TCSLib

272 Boolean Analysis — Duality

272.1 Overview

The noise kernel has total mass one in either variable and contracts each norm of exponent at least one. Self-adjointness yields conjugate operator bounds and the hypercontractive estimate crossing exponent two.

272.2 Declarations

Lemma 272.1 Noise kernel rows sum to one
✓

Let \(n\) be a natural number, let \(0\le \rho \le 1\), and fix \(x\in \{ 0,1\} ^n\). The noise kernel \(K_\rho (x,y)=\prod _{i=1}^{n}\frac{1+\rho \, \sigma (x_i)\, \sigma (y_i)}{2}\), where \(\sigma (0)=1\) and \(\sigma (1)=-1\), satisfies

\[ \sum _{y\in \{ 0,1\} ^n}K_\rho (x,y)=1. \]
Lemma 272.2 Noise kernel sums to one over the cube
✓

Let \(n\) be a natural number, let \(0\le \rho \le 1\), and fix \(y\in \{ 0,1\} ^n\). The noise kernel \(K_\rho (x,y)=\prod _{i=1}^{n}\frac{1+\rho \, \sigma (x_i)\, \sigma (y_i)}{2}\), where \(\sigma (0)=1\) and \(\sigma (1)=-1\), satisfies

\[ \sum _{x\in \{ 0,1\} ^n}K_\rho (x,y)=1. \]

Thus \(K_\rho (\cdot ,y)\) is a probability distribution on the cube.

Let \(n\) be a natural number, and let \(T_\rho \) denote the noise operator with parameter \(\rho \in [0,1]\) acting on Boolean functions \(f : \{ 0,1\} ^n \to \mathbb {R}\). Then for every real exponent \(s \ge 1\) and every point \(x \in \{ 0,1\} ^n\),

\[ \left\lvert (T_\rho f)(x)\right\rvert ^{s} \; \le \; \sum _{y \in \{ 0,1\} ^n} K_\rho (x,y)\, \left\lvert f(y)\right\rvert ^{s}, \]

where \(K_\rho (x,y) = \prod _{i} \tfrac {1 + \rho \, \sigma (x_i)\, \sigma (y_i)}{2}\) is the noise kernel, with \(\sigma (0)=1\) and \(\sigma (1)=-1\). For \(d\in \mathbb N\), \(h:\{ 0,1\} ^d\to \mathbb {R}\), and \(\tau \in \mathbb {R}\), the spectral noise operator is \(T_\tau h(x)=\sum _{S\subseteq [d]}\tau ^{|S|}\widehat h(S)\chi _S(x)\), where \(\chi _S(x)=(-1)^{\sum _{i\in S}x_i}\) and \(\widehat h(S)=\mathbb {E}[h\chi _S]\) under the uniform measure.

Lemma 272.4 Contractivity of the noise operator in \(L^s\)
✓

Let \(n\) be a natural number, let \(s \ge 1\) and \(\rho \in [0,1]\), and let \(f : \{ 0,1\} ^n \to \mathbb {R}\) be a Boolean function. Then, with expectations taken under the uniform measure on the cube, the noise operator \(T_\rho \) satisfies

\[ \bigl(\mathbb {E}\bigl[\left\lvert T_\rho f\right\rvert ^{s}\bigr]\bigr)^{1/s} \; \le \; \bigl(\mathbb {E}\bigl[\left\lvert f\right\rvert ^{s}\bigr]\bigr)^{1/s}. \]

For \(d\in \mathbb N\), \(h:\{ 0,1\} ^d\to \mathbb {R}\), and \(\tau \in \mathbb {R}\), the spectral noise operator is \(T_\tau h(x)=\sum _{S\subseteq [d]}\tau ^{|S|}\widehat h(S)\chi _S(x)\), where \(\chi _S(x)=(-1)^{\sum _{i\in S}x_i}\) and \(\widehat h(S)=\mathbb {E}[h\chi _S]\) under the uniform measure.

Fix a natural number \(n\), exponents \(p,q\gt 1\) and a noise parameter \(\rho \) with \(0 \le \rho \le 1\), and write \(p' = p/(p-1)\) and \(q' = q/(q-1)\) for the Hölder conjugates. Then the noise operator \(T_\rho \) on Boolean functions \(f : \{ 0,1\} ^n \to \mathbb {R}\) satisfies the \((p \to q)\) bound

\[ \bigl(\mathbb {E}\bigl[\, \left\lvert T_\rho f\right\rvert ^q\, \bigr]\bigr)^{1/q} \; \le \; \bigl(\mathbb {E}\bigl[\, \left\lvert f\right\rvert ^p\, \bigr]\bigr)^{1/p} \qquad \text{for every } f \]

if and only if it satisfies the conjugate \((q' \to p')\) bound

\[ \bigl(\mathbb {E}\bigl[\, \left\lvert T_\rho f\right\rvert ^{p'}\, \bigr]\bigr)^{1/p'} \; \le \; \bigl(\mathbb {E}\bigl[\, \left\lvert f\right\rvert ^{q'}\, \bigr]\bigr)^{1/q'} \qquad \text{for every } f, \]

where all norms are taken with respect to the uniform measure on \(\{ 0,1\} ^n\). For \(d\in \mathbb N\), \(h:\{ 0,1\} ^d\to \mathbb {R}\), and \(\tau \in \mathbb {R}\), the spectral noise operator is \(T_\tau h(x)=\sum _{S\subseteq [d]}\tau ^{|S|}\widehat h(S)\chi _S(x)\), where \(\chi _S(x)=(-1)^{\sum _{i\in S}x_i}\) and \(\widehat h(S)=\mathbb {E}[h\chi _S]\) under the uniform measure.

Lemma 272.6 Square root of a ratio bounded by one
✓

Let \(a\) and \(b\) be real numbers with \(0 \le a\), \(0 \lt b\), and \(a \le b\). Then \(\sqrt{a/b} \le 1\).

Lemma 272.7 Noise-parameter identity
✓

Let \(p, u \in \mathbb {R}\) with \(u \gt 1\). Then

\[ \Bigl(\frac{u}{u-1} - 1\Bigr)(p-1) \; =\; \frac{p-1}{u-1}. \]

Fix a natural number \(n\). Let \(f:\{ 0,1\} ^n\to \mathbb {R}\) be a Boolean function, and let \(p,u\in \mathbb {R}\) satisfy \(1\le p\le 2\le u\). Set \(\rho =\sqrt{(p-1)/(u-1)}\) and let \(T_\rho \) denote the noise operator with parameter \(\rho \). Then

\[ \bigl(\mathbb {E}\bigl[\, \left\lvert T_\rho f\right\rvert ^{\, u}\, \bigr]\bigr)^{1/u} \; \le \; \bigl(\mathbb {E}\bigl[\, \left\lvert f\right\rvert ^{\, p}\, \bigr]\bigr)^{1/p}, \]

where the expectations are taken with respect to the uniform measure on \(\{ 0,1\} ^n\). For \(d\in \mathbb N\), \(h:\{ 0,1\} ^d\to \mathbb {R}\), and \(\tau \in \mathbb {R}\), the spectral noise operator is \(T_\tau h(x)=\sum _{S\subseteq [d]}\tau ^{|S|}\widehat h(S)\chi _S(x)\), where \(\chi _S(x)=(-1)^{\sum _{i\in S}x_i}\) and \(\widehat h(S)=\mathbb {E}[h\chi _S]\) under the uniform measure.

Lemma 272.9 Interpolation parameters for the low-norms case
✓

Let \(p, u, \rho \) be real numbers with \(1 \lt p \lt u \lt 2\), and suppose \(\rho ^2 = (p-1)/(u-1)\). Then there exist real numbers \(\theta \in (0,1)\) and \(s \gt 0\) such that

\[ \frac{1}{p} = \frac{\theta }{1+\rho ^2} + \frac{1-\theta }{s} \qquad \text{and}\qquad \frac{1}{u} = \frac{\theta }{2} + \frac{1-\theta }{s}. \]