝ) (hdegree : FABL.fourierDegree f ≤ k) : (Finset.univ.expect fun x => f x ^ 4) ≤ 9 ^ k * (Finset.univ.expect fun x => f x ^ 2) ^ 2
The Bonami Lemma in fourth-moment form.
-
FABL.twoQHypercontractivity_of_two_lt[complete] -
FABL.twoQHypercontractivity[complete]
$(2,q)$-Hypercontractivity Theorem. For 2\le q\le\infty,
\left\|T_{1/\sqrt{q-1}}f\right\|_q\le\|f\|_2.
Consequently, if \deg(f)\le k, then
\|f\|_q\le(\sqrt{q-1})^k\|f\|_2.
Lean code for Theorem9.1.3●2 theorems
Associated Lean declarations
-
FABL.twoQHypercontractivity_of_two_lt[complete]
-
FABL.twoQHypercontractivity[complete]
-
FABL.twoQHypercontractivity_of_two_lt[complete] -
FABL.twoQHypercontractivity[complete]
-
theoremdefined in FABL/Chapter09/ProductHypercontractivity.leancomplete
theorem FABL.twoQHypercontractivity_of_two_lt {n : ℕ} (q : ℝ) (hq : 2 < q) (f : FABL.SignCube n → ℝ) : FABL.uniformLpNorm q ((FABL.noiseOperator (1 / √(q - 1))) f) ≤ FABL.uniformLpNorm 2 f
theorem FABL.twoQHypercontractivity_of_two_lt {n : ℕ} (q : ℝ) (hq : 2 < q) (f : FABL.SignCube n → ℝ) : FABL.uniformLpNorm q ((FABL.noiseOperator (1 / √(q - 1))) f) ≤ FABL.uniformLpNorm 2 f
The dual `(2,q)` theorem for `q > 2`.
-
theoremdefined in FABL/Chapter09/ProductHypercontractivity.leancomplete
theorem FABL.twoQHypercontractivity {n : ℕ} (q : ℝ) (hq : 2 ≤ q) (f : FABL.SignCube n → ℝ) : FABL.uniformLpNorm q ((FABL.noiseOperator (1 / √(q - 1))) f) ≤ FABL.uniformLpNorm 2 f
theorem FABL.twoQHypercontractivity {n : ℕ} (q : ℝ) (hq : 2 ≤ q) (f : FABL.SignCube n → ℝ) : FABL.uniformLpNorm q ((FABL.noiseOperator (1 / √(q - 1))) f) ≤ FABL.uniformLpNorm 2 f
The full finite-exponent `n`-bit `(2,q)`-Hypercontractivity Theorem.
-
FABL.pTwoHypercontractivity[complete] -
FABL.noiseStability_pSubOne_le_uniformLpNorm_sq[complete]
$(p,2)$-Hypercontractivity Theorem. For 1\le p\le2,
\|T_{\sqrt{p-1}}f\|_2\le\|f\|_p.
Equivalently, for 0\le\rho\le1,
\operatorname{Stab}_\rho[f]\le\|f\|_{1+\rho}^2.
Lean code for Theorem9.1.4●2 theorems
Associated Lean declarations
-
FABL.pTwoHypercontractivity[complete]
-
FABL.noiseStability_pSubOne_le_uniformLpNorm_sq[complete]
-
FABL.pTwoHypercontractivity[complete] -
FABL.noiseStability_pSubOne_le_uniformLpNorm_sq[complete]
-
theoremdefined in FABL/Chapter09/ProductHypercontractivity.leancomplete
theorem FABL.pTwoHypercontractivity {n : ℕ} (p : ℝ) (hp1 : 1 ≤ p) (hp2 : p ≤ 2) (f : FABL.SignCube n → ℝ) : FABL.uniformLpNorm 2 ((FABL.noiseOperator √(p - 1)) f) ≤ FABL.uniformLpNorm p f
theorem FABL.pTwoHypercontractivity {n : ℕ} (p : ℝ) (hp1 : 1 ≤ p) (hp2 : p ≤ 2) (f : FABL.SignCube n → ℝ) : FABL.uniformLpNorm 2 ((FABL.noiseOperator √(p - 1)) f) ≤ FABL.uniformLpNorm p f
The full `n`-bit `(p,2)`-Hypercontractivity Theorem with the sharp correlation.
-
theoremdefined in FABL/Chapter09/ProductHypercontractivity.leancomplete
theorem FABL.noiseStability_pSubOne_le_uniformLpNorm_sq (n : ℕ) (p : ℝ) (hp1 : 1 ≤ p) (hp2 : p ≤ 2) (f : FABL.SignCube n → ℝ) : FABL.uniformInner f ((FABL.noiseOperator (p - 1)) f) ≤ FABL.uniformLpNorm p f ^ 2
theorem FABL.noiseStability_pSubOne_le_uniformLpNorm_sq (n : ℕ) (p : ℝ) (hp1 : 1 ≤ p) (hp2 : p ≤ 2) (f : FABL.SignCube n → ℝ) : FABL.uniformInner f ((FABL.noiseOperator (p - 1)) f) ≤ FABL.uniformLpNorm p f ^ 2
Stability form of the full `n`-bit `(p,2)` theorem.
-
FABL.IsReasonable[complete] -
FABL.isReasonable_const_mul_iff[complete]
Definition 9.1. For B\ge1, a real random variable X is
B-reasonable when
\mathbb E[X^4]\le B\mathbb E[X^2]^2.
Equivalently,
\|X\|_4\le B^{1/4}\|X\|_2.
Lean code for Definition9.1.5●2 declarations
Associated Lean declarations
-
FABL.IsReasonable[complete]
-
FABL.isReasonable_const_mul_iff[complete]
-
FABL.IsReasonable[complete] -
FABL.isReasonable_const_mul_iff[complete]
-
defdefined in FABL/Chapter09/Reasonableness.leancomplete
def FABL.IsReasonable.{u_1} {Ω : Type u_1} [Fintype Ω] (p : PMF Ω) (B : ℝ) (X : Ω → ℝ) : Prop
def FABL.IsReasonable.{u_1} {Ω : Type u_1} [Fintype Ω] (p : PMF Ω) (B : ℝ) (X : Ω → ℝ) : Prop
O'Donnell, Definition 9.1: a finite real random variable is `B`-reasonable when `B ≥ 1` and its fourth moment is at most `B` times the square of its second moment.
-
theoremdefined in FABL/Chapter09/Reasonableness.leancomplete
theorem FABL.isReasonable_const_mul_iff.{u_1} {Ω : Type u_1} [Fintype Ω] (p : PMF Ω) (B c : ℝ) (hc : c ≠ 0) (X : Ω → ℝ) : (FABL.IsReasonable p B fun x => c * X x) ↔ FABL.IsReasonable p B X
theorem FABL.isReasonable_const_mul_iff.{u_1} {Ω : Type u_1} [Fintype Ω] (p : PMF Ω) (B c : ℝ) (hc : c ≠ 0) (X : Ω → ℝ) : (FABL.IsReasonable p B fun x => c * X x) ↔ FABL.IsReasonable p B X
Scaling by a nonzero constant preserves reasonableness.
-
FABL.uniformSign_isReasonable[complete] -
FABL.sparseBernoulli_isReasonable[complete] -
FABL.sparseBernoulli_not_reasonable_of_lt[complete] -
FABL.IsReasonableMeasure[complete] -
FABL.standardGaussian_secondMoment[complete] -
FABL.standardGaussian_fourthMoment[complete] -
FABL.standardGaussian_isThreeReasonable[complete] -
FABL.uniformIntervalMoment[complete] -
FABL.uniformIntervalMeasure[complete] -
FABL.integral_pow_uniformIntervalMeasure[complete] -
FABL.uniformIntervalMoment_two[complete] -
FABL.uniformIntervalMoment_four[complete] -
FABL.uniformInterval_isNineFifthsReasonable[complete] -
FABL.uniformIntervalMeasure_isNineFifthsReasonable[complete]
Example 9.2. A uniform sign is 1-reasonable, a standard Gaussian is
3-reasonable, and the uniform law on [-1,1] is
9/5-reasonable. If Y=1 with probability 2^{-n} and Y=0
otherwise, then Y is not B-reasonable unless B\ge2^n.
Lean code for Lemma9.1.6●14 declarations
Associated Lean declarations
-
FABL.uniformSign_isReasonable[complete]
-
FABL.sparseBernoulli_isReasonable[complete]
-
FABL.sparseBernoulli_not_reasonable_of_lt[complete]
-
FABL.IsReasonableMeasure[complete]
-
FABL.standardGaussian_secondMoment[complete]
-
FABL.standardGaussian_fourthMoment[complete]
-
FABL.standardGaussian_isThreeReasonable[complete]
-
FABL.uniformIntervalMoment[complete]
-
FABL.uniformIntervalMeasure[complete]
-
FABL.integral_pow_uniformIntervalMeasure[complete]
-
FABL.uniformIntervalMoment_two[complete]
-
FABL.uniformIntervalMoment_four[complete]
-
FABL.uniformInterval_isNineFifthsReasonable[complete]
-
FABL.uniformIntervalMeasure_isNineFifthsReasonable[complete]
-
FABL.uniformSign_isReasonable[complete] -
FABL.sparseBernoulli_isReasonable[complete] -
FABL.sparseBernoulli_not_reasonable_of_lt[complete] -
FABL.IsReasonableMeasure[complete] -
FABL.standardGaussian_secondMoment[complete] -
FABL.standardGaussian_fourthMoment[complete] -
FABL.standardGaussian_isThreeReasonable[complete] -
FABL.uniformIntervalMoment[complete] -
FABL.uniformIntervalMeasure[complete] -
FABL.integral_pow_uniformIntervalMeasure[complete] -
FABL.uniformIntervalMoment_two[complete] -
FABL.uniformIntervalMoment_four[complete] -
FABL.uniformInterval_isNineFifthsReasonable[complete] -
FABL.uniformIntervalMeasure_isNineFifthsReasonable[complete]
-
theoremdefined in FABL/Chapter09/Reasonableness.leancomplete
theorem FABL.uniformSign_isReasonable : FABL.IsReasonable (FABL.uniformPMF FABL.Sign) 1 FABL.signValue
theorem FABL.uniformSign_isReasonable : FABL.IsReasonable (FABL.uniformPMF FABL.Sign) 1 FABL.signValue
O'Donnell, Example 9.2: a uniformly random sign is `1`-reasonable.
-
theoremdefined in FABL/Chapter09/Reasonableness.leancomplete
theorem FABL.sparseBernoulli_isReasonable (n : ℕ) : let q := 1 / 2 ^ n; have hq := ⋯; FABL.IsReasonable (FABL.pBiasedBitPMF q hq) (2 ^ n) FABL.zeroOneValue
theorem FABL.sparseBernoulli_isReasonable (n : ℕ) : let q := 1 / 2 ^ n; have hq := ⋯; FABL.IsReasonable (FABL.pBiasedBitPMF q hq) (2 ^ n) FABL.zeroOneValue
O'Donnell, Example 9.2: the Bernoulli variable of success probability `2⁻ⁿ` is `2ⁿ`-reasonable.
-
theoremdefined in FABL/Chapter09/Reasonableness.leancomplete
theorem FABL.sparseBernoulli_not_reasonable_of_lt (n : ℕ) {B : ℝ} (hB : B < 2 ^ n) : let q := 1 / 2 ^ n; have hq := ⋯; ¬FABL.IsReasonable (FABL.pBiasedBitPMF q hq) B FABL.zeroOneValue
theorem FABL.sparseBernoulli_not_reasonable_of_lt (n : ℕ) {B : ℝ} (hB : B < 2 ^ n) : let q := 1 / 2 ^ n; have hq := ⋯; ¬FABL.IsReasonable (FABL.pBiasedBitPMF q hq) B FABL.zeroOneValue
The same sparse Bernoulli example cannot be `B`-reasonable with `B < 2ⁿ`.
-
defdefined in FABL/Chapter09/ContinuousReasonableness.leancomplete
def FABL.IsReasonableMeasure.{u_1} {Ω : Type u_1} [MeasurableSpace Ω] (μ : MeasureTheory.Measure Ω) (B : ℝ) (X : Ω → ℝ) : Prop
def FABL.IsReasonableMeasure.{u_1} {Ω : Type u_1} [MeasurableSpace Ω] (μ : MeasureTheory.Measure Ω) (B : ℝ) (X : Ω → ℝ) : Prop
Measure-theoretic extension of Definition 9.1 for the continuous examples in the book.
-
theoremdefined in FABL/Chapter09/ContinuousReasonableness.leancomplete
theorem FABL.standardGaussian_secondMoment : ∫ (x : ℝ), x ^ 2 ∂ProbabilityTheory.gaussianReal 0 1 = 1
theorem FABL.standardGaussian_secondMoment : ∫ (x : ℝ), x ^ 2 ∂ProbabilityTheory.gaussianReal 0 1 = 1
The standard Gaussian has second moment one.
-
theoremdefined in FABL/Chapter09/ContinuousReasonableness.leancomplete
theorem FABL.standardGaussian_fourthMoment : ∫ (x : ℝ), x ^ 4 ∂ProbabilityTheory.gaussianReal 0 1 = 3
theorem FABL.standardGaussian_fourthMoment : ∫ (x : ℝ), x ^ 4 ∂ProbabilityTheory.gaussianReal 0 1 = 3
The standard Gaussian has fourth moment three.
-
theoremdefined in FABL/Chapter09/ContinuousReasonableness.leancomplete
theorem FABL.standardGaussian_isThreeReasonable : FABL.IsReasonableMeasure (ProbabilityTheory.gaussianReal 0 1) 3 fun x => x
theorem FABL.standardGaussian_isThreeReasonable : FABL.IsReasonableMeasure (ProbabilityTheory.gaussianReal 0 1) 3 fun x => x
Example 9.2: a standard Gaussian is `3`-reasonable.
-
defdefined in FABL/Chapter09/ContinuousReasonableness.leancomplete
def FABL.uniformIntervalMoment (m : ℕ) : ℝ
def FABL.uniformIntervalMoment (m : ℕ) : ℝ
The normalized `m`-th moment of the uniform law on `[-1,1]`.
-
defdefined in FABL/Chapter09/ContinuousReasonableness.leancomplete
def FABL.uniformIntervalMeasure : MeasureTheory.Measure ℝ
def FABL.uniformIntervalMeasure : MeasureTheory.Measure ℝ
The uniform probability measure on the interval `[-1,1]`.
-
theoremdefined in FABL/Chapter09/ContinuousReasonableness.leancomplete
theorem FABL.integral_pow_uniformIntervalMeasure (m : ℕ) : ∫ (x : ℝ), x ^ m ∂FABL.uniformIntervalMeasure = FABL.uniformIntervalMoment m
theorem FABL.integral_pow_uniformIntervalMeasure (m : ℕ) : ∫ (x : ℝ), x ^ m ∂FABL.uniformIntervalMeasure = FABL.uniformIntervalMoment m
Integration against the uniform interval measure is normalized interval integration.
-
theoremdefined in FABL/Chapter09/ContinuousReasonableness.leancomplete
theorem FABL.uniformIntervalMoment_two : FABL.uniformIntervalMoment 2 = 1 / 3
theorem FABL.uniformIntervalMoment_two : FABL.uniformIntervalMoment 2 = 1 / 3
The uniform law on `[-1,1]` has second moment `1/3`.
-
theoremdefined in FABL/Chapter09/ContinuousReasonableness.leancomplete
theorem FABL.uniformIntervalMoment_four : FABL.uniformIntervalMoment 4 = 1 / 5
theorem FABL.uniformIntervalMoment_four : FABL.uniformIntervalMoment 4 = 1 / 5
The uniform law on `[-1,1]` has fourth moment `1/5`.
-
theoremdefined in FABL/Chapter09/ContinuousReasonableness.leancomplete
theorem FABL.uniformInterval_isNineFifthsReasonable : FABL.uniformIntervalMoment 4 ≤ 9 / 5 * FABL.uniformIntervalMoment 2 ^ 2
theorem FABL.uniformInterval_isNineFifthsReasonable : FABL.uniformIntervalMoment 4 ≤ 9 / 5 * FABL.uniformIntervalMoment 2 ^ 2
The normalized interval moments satisfy the sharp `9/5` reasonableness identity.
-
theoremdefined in FABL/Chapter09/ContinuousReasonableness.leancomplete
theorem FABL.uniformIntervalMeasure_isNineFifthsReasonable : FABL.IsReasonableMeasure FABL.uniformIntervalMeasure (9 / 5) fun x => x
theorem FABL.uniformIntervalMeasure_isNineFifthsReasonable : FABL.IsReasonableMeasure FABL.uniformIntervalMeasure (9 / 5) fun x => x
Example 9.2: the uniform probability law on `[-1,1]` is `9/5`-reasonable.
Proposition 9.3. If X\not\equiv0 is B-reasonable, then for every
t>0,
\Pr[|X|\ge t\|X\|_2]\le\frac{B}{t^4}.
Lean code for Theorem9.1.7●1 theorem
Associated Lean declarations
-
FABL.reasonable_tail_le[complete]
-
FABL.reasonable_tail_le[complete]
-
theoremdefined in FABL/Chapter09/Reasonableness.leancomplete
theorem FABL.reasonable_tail_le.{u_1} {Ω : Type u_1} [Fintype Ω] (p : PMF Ω) (B : ℝ) (X : Ω → ℝ) (hX : FABL.IsReasonable p B X) (hne : FABL.NonzeroUnder p X) (t : ℝ) (ht : 0 < t) : (FABL.pmfEventProbability p fun x => t * √(FABL.pmfExpectation p fun y => X y ^ 2) ≤ |X x|) ≤ B / t ^ 4
theorem FABL.reasonable_tail_le.{u_1} {Ω : Type u_1} [Fintype Ω] (p : PMF Ω) (B : ℝ) (X : Ω → ℝ) (hX : FABL.IsReasonable p B X) (hne : FABL.NonzeroUnder p X) (t : ℝ) (ht : 0 < t) : (FABL.pmfEventProbability p fun x => t * √(FABL.pmfExpectation p fun y => X y ^ 2) ≤ |X x|) ≤ B / t ^ 4
O'Donnell, Proposition 9.3: a reasonable random variable has a fourth-moment tail bound.
-
FABL.pmfPaleyZygmund[complete] -
FABL.reasonable_anticoncentration[complete]
Proposition 9.4. If X\not\equiv0 is B-reasonable, then for every
t\in[0,1],
\Pr[|X|>t\|X\|_2]
\ge\frac{(1-t^2)^2}{B}.
Lean code for Theorem9.1.8●2 theorems
Associated Lean declarations
-
FABL.pmfPaleyZygmund[complete]
-
FABL.reasonable_anticoncentration[complete]
-
FABL.pmfPaleyZygmund[complete] -
FABL.reasonable_anticoncentration[complete]
-
theoremdefined in FABL/Chapter09/Reasonableness.leancomplete
theorem FABL.pmfPaleyZygmund.{u_1} {Ω : Type u_1} [Fintype Ω] (p : PMF Ω) (Y : Ω → ℝ) (hY : ∀ (x : Ω), 0 ≤ Y x) (θ : ℝ) (hθ₀ : 0 ≤ θ) (hθ₁ : θ < 1) : (1 - θ) ^ 2 * FABL.pmfExpectation p Y ^ 2 ≤ (FABL.pmfEventProbability p fun x => θ * FABL.pmfExpectation p Y < Y x) * FABL.pmfExpectation p fun x => Y x ^ 2
theorem FABL.pmfPaleyZygmund.{u_1} {Ω : Type u_1} [Fintype Ω] (p : PMF Ω) (Y : Ω → ℝ) (hY : ∀ (x : Ω), 0 ≤ Y x) (θ : ℝ) (hθ₀ : 0 ≤ θ) (hθ₁ : θ < 1) : (1 - θ) ^ 2 * FABL.pmfExpectation p Y ^ 2 ≤ (FABL.pmfEventProbability p fun x => θ * FABL.pmfExpectation p Y < Y x) * FABL.pmfExpectation p fun x => Y x ^ 2
Finite-PMF Paley--Zygmund inequality, with the strict event used in Proposition 9.4.
-
theoremdefined in FABL/Chapter09/Reasonableness.leancomplete
theorem FABL.reasonable_anticoncentration.{u_1} {Ω : Type u_1} [Fintype Ω] (p : PMF Ω) (B : ℝ) (X : Ω → ℝ) (hX : FABL.IsReasonable p B X) (hne : FABL.NonzeroUnder p X) (t : ℝ) (ht : t ∈ Set.Icc 0 1) : (1 - t ^ 2) ^ 2 / B ≤ FABL.pmfEventProbability p fun x => t * √(FABL.pmfExpectation p fun y => X y ^ 2) < |X x|
theorem FABL.reasonable_anticoncentration.{u_1} {Ω : Type u_1} [Fintype Ω] (p : PMF Ω) (B : ℝ) (X : Ω → ℝ) (hX : FABL.IsReasonable p B X) (hne : FABL.NonzeroUnder p X) (t : ℝ) (ht : t ∈ Set.Icc 0 1) : (1 - t ^ 2) ^ 2 / B ≤ FABL.pmfEventProbability p fun x => t * √(FABL.pmfExpectation p fun y => X y ^ 2) < |X x|
O'Donnell, Proposition 9.4: a reasonable random variable has strict-event anticoncentration.
-
FABL.pmfPositiveSupport[complete] -
FABL.minimumPositiveAtom[complete] -
FABL.minimumPositiveAtom_pos[complete] -
FABL.reasonable_of_minimumPositiveAtom[complete]
Proposition 9.5. Let X be a discrete random variable with finite range
and probability mass function \pi. If
\lambda=\min_{x\in\operatorname{range}(X)}\Pr[X=x],
then X is 1/\lambda-reasonable.
Lean code for Theorem9.1.9●4 declarations
Associated Lean declarations
-
FABL.pmfPositiveSupport[complete]
-
FABL.minimumPositiveAtom[complete]
-
FABL.minimumPositiveAtom_pos[complete]
-
FABL.reasonable_of_minimumPositiveAtom[complete]
-
FABL.pmfPositiveSupport[complete] -
FABL.minimumPositiveAtom[complete] -
FABL.minimumPositiveAtom_pos[complete] -
FABL.reasonable_of_minimumPositiveAtom[complete]
-
defdefined in FABL/Chapter09/Reasonableness.leancomplete
def FABL.pmfPositiveSupport.{u_1} {Ω : Type u_1} [Fintype Ω] (p : PMF Ω) : Finset Ω
def FABL.pmfPositiveSupport.{u_1} {Ω : Type u_1} [Fintype Ω] (p : PMF Ω) : Finset Ω
The finite support of a PMF, excluding ambient values of mass zero.
-
defdefined in FABL/Chapter09/Reasonableness.leancomplete
def FABL.minimumPositiveAtom.{u_1} {Ω : Type u_1} [Fintype Ω] (p : PMF Ω) : ℝ
def FABL.minimumPositiveAtom.{u_1} {Ω : Type u_1} [Fintype Ω] (p : PMF Ω) : ℝ
The smallest positive atom of a PMF on a finite ambient type.
-
theoremdefined in FABL/Chapter09/Reasonableness.leancomplete
theorem FABL.minimumPositiveAtom_pos.{u_1} {Ω : Type u_1} [Fintype Ω] (p : PMF Ω) : 0 < FABL.minimumPositiveAtom p
theorem FABL.minimumPositiveAtom_pos.{u_1} {Ω : Type u_1} [Fintype Ω] (p : PMF Ω) : 0 < FABL.minimumPositiveAtom p
On a finite space, the minimum positive atom is strictly positive.
-
theoremdefined in FABL/Chapter09/Reasonableness.leancomplete
theorem FABL.reasonable_of_minimumPositiveAtom.{u_1} {Ω : Type u_1} [Fintype Ω] (p : PMF Ω) (X : Ω → ℝ) : FABL.IsReasonable p (1 / FABL.minimumPositiveAtom p) X
theorem FABL.reasonable_of_minimumPositiveAtom.{u_1} {Ω : Type u_1} [Fintype Ω] (p : PMF Ω) (X : Ω → ℝ) : FABL.IsReasonable p (1 / FABL.minimumPositiveAtom p) X
O'Donnell, Proposition 9.5: a finite discrete random variable is `1 / minimumPositiveAtom`-reasonable.
For the decomposition f=x_nd+e, where d=D_nf and e=E_nf do not
depend on x_n, independence and the sign moments give
\mathbb E[f^4]
=\mathbb E[d^4]+6\mathbb E[d^2e^2]+\mathbb E[e^4].
\tag{9.1}
Lean code for Lemma9.1.10●1 theorem
Associated Lean declarations
-
FABL.equation9_1_expect_fourth[complete]
-
FABL.equation9_1_expect_fourth[complete]
-
theoremdefined in FABL/Chapter09/Reasonableness.leancomplete
theorem FABL.equation9_1_expect_fourth {n : ℕ} (d e : FABL.SignCube n → ℝ) (f : FABL.SignCube (n + 1) → ℝ) (hf : ∀ (b : FABL.Sign) (x : Fin n → FABL.Sign), f (Fin.cons b x) = FABL.signValue b * d x + e x) : (Finset.univ.expect fun x => f x ^ 4) = ((Finset.univ.expect fun x => d x ^ 4) + 6 * Finset.univ.expect fun x => d x ^ 2 * e x ^ 2) + Finset.univ.expect fun x => e x ^ 4
theorem FABL.equation9_1_expect_fourth {n : ℕ} (d e : FABL.SignCube n → ℝ) (f : FABL.SignCube (n + 1) → ℝ) (hf : ∀ (b : FABL.Sign) (x : Fin n → FABL.Sign), f (Fin.cons b x) = FABL.signValue b * d x + e x) : (Finset.univ.expect fun x => f x ^ 4) = ((Finset.univ.expect fun x => d x ^ 4) + 6 * Finset.univ.expect fun x => d x ^ 2 * e x ^ 2) + Finset.univ.expect fun x => e x ^ 4
Equation (9.1), delegated to the canonical odd/even fourth-moment identity from the FKN development.
In the same decomposition,
\mathbb E[f^2]=\mathbb E[d^2]+\mathbb E[e^2].
\tag{9.2}
Lean code for Lemma9.1.11●1 theorem
Associated Lean declarations
-
FABL.equation9_2_expect_sq[complete]
-
FABL.equation9_2_expect_sq[complete]
-
theoremdefined in FABL/Chapter09/Reasonableness.leancomplete
theorem FABL.equation9_2_expect_sq {n : ℕ} (d e : FABL.SignCube n → ℝ) (f : FABL.SignCube (n + 1) → ℝ) (hf : ∀ (b : FABL.Sign) (x : Fin n → FABL.Sign), f (Fin.cons b x) = FABL.signValue b * d x + e x) : (Finset.univ.expect fun x => f x ^ 2) = (Finset.univ.expect fun x => d x ^ 2) + Finset.univ.expect fun x => e x ^ 2
theorem FABL.equation9_2_expect_sq {n : ℕ} (d e : FABL.SignCube n → ℝ) (f : FABL.SignCube (n + 1) → ℝ) (hf : ∀ (b : FABL.Sign) (x : Fin n → FABL.Sign), f (Fin.cons b x) = FABL.signValue b * d x + e x) : (Finset.univ.expect fun x => f x ^ 2) = (Finset.univ.expect fun x => d x ^ 2) + Finset.univ.expect fun x => e x ^ 2
Equation (9.2), delegated to the canonical odd/even slicing identity from the FKN development.
-
FABL.multilinearEvenCoefficients[complete] -
FABL.multilinearOddCoefficients[complete] -
FABL.productMultilinearPolynomial[complete] -
FABL.MultilinearDegreeAtMost[complete] -
FABL.productMultilinearRandomVariable[complete] -
FABL.reasonableProductPolynomial_isReasonable[complete]
Corollary 9.6. Let X_1,ldots,X_n be independent random variables
satisfying
\mathbb E[X_i]=\mathbb E[X_i^3]=0
and suppose every X_i is B-reasonable. If F is a multilinear
polynomial of degree at most k, then
F(X_1,ldots,X_n) is \max(B,9)^k-reasonable.
Lean code for Theorem9.1.12●6 declarations
Associated Lean declarations
-
FABL.multilinearEvenCoefficients[complete]
-
FABL.multilinearOddCoefficients[complete]
-
FABL.productMultilinearPolynomial[complete]
-
FABL.MultilinearDegreeAtMost[complete]
-
FABL.productMultilinearRandomVariable[complete]
-
FABL.reasonableProductPolynomial_isReasonable[complete]
-
FABL.multilinearEvenCoefficients[complete] -
FABL.multilinearOddCoefficients[complete] -
FABL.productMultilinearPolynomial[complete] -
FABL.MultilinearDegreeAtMost[complete] -
FABL.productMultilinearRandomVariable[complete] -
FABL.reasonableProductPolynomial_isReasonable[complete]
-
defdefined in FABL/Chapter09/ReasonableProductPolynomials.leancomplete
def FABL.multilinearEvenCoefficients {n : ℕ} (a : Finset (Fin (n + 1)) → ℝ) : Finset (Fin n) → ℝ
def FABL.multilinearEvenCoefficients {n : ℕ} (a : Finset (Fin (n + 1)) → ℝ) : Finset (Fin n) → ℝ
Coefficients of terms not containing the first variable.
-
defdefined in FABL/Chapter09/ReasonableProductPolynomials.leancomplete
def FABL.multilinearOddCoefficients {n : ℕ} (a : Finset (Fin (n + 1)) → ℝ) : Finset (Fin n) → ℝ
def FABL.multilinearOddCoefficients {n : ℕ} (a : Finset (Fin (n + 1)) → ℝ) : Finset (Fin n) → ℝ
Coefficients of terms containing the first variable, after removing that variable.
-
defdefined in FABL/Chapter09/ReasonableProductPolynomials.leancomplete
def FABL.productMultilinearPolynomial {n : ℕ} : (Finset (Fin n) → ℝ) → (Fin n → ℝ) → ℝ
def FABL.productMultilinearPolynomial {n : ℕ} : (Finset (Fin n) → ℝ) → (Fin n → ℝ) → ℝ
Evaluation of a multilinear polynomial from its square-free monomial coefficients. At arity `n + 1`, the definition is the canonical decomposition `x₀ * D + E`.
-
defdefined in FABL/Chapter09/ReasonableProductPolynomials.leancomplete
def FABL.MultilinearDegreeAtMost {n : ℕ} (a : Finset (Fin n) → ℝ) (k : ℕ) : Prop
def FABL.MultilinearDegreeAtMost {n : ℕ} (a : Finset (Fin n) → ℝ) (k : ℕ) : Prop
The coefficient-support formulation of degree at most `k`.
-
defdefined in FABL/Chapter09/ReasonableProductPolynomials.leancomplete
def FABL.productMultilinearRandomVariable.{u_1} {n : ℕ} {Ω : Fin n → Type u_1} (a : Finset (Fin n) → ℝ) (X : (i : Fin n) → Ω i → ℝ) (ω : (i : Fin n) → Ω i) : ℝ
def FABL.productMultilinearRandomVariable.{u_1} {n : ℕ} {Ω : Fin n → Type u_1} (a : Finset (Fin n) → ℝ) (X : (i : Fin n) → Ω i → ℝ) (ω : (i : Fin n) → Ω i) : ℝ
Evaluate a coefficient-form multilinear polynomial on coordinate random variables.
-
theoremdefined in FABL/Chapter09/ReasonableProductPolynomials.leancomplete
theorem FABL.reasonableProductPolynomial_isReasonable.{u_1} {n : ℕ} {Ω : Fin n → Type u_1} [(i : Fin n) → Fintype (Ω i)] (p : (i : Fin n) → PMF (Ω i)) (X : (i : Fin n) → Ω i → ℝ) (B : ℝ) (a : Finset (Fin n) → ℝ) (k : ℕ) (hmean : ∀ (i : Fin n), FABL.pmfExpectation (p i) (X i) = 0) (hthird : ∀ (i : Fin n), (FABL.pmfExpectation (p i) fun z => X i z ^ 3) = 0) (hreasonable : ∀ (i : Fin n), FABL.IsReasonable (p i) B (X i)) (hdegree : FABL.MultilinearDegreeAtMost a k) : FABL.IsReasonable (FABL.independentProductPMF p) (max B 9 ^ k) (FABL.productMultilinearRandomVariable a X)
theorem FABL.reasonableProductPolynomial_isReasonable.{u_1} {n : ℕ} {Ω : Fin n → Type u_1} [(i : Fin n) → Fintype (Ω i)] (p : (i : Fin n) → PMF (Ω i)) (X : (i : Fin n) → Ω i → ℝ) (B : ℝ) (a : Finset (Fin n) → ℝ) (k : ℕ) (hmean : ∀ (i : Fin n), FABL.pmfExpectation (p i) (X i) = 0) (hthird : ∀ (i : Fin n), (FABL.pmfExpectation (p i) fun z => X i z ^ 3) = 0) (hreasonable : ∀ (i : Fin n), FABL.IsReasonable (p i) B (X i)) (hdegree : FABL.MultilinearDegreeAtMost a k) : FABL.IsReasonable (FABL.independentProductPMF p) (max B 9 ^ k) (FABL.productMultilinearRandomVariable a X)
Corollary 9.6: a degree-at-most-`k` multilinear polynomial in independent centered, zero-third-moment, `B`-reasonable finite random variables is `max(B,9)^k`-reasonable.
Theorem 9.7. Let f:\{-1,1\}^n\to\mathbb R be nonconstant of degree
at most k, write \mu=\mathbb E[f], and let
\sigma^2=\operatorname{Var}[f]. Then
\Pr\left[|f(x)-\mu|>\frac12\sigma\right]
\ge\frac1{16}\,9^{-k}.
Lean code for Theorem9.1.13●1 theorem
Associated Lean declarations
-
FABL.lowDegree_anticoncentration[complete]
-
FABL.lowDegree_anticoncentration[complete]
-
theoremdefined in FABL/Chapter09/Reasonableness.leancomplete
theorem FABL.lowDegree_anticoncentration {n : ℕ} (f : FABL.SignCube n → ℝ) (k : ℕ) (hdegree : FABL.fourierDegree f ≤ k) (hf : ¬∃ c, ∀ (x : FABL.SignCube n), f x = c) : 9 / 16 / 9 ^ k ≤ FABL.uniformProbability fun x => 1 / 2 * √(FABL.variance f) < |f x - FABL.mean f|
theorem FABL.lowDegree_anticoncentration {n : ℕ} (f : FABL.SignCube n → ℝ) (k : ℕ) (hdegree : FABL.fourierDegree f ≤ k) (hf : ¬∃ c, ∀ (x : FABL.SignCube n), f x = c) : 9 / 16 / 9 ^ k ≤ FABL.uniformProbability fun x => 1 / 2 * √(FABL.variance f) < |f x - FABL.mean f|
O'Donnell, Theorem 9.7: a nonconstant degree-`k` function differs from its mean by more than half a standard deviation with probability at least `(9/16) / 9^k`. This is the book's `(1/16) * 9^(1-k)` constant, written without a negative natural exponent.
-
FABL.expect_centeredSingletonSquare_sq_le[complete] -
FABL.fkn_small[complete]
FKN Theorem, proof in Section 9.1. If
f:\{-1,1\}^n\to\{-1,1\} satisfies
W^1[f]=1-\delta, then f is O(\delta)-close to
\pm\chi_i for some coordinate i. This proof reuses the canonical
Chapter 2 declaration rather than creating a second public theorem.
Lean code for Theorem9.1.14●2 theorems
Associated Lean declarations
-
FABL.expect_centeredSingletonSquare_sq_le[complete]
-
FABL.fkn_small[complete]
-
FABL.expect_centeredSingletonSquare_sq_le[complete] -
FABL.fkn_small[complete]
-
theoremdefined in FABL/Chapter02/FKN.leancomplete
theorem FABL.expect_centeredSingletonSquare_sq_le {n : ℕ} (f : FABL.BooleanFunction n) (δ : ℝ) (hδ₀ : 0 ≤ δ) (hδ : δ ≤ 1 / 1600) (hweight : FABL.fourierWeightAtLevel 1 f.toReal = 1 - δ) : (Finset.univ.expect fun x => FABL.centeredLinearSquare (fun i => FABL.fourierCoeff f.toReal {i}) x ^ 2) ≤ 6400 * δ
theorem FABL.expect_centeredSingletonSquare_sq_le {n : ℕ} (f : FABL.BooleanFunction n) (δ : ℝ) (hδ₀ : 0 ≤ δ) (hδ : δ ≤ 1 / 1600) (hweight : FABL.fourierWeightAtLevel 1 f.toReal = 1 - δ) : (Finset.univ.expect fun x => FABL.centeredLinearSquare (fun i => FABL.fourierCoeff f.toReal {i}) x ^ 2) ≤ 6400 * δ
The main analytic estimate in the proof of FKN: the square of the level-one part has variance at most `6400 * δ`.
-
theoremdefined in FABL/Chapter02/FKN.leancomplete
theorem FABL.fkn_small {n : ℕ} (f : FABL.BooleanFunction n) (hn : 0 < n) (δ : ℝ) (hδ₀ : 0 ≤ δ) (hδ : δ ≤ 1 / 1600) (hweight : FABL.fourierWeightAtLevel 1 f.toReal = 1 - δ) : ∃ i, FABL.relativeHammingDist f (FABL.dictator i) ≤ 1601 * δ ∨ FABL.relativeHammingDist f (-FABL.dictator i) ≤ 1601 * δ
theorem FABL.fkn_small {n : ℕ} (f : FABL.BooleanFunction n) (hn : 0 < n) (δ : ℝ) (hδ₀ : 0 ≤ δ) (hδ : δ ≤ 1 / 1600) (hweight : FABL.fourierWeightAtLevel 1 f.toReal = 1 - δ) : ∃ i, FABL.relativeHammingDist f (FABL.dictator i) ≤ 1601 * δ ∨ FABL.relativeHammingDist f (-FABL.dictator i) ≤ 1601 * δ
The explicit small-error FKN theorem.