TCSLib

328 Boolean Analysis — Majority

328.1 Overview

This chapter records exact Fourier coefficients, relations between Fourier levels, and dimensional monotonicity and asymptotics for majority on an odd number of variables. Boolean inputs are expressed as signs, with false encoded by \(1\) and true by \(-1\); majority returns the sign held by more than half of the coordinates, so no tie occurs.

328.2 Declarations

Definition 328.1 Limiting majority level weights
✓

For each \(k\in \mathbb N\), define the real number

\[ L_k=\begin{cases} \displaystyle \frac{4}{\pi k2^k}\binom {k-1}{(k-1)/2},& k\text{ odd},\\ 0,& k\text{ even}. \end{cases} \]
Lemma 328.2 Majority takes sign values
✓

Let \(m\in \mathbb N\), put \(n=2m+1\), and define \(M_m:\{ \pm 1\} ^n\to \mathbb R\) by \(M_m(x)=1\) when more than \(m\) coordinates equal \(1\), and \(M_m(x)=-1\) otherwise. Then \(M_m(x)\in \{ 1,-1\} \) for every \(x\in \{ \pm 1\} ^n\).

Lemma 328.3 Oddness of majority
✓

Let \(m\in \mathbb N\), put \(n=2m+1\), and define \(M_m:\{ \pm 1\} ^n\to \mathbb R\) by \(M_m(x)=1\) when more than \(m\) coordinates equal \(1\), and \(M_m(x)=-1\) otherwise. Then \(M_m(-x)=-M_m(x)\) for every \(x\in \{ \pm 1\} ^n\), where \(-x\) negates every coordinate.

Lemma 328.4 Nonnegative Fourier level weights
✓

Let \(n\in \mathbb N\) and \(f:\{ \pm 1\} ^n\to \mathbb R\); for \(S\subseteq \{ 1,\ldots ,n\} \) define \(\widehat f(S)=2^{-n}\sum _{x\in \{ \pm 1\} ^n}f(x)\prod _{i\in S}x_i\), and for every \(k\in \mathbb N\) define \(W_k(f)=\sum _{|S|=k}\widehat f(S)^2\). Then \(W_k(f)\ge 0\) for every \(k\in \mathbb N\).

Lemma 328.5 Vanishing levels above the dimension
✓

Let \(n\in \mathbb N\) and \(f:\{ \pm 1\} ^n\to \mathbb R\); for \(S\subseteq \{ 1,\ldots ,n\} \) define \(\widehat f(S)=2^{-n}\sum _{x\in \{ \pm 1\} ^n}f(x)\prod _{i\in S}x_i\), and for every \(k\in \mathbb N\) define \(W_k(f)=\sum _{|S|=k}\widehat f(S)^2\). Then \(W_k(f)=0\) for every \(k\in \mathbb N\) with \(n\lt k\).

Lemma 328.6 Fourier sums grouped by degree
✓

Let \(n\in \mathbb N\) and \(f:\{ \pm 1\} ^n\to \mathbb R\); for \(S\subseteq \{ 1,\ldots ,n\} \) define \(\widehat f(S)=2^{-n}\sum _{x\in \{ \pm 1\} ^n}f(x)\prod _{i\in S}x_i\), and for every \(k\in \mathbb N\) define \(W_k(f)=\sum _{|S|=k}\widehat f(S)^2\). For every \(g:\mathbb N\to \mathbb R\),

\[ \sum _{S\subseteq \{ 1,\ldots ,n\} }g(|S|)\widehat f(S)^2 =\sum _{k=0}^n g(k)W_k(f). \]
Lemma 328.7 Stability generating function
✓

Let \(n\in \mathbb N\), \(\rho \in \mathbb R\), and \(f:\{ \pm 1\} ^n\to \mathbb R\); for \(S\subseteq \{ 1,\ldots ,n\} \) define \(\chi _S(x)=\prod _{i\in S}x_i\), \(\widehat f(S)=2^{-n}\sum _{x\in \{ \pm 1\} ^n}f(x)\chi _S(x)\), \(W_k(f)=\sum _{|S|=k}\widehat f(S)^2\) for \(k\in \mathbb N\), and \(T_\rho f(x)=\sum _{S\subseteq \{ 1,\ldots ,n\} }\rho ^{|S|}\widehat f(S)\chi _S(x)\). The noise stability \(\operatorname {Stab}_\rho (f)=2^{-n}\sum _{x\in \{ \pm 1\} ^n}f(x)T_\rho f(x)\) satisfies

\[ \operatorname {Stab}_\rho (f)=\sum _{k=0}^n\rho ^kW_k(f). \]
Lemma 328.8 Unit total Fourier weight
✓

Let \(n\in \mathbb N\) and \(f:\{ \pm 1\} ^n\to \mathbb R\) satisfy \(f(x)\in \{ 1,-1\} \) for every \(x\); for \(S\subseteq \{ 1,\ldots ,n\} \) define \(\widehat f(S)=2^{-n}\sum _{x\in \{ \pm 1\} ^n}f(x)\prod _{i\in S}x_i\) and \(W_k(f)=\sum _{|S|=k}\widehat f(S)^2\) for \(k\in \mathbb N\). Then \(\sum _{k=0}^n W_k(f)=1\).

Lemma 328.9 Prefix domination for geometric weights
✓

Let \(N\in \mathbb N\), \(\rho \in [0,1]\), and \(a,b:\mathbb N\to \mathbb R\) satisfy \(\sum _{k=0}^{j-1}b(k)\le \sum _{k=0}^{j-1}a(k)\) for every integer \(0\le j\le N\) and \(\sum _{k=0}^{N-1}b(k)=\sum _{k=0}^{N-1}a(k)\), with sums over no indices interpreted as zero. Then

\[ \sum _{k=0}^{N-1}\rho ^k b(k)\le \sum _{k=0}^{N-1}\rho ^k a(k). \]
Theorem 328.10 Vanishing even majority coefficients
✓

Let \(m\in \mathbb N\), put \(n=2m+1\), and define \(M_m:\{ \pm 1\} ^n\to \mathbb R\) by \(M_m(x)=1\) when more than \(m\) coordinates equal \(1\), and \(M_m(x)=-1\) otherwise. For \(S\subseteq \{ 1,\ldots ,n\} \) define \(\widehat M_m(S)=2^{-n}\sum _{x\in \{ \pm 1\} ^n}M_m(x)\prod _{i\in S}x_i\). If \(S\subseteq \{ 1,\ldots ,n\} \) has even cardinality, then \(\widehat M_m(S)=0\).

Theorem 328.11 Exact odd majority coefficients
✓

Let \(m\in \mathbb N\), put \(n=2m+1\), and define \(M_m:\{ \pm 1\} ^n\to \mathbb R\) by \(M_m(x)=1\) when more than \(m\) coordinates equal \(1\), and \(M_m(x)=-1\) otherwise. For \(S\subseteq \{ 1,\ldots ,n\} \) define \(\widehat M_m(S)=2^{-n}\sum _{x\in \{ \pm 1\} ^n}M_m(x)\prod _{i\in S}x_i\). If \(S\subseteq \{ 1,\ldots ,n\} \) has cardinality \(2r+1\) for an integer \(r\ge 0\), then

\[ \widehat M_m(S)=(-1)^r\frac{\binom mr}{\binom {2m}{2r}} \frac{\binom {2m}{m}}{2^{2m}}. \]
Lemma 328.12 Symmetry of binomial ratios
✓

For all \(m,p,q\in \mathbb N\) with \(p+q=m\), the binomial coefficients satisfy the real identity

\[ \frac{\binom mq}{\binom {2m}{2q}} =\frac{\binom mp}{\binom {2m}{2p}}. \]
Theorem 328.13 Duality of majority coefficients
✓

Let \(m\in \mathbb N\), put \(n=2m+1\), and define \(M_m:\{ \pm 1\} ^n\to \mathbb R\) by \(M_m(x)=1\) when more than \(m\) coordinates equal \(1\), and \(M_m(x)=-1\) otherwise. For \(S\subseteq \{ 1,\ldots ,n\} \) define \(\widehat M_m(S)=2^{-n}\sum _{x\in \{ \pm 1\} ^n}M_m(x)\prod _{i\in S}x_i\). If \(S,T\subseteq \{ 1,\ldots ,n\} \) satisfy \(|S|+|T|=2m+2\), then

\[ \widehat M_m(S)=(-1)^m\widehat M_m(T). \]
Lemma 328.14 Vanishing even majority weights
✓

Let \(m\in \mathbb N\), put \(n=2m+1\), and define \(M_m:\{ \pm 1\} ^n\to \mathbb R\) by \(M_m(x)=1\) when more than \(m\) coordinates equal \(1\), and \(M_m(x)=-1\) otherwise. For \(S\subseteq \{ 1,\ldots ,n\} \) define \(\widehat M_m(S)=2^{-n}\sum _{x\in \{ \pm 1\} ^n}M_m(x)\prod _{i\in S}x_i\), and for every \(k\in \mathbb N\) define \(W_k(M_m)=\sum _{|S|=k}\widehat M_m(S)^2\). If \(k\in \mathbb N\) is even, then \(W_k(M_m)=0\).

Lemma 328.15 Exact odd majority level weights
✓

Let \(m\in \mathbb N\), put \(n=2m+1\), and define \(M_m:\{ \pm 1\} ^n\to \mathbb R\) by \(M_m(x)=1\) when more than \(m\) coordinates equal \(1\), and \(M_m(x)=-1\) otherwise. For \(S\subseteq \{ 1,\ldots ,n\} \) define \(\widehat M_m(S)=2^{-n}\sum _{x\in \{ \pm 1\} ^n}M_m(x)\prod _{i\in S}x_i\), and for every \(k\in \mathbb N\) define \(W_k(M_m)=\sum _{|S|=k}\widehat M_m(S)^2\). For every odd \(k\in \mathbb N\), put \(r=(k-1)/2\); then

\[ W_k(M_m)=\binom {2m+1}{k} \left(\frac{\binom mr}{\binom {2m}{k-1}} \frac{\binom {2m}{m}}{2^{2m}}\right)^2, \]

where \(\binom ab=0\) for \(b\gt a\) and a real quotient with zero denominator is interpreted as zero.

Theorem 328.16 Duality of majority level weights
✓

Let \(m\in \mathbb N\), put \(n=2m+1\), and define \(M_m:\{ \pm 1\} ^n\to \mathbb R\) by \(M_m(x)=1\) when more than \(m\) coordinates equal \(1\), and \(M_m(x)=-1\) otherwise. For \(S\subseteq \{ 1,\ldots ,n\} \) define \(\widehat M_m(S)=2^{-n}\sum _{x\in \{ \pm 1\} ^n}M_m(x)\prod _{i\in S}x_i\), and for every \(k\in \mathbb N\) define \(W_k(M_m)=\sum _{|S|=k}\widehat M_m(S)^2\). For every \(k\in \mathbb N\) with \(k\le 2m+1\),

\[ W_{2m+2-k}(M_m)=\frac{k}{2m+2-k}W_k(M_m). \]
Lemma 328.17 Two-step binomial recurrence
✓

For all \(N,k,a\in \mathbb N\) satisfying \(N=k+a\), the binomial coefficients satisfy

\[ \binom {N+2}{k}(a+2)(a+1)=\binom Nk(N+2)(N+1). \]
Theorem 328.18 Strict decrease of odd level weights
✓

For each \(q\in \mathbb N\), define \(M_q:\{ \pm 1\} ^{2q+1}\to \mathbb R\) to equal \(1\) when more than \(q\) coordinates equal \(1\), and \(-1\) otherwise; for \(S\subseteq \{ 1,\ldots ,2q+1\} \) define \(\widehat M_q(S)=2^{-(2q+1)}\sum _{x\in \{ \pm 1\} ^{2q+1}}M_q(x)\prod _{i\in S}x_i\), and for \(k\in \mathbb N\) define \(W_k(M_q)=\sum _{|S|=k}\widehat M_q(S)^2\). For every \(m,k\in \mathbb N\) with \(k\) odd and \(k\le 2m+1\), one has \(W_k(M_{m+1})\lt W_k(M_m)\).

Theorem 328.19 Limit and finite-dimensional error bound
✓

For each \(q\in \mathbb N\), define \(M_q:\{ \pm 1\} ^{2q+1}\to \mathbb R\) to equal \(1\) when more than \(q\) coordinates equal \(1\), and \(-1\) otherwise; for \(S\subseteq \{ 1,\ldots ,2q+1\} \) define \(\widehat M_q(S)=2^{-(2q+1)}\sum _{x\in \{ \pm 1\} ^{2q+1}}M_q(x)\prod _{i\in S}x_i\), and for \(k\in \mathbb N\) define \(W_k(M_q)=\sum _{|S|=k}\widehat M_q(S)^2\). For each fixed odd integer \(k\ge 1\), set \(L_k=\frac{4}{\pi k2^k}\binom {k-1}{(k-1)/2}\); then \(\lim _{m\to \infty }W_k(M_m)=L_k\), and for every \(m\in \mathbb N\) satisfying \(k\le \lfloor (2m+1)/2\rfloor \) one has

\[ L_k\le W_k(M_m)\le \left(1+\frac{2k}{2m+1}\right)L_k. \]
Lemma 328.20 Decrease of cumulative majority weights
✓

For each \(q\in \mathbb N\), define \(M_q:\{ \pm 1\} ^{2q+1}\to \mathbb R\) to equal \(1\) when more than \(q\) coordinates equal \(1\), and \(-1\) otherwise; for \(S\subseteq \{ 1,\ldots ,2q+1\} \) define \(\widehat M_q(S)=2^{-(2q+1)}\sum _{x\in \{ \pm 1\} ^{2q+1}}M_q(x)\prod _{i\in S}x_i\), and for \(k\in \mathbb N\) define \(W_k(M_q)=\sum _{|S|=k}\widehat M_q(S)^2\). For every \(m,j\in \mathbb N\) satisfying \(j\le 2(m+1)+2\),

\[ \sum _{k=0}^{j-1}W_k(M_{m+1})\le \sum _{k=0}^{j-1}W_k(M_m), \]

where both sums are zero when \(j=0\).

Theorem 328.21 Decrease of majority noise stability
✓

Let \(m\in \mathbb N\) and \(\rho \in [0,1]\), and for each \(q\in \mathbb N\) define \(M_q:\{ \pm 1\} ^{2q+1}\to \mathbb R\) to equal \(1\) when more than \(q\) coordinates equal \(1\), and \(-1\) otherwise. For any \(d\in \mathbb N\) and \(f:\{ \pm 1\} ^d\to \mathbb R\), define \(\operatorname {Stab}_\rho (f)=\mathbb E[f(X)f(Y)]\), where \(X\) is uniform on \(\{ \pm 1\} ^d\) and, independently in each coordinate, \(Y_i=-X_i\) with probability \((1-\rho )/2\) and \(Y_i=X_i\) otherwise; then

\[ \operatorname {Stab}_\rho (M_{m+1})\le \operatorname {Stab}_\rho (M_m). \]
Theorem 328.22 Uniform majority spectral asymptotics
✓

For each \(m\in \mathbb N\), define \(M_m:\{ \pm 1\} ^{2m+1}\to \mathbb R\) to equal \(1\) when more than \(m\) coordinates equal \(1\), and \(-1\) otherwise; for \(S\subseteq \{ 1,\ldots ,2m+1\} \) define \(\widehat M_m(S)=2^{-(2m+1)}\sum _{x\in \{ \pm 1\} ^{2m+1}}M_m(x)\prod _{i\in S}x_i\), \(W_k(M_m)=\sum _{|S|=k}\widehat M_m(S)^2\), and \(W_{\gt k}(M_m)=\sum _{|S|\gt k}\widehat M_m(S)^2\) for \(k\in \mathbb N\), and put \(A=(2/\pi )^{3/2}\). There exists one real constant \(C\ge 0\) such that for every \(m,k\in \mathbb N\) with \(k\) odd, \(k\gt 0\), and \(2k^2\le 2m+1\), both estimates hold:

\[ \left|\frac{W_k(M_m)}{A k^{-3/2}}-1\right|\le \frac Ck, \qquad \left|\frac{W_{\gt k}(M_m)}{A k^{-1/2}}-1\right|\le \frac Ck. \]