Topics in Advanced AlgorithmsFall 2026

Chapter 3

Fourier Analysis of Boolean Functions

Arrow’s Theorem and the Power of Small Coalitions

How can a group combine individual inputs into a sensible collective decision while sharing power? This chapter studies two concrete versions of that question and develops the Fourier analysis of Boolean functions to answer them.

In our first problem, each of nn voters ranks three candidates from most to least preferred, with no ties, and we want to aggregate their ballots into a single transitive ranking that respects unanimous preferences. A natural procedure is to apply one binary comparison rule f:{−1,1}n→{−1,1}f:\{-1,1\}^n\to\{-1,1\} to each pair of candidates—AA versus BB, BB versus CC, and CC versus AA—and assemble the three pairwise winners. As Condorcet observed for majority rule, three transitive ballots can still produce a cyclic collective outcome (A>B>C>AA>B>C>A). Is this flaw specific to majority, or can a cleverer rule avoid cycles altogether? Arrow’s theorem, proved here in the three-candidate common-rule setting, shows a stark rigidity: any unanimous pairwise procedure that produces a valid ranking on every profile must be a dictatorship (f(x)=xif(x)=x_i), concentrating all decision power in a single fixed voter.

If demanding universal consistency of rankings forces all power into one person’s hands, what happens when we drop the ranking structure and try to spread power as evenly as possible across a single binary decision? In our second problem, honest participants broadcast independent fair signs to generate a shared random bit f(X)∈{−1,1}f(X)\in\{-1,1\}, while a small dishonest coalition waits to see the honest broadcasts before choosing its own signs. Even when no single participant has much say over the outcome, the Kahn–Kalai–Linial (KKL) theorem shows that some coalition of sublinear size O(n/log⁡n)O(n/\log n) can still force the output with probability 0.990.99.

Despite their different models, both results rest on the same analytic foundation: expanding a Boolean rule in the orthonormal basis of product characters χS(x)=∏i∈Sxi\chi_S(x)=\prod_{i\in S}x_i and tracking how its Fourier spectrum responds to coordinatewise noise TρT_\rho. In the first meeting, negative correlation (ρ=−1/3\rho=-1/3) arises directly from the transitivity of individual rankings and turns cycle-freeness into an exact spectral constraint. In the second meeting, positive noise (ρ=1/3\rho=1/\sqrt3) serves as a smoothing operator whose hypercontractive inequality prevents rare pivotal events from hiding variance on low Fourier degrees.

Reading guide. Course Content is organized into two 120-minute meetings. In the first meeting, §§1.1–1.6 develop the Fourier expansion of Boolean functions, concrete coefficient calculations, and noise stability to prove the classroom version of Arrow’s theorem. In the second meeting, §§1.7–1.16 turn to binary collective coin flipping, individual influence, and coalition manipulation: §§1.7–1.10 formulate the attack and show why elementary Fourier bounds stall at 1/n1/n, §§1.11–1.14 isolate the missing low-degree rarity estimate and prove it via noise smoothing and hypercontractivity, and §§1.15–1.16 establish the KKL theorem and the sublinear coalition attack.

Reading Contents preserves Blum–Luby–Rubinfeld (BLR) linearity testing and local correction as a self-contained algorithmic application of the character basis, followed by supporting calculations and extensions for both classroom stories. Further Topics gives background, main statements, and references for five continuations, including robust Arrow via the Friedgut–Kalai–Naor (FKN) theorem.

We assume basic probability and linear algebra, including the Cauchy–Schwarz inequality; the specific finite-space Hölder inequality used in §1.12 is stated and proved in Reading Contents. Throughout Course Content, inputs are sign vectors x∈{−1,1}nx\in\{-1,1\}^n, Boolean outputs lie in {−1,1}\{-1,1\}, expectations are taken under the uniform product measure unless another distribution is specified, and logarithms are natural unless a base is displayed. Reading Contents §2.1 uses bits u∈{0,1}nu\in\{0,1\}^n for linear maps over F2={0,1}\mathbb F_2=\{0,1\} with addition modulo two, related to signs by xi=(−1)uix_i=(-1)^{u_i}.

I. Course Content

1.1 A mathematical model of voting and collective rankings

We want to combine the preferences of nn voters into one collective ranking of three candidates. Before testing any specific voting rule, let us specify the legal inputs, the desired output space, and the pairwise aggregation procedure we will analyze.

Ballots, profiles, and the desired output

Fix an integer n≥1n\ge1. The voters are labelled by [n]={1,…,n}[n]=\{1,\ldots,n\}, and the candidate set is C={A,B,C}\mathcal C=\{A,B,C\}.

A strict ranking is an ordering of all three candidates from most to least preferred, with no ties. In relation notation, voter ii’s preference ≻i\succ_i is irreflexive (U̸≻iUU\not\succ_i U for each candidate UU), compares every two distinct candidates in exactly one direction, and is transitive: U≻iV and V≻iW⟹U≻iW. U\succ_i V\text{ and }V\succ_i W\quad\Longrightarrow\quad U\succ_i W. Write A>B>CA>B>C for the ranking that places AA first, BB second, and CC third. There are 3!=63!=6 such rankings. Denote their set by L(C)\mathcal L(\mathcal C).

Definition 1 (Ballot and preference profile) Voter ii submits a ballot πi∈L(C)\pi_i\in\mathcal L(\mathcal C), consisting of a complete strict ranking of the candidates. A preference profile is the ordered tuple of all nn ballots, π=(π1,…,πn)∈L(C)n. \boldsymbol\pi=(\pi_1,\ldots,\pi_n)\in\mathcal L(\mathcal C)^n. Every such profile is allowed, giving 6n6^n possible profiles in total.

The desired collective output is another strict ranking of the three candidates. A procedure that always supplies such an output is a social ranking rule, a map F:L(C)n⟶L(C). \mathsf F:\mathcal L(\mathcal C)^n\longrightarrow\mathcal L(\mathcal C). Its input is the entire profile; its output tells society how to rank all three candidates. A natural way to build such a rule is to decide each pair of candidates separately. However, three separate pairwise decisions can form a cycle, so we must distinguish the raw pairwise output from the set of valid rankings L(C)\mathcal L(\mathcal C).

Turn each ballot into three binary comparisons

For a fixed profile, we record voter ii’s pairwise preferences along the cyclic orientations AA versus BB, BB versus CC, and CC versus AA: xi={+1,A≻iB,−1,B≻iA,yi={+1,B≻iC,−1,C≻iB,zi={+1,C≻iA,−1,A≻iC. \begin{aligned} x_i&=\begin{cases}+1,&A\succ_i B,\\-1,&B\succ_i A,\end{cases} &y_i&=\begin{cases}+1,&B\succ_i C,\\-1,&C\succ_i B,\end{cases} &z_i&=\begin{cases}+1,&C\succ_i A,\\-1,&A\succ_i C.\end{cases} \end{aligned} Orienting the three pairs cyclically (A→B→C→AA\to B\to C\to A) makes the forbidden cycles easy to spot: (xi,yi,zi)=(+1,+1,+1)(x_i,y_i,z_i)=(+1,+1,+1) means A≻iB≻iC≻iAA\succ_i B\succ_i C\succ_i A, while (−1,−1,−1)(-1,-1,-1) gives the reverse cycle A≺iB≺iC≺iAA\prec_i B\prec_i C\prec_i A. Neither can occur for a transitive voter. Checking all 3!=63!=6 rankings shows that the remaining six sign triples all occur:

A > B > C A B C (+1, +1, −1) Cycle A B C (+1, +1, +1)

Orienting pairs along A→B→C→AA\to B\to C\to A turns the two directed cycles into the constant triples (+1,+1,+1)(+1,+1,+1) and (−1,−1,−1)(-1,-1,-1).

From Condorcet to Arrow and Kalai. Nicolas de Condorcet (Essai sur l’application de l’analyse à la probabilité des décisions rendues à la pluralité des voix, 1785) observed that pairwise majority voting can cycle even when every voter’s ranking is transitive. Kenneth Arrow (1950; Social Choice and Individual Values, 1951) proved that no non-dictatorial rule satisfies unanimity and independence of irrelevant alternatives. Georges-Théodule Guilbaud (1952) calculated the uniform-profile cycle probability of majority rule, and Gil Kalai (2002) showed that Fourier analysis on {−1,1}n\{-1,1\}^n turns Arrow’s theorem into a spectral identity.

Individual ranking xix_i yiy_i ziz_i
A>B>CA>B>C +1+1 +1+1 −1-1
A>C>BA>C>B +1+1 −1-1 −1-1
B>A>CB>A>C −1-1 +1+1 −1-1
B>C>AB>C>A −1-1 +1+1 +1+1
C>A>BC>A>B +1+1 −1-1 +1+1
C>B>AC>B>A −1-1 −1-1 +1+1

Thus strict rankings correspond bijectively to the six non-constant sign triples: R={−1,1}3∖{(+1,+1,+1),(−1,−1,−1)}. \mathcal R=\{-1,1\}^3\setminus\{(+1,+1,+1),(-1,-1,-1)\}.

Collecting the comparisons across all nn voters gives three column vectors in {−1,1}n\{-1,1\}^n: x=(x1,…,xn),y=(y1,…,yn),z=(z1,…,zn). x=(x_1,\ldots,x_n),\qquad y=(y_1,\ldots,y_n),\qquad z=(z_1,\ldots,z_n). Equivalently, a legal profile π∈L(C)n\boldsymbol\pi\in\mathcal L(\mathcal C)^n is an n×3n\times 3 sign matrix with columns x,y,zx,y,z whose iith row (xi,yi,zi)(x_i,y_i,z_i) belongs to R\mathcal R for every voter i∈[n]i\in[n].

The pairwise procedure and its requirements

A Boolean comparison rule is a function f:{−1,1}n→{−1,1}f:\{-1,1\}^n\to\{-1,1\}. Its input is the column of nn voters’ signs for one pair of candidates; output +1+1 means society prefers the first candidate in that pair’s stated orientation, and −1-1 means society prefers the second.

Applying the same rule ff to all three comparison columns defines the pairwise aggregation procedure Hf:L(C)n⟶{−1,1}3,Hf(π)=(f(x),f(y),f(z)). \mathsf H_f:\mathcal L(\mathcal C)^n\longrightarrow\{-1,1\}^3, \qquad \mathsf H_f(\boldsymbol\pi)=(f(x),f(y),f(z)). Its codomain {−1,1}3\{-1,1\}^3 has eight triples. Whenever Hf(π)∈R\mathsf H_f(\boldsymbol\pi)\in\mathcal R, the encoding table translates the triple back into a unique social ranking in L(C)\mathcal L(\mathcal C). When Hf(π)\mathsf H_f(\boldsymbol\pi) is (+1,+1,+1)(+1,+1,+1) or (−1,−1,−1)(-1,-1,-1), the pairwise decisions form a cycle and fail to define a ranking.

By construction, Hf\mathsf H_f satisfies pairwise independence of irrelevant alternatives (IIA): society’s preference between any two candidates depends only on how the voters compare that specific pair (for instance, the collective preference between AA and BB is f(x)f(x), which depends only on column xx). Using the same Boolean function ff across all three oriented pairs is an additional symmetry assumption—the common-rule restriction—that keeps our classroom model clean; the general Arrow theorem allows three separate rules fAB,fBC,fCAf_{AB},f_{BC},f_{CA}.

Definition 2 (The design requirements) The common-rule procedure Hf\mathsf H_f is:

  • unanimous if it respects unanimous agreement on any pair: f(1,…,1)=1,f(−1,…,−1)=−1; f(1,\ldots,1)=1,\qquad f(-1,\ldots,-1)=-1;
  • consistent at every profile if it never outputs a cycle: Hf(π)∈Rfor every π∈L(C)n; \mathsf H_f(\boldsymbol\pi)\in\mathcal R \quad\text{for every }\boldsymbol\pi\in\mathcal L(\mathcal C)^n;
  • a dictatorship if there is a single fixed voter j∈[n]j\in[n] such that f(u)=ujfor every u∈{−1,1}n, f(u)=u_j\quad\text{for every }u\in\{-1,1\}^n, in which case Hf(π)=(xj,yj,zj)\mathsf H_f(\boldsymbol\pi)=(x_j,y_j,z_j) always copies voter jj’s ballot πj\pi_j. A rule is non-dictatorial if no such jj exists.

Our design question is now completely specified: does there exist a unanimous, non-dictatorial Boolean rule ff that is consistent at every profile? Note that ff is not required to treat voters symmetrically—it may assign arbitrary weights or roles to different coordinates—yet we ask whether avoiding cycles forces all power into a single coordinate.

A proposed rule that fails the consistency requirement

Example 1 (Pairwise majority creates a cycle) For n=3n=3, let f=Maj⁡3f=\operatorname{Maj}_3, which outputs +1+1 when at least two of its three input signs are +1+1, and −1-1 otherwise. Consider the cyclic profile:

Voter Ballot xix_i (AA vs. BB) yiy_i (BB vs. CC) ziz_i (CC vs. AA)
1 A>B>CA>B>C +1+1 +1+1 −1-1
2 B>C>AB>C>A −1-1 +1+1 +1+1
3 C>A>BC>A>B +1+1 −1-1 +1+1

Reading down the columns gives x=(+1,−1,+1)x=(+1,-1,+1), y=(+1,+1,−1)y=(+1,+1,-1), and z=(−1,+1,+1)z=(-1,+1,+1). Since each column contains two +1+1 signs and one −1-1 sign, majority outputs +1+1 on all three: HMaj⁡3(π)=(+1,+1,+1)∉R. \mathsf H_{\operatorname{Maj}_3}(\boldsymbol\pi)=(+1,+1,+1)\notin\mathcal R. Society prefers AA to BB, BB to CC, and CC to AA. Every individual ballot is transitive, yet pairwise majority produces a cycle—Condorcet’s paradox.

Majority is unanimous and non-dictatorial, so this single profile disqualifies it. What about electoral colleges, weighted voting, or more intricate Boolean rules? To test consistency across all 22n2^{2^n} possible rules without inspecting 6n6^n profiles by hand, the next section converts universal consistency into an expectation calculation [3,6].

1.2 Random rankings turn consistency into a correlation calculation

Instead of checking all 6n6^n profiles one by one, suppose each voter i∈[n]i\in[n] independently draws a ballot Πi\Pi_i uniformly at random from the six rankings in L(C)\mathcal L(\mathcal C) (the impartial culture distribution). Because every profile π∈L(C)n\boldsymbol\pi\in\mathcal L(\mathcal C)^n has strictly positive probability 6−n6^{-n}, a rule ff is consistent at every profile if and only if P(Hf(Π)∉R)=0. \mathbb P(\mathsf H_f(\boldsymbol\Pi)\notin\mathcal R)=0. Randomness here is purely a proof device: we use a uniform distribution over profiles to test a deterministic, worst-case requirement.

Let X=x(Π)X=x(\boldsymbol\Pi), Y=y(Π)Y=y(\boldsymbol\Pi), and Z=z(Π)Z=z(\boldsymbol\Pi) be the three random comparison columns generated by Π\boldsymbol\Pi. Look first at a single column, say XX: for each voter ii, exactly three of the six rankings place AA above BB (Xi=+1X_i=+1) and three place BB above AA (Xi=−1X_i=-1). Since voters are independent, XX is distributed uniformly on the hypercube {−1,1}n\{-1,1\}^n, and the same holds marginally for YY and ZZ.

Across columns, however, a single voter’s signs are coupled by transitivity. When do XiX_i and YiY_i agree? Having Xi=Yi=+1X_i=Y_i=+1 means A≻iBA\succ_i B and B≻iCB\succ_i C, which by transitivity forces the single ranking A>B>CA>B>C; similarly, Xi=Yi=−1X_i=Y_i=-1 forces C>B>AC>B>A. In the other four rankings—whenever BB is ranked first or last rather than in the middle—XiX_i and YiY_i must have opposite signs. Therefore, for each voter ii, EXi=EYi=0,E[XiYi]=(+1)⋅26+(−1)⋅46=−13. \mathbb E X_i=\mathbb E Y_i=0,\qquad \mathbb E[X_iY_i]=(+1)\cdot\frac26+(-1)\cdot\frac46=-\frac13. By symmetry of the three cyclic orientations, (Yi,Zi)(Y_i,Z_i) and (Zi,Xi)(Z_i,X_i) follow the exact same joint distribution. Because different voters choose their rankings independently, the pair of vectors (X,Y)(X,Y) consists of nn independent coordinate pairs (Xi,Yi)(X_i,Y_i) with correlation −1/3-1/3.

Now we need an algebraic expression for the event that Hf(Π)=(f(X),f(Y),f(Z))\mathsf H_f(\boldsymbol\Pi)=(f(X),f(Y),f(Z)) forms a cycle. For any triple of signs (a,b,c)∈{−1,1}3(a,b,c)\in\{-1,1\}^3, consider the sum of the three pairwise products ab+bc+caab+bc+ca:

Rescaling {−1,3}\{-1,3\} to {0,1}\{0,1\} gives the exact polynomial indicator of a cycle: 1cycle(a,b,c)=1+ab+bc+ca4. \mathbf 1_{\mathrm{cycle}}(a,b,c) =\frac{1+ab+bc+ca}{4}. Notice what this quadratic formula buys us: even though the three vectors (X,Y,Z)(X,Y,Z) are not jointly independent, the indicator is a sum of pairwise products! Taking expectations and using the identical marginal distribution of (X,Y)(X,Y), (Y,Z)(Y,Z), and (Z,X)(Z,X), we obtain P(cycle)=1+E[f(X)f(Y)]+E[f(Y)f(Z)]+E[f(Z)f(X)]4=14+34E[f(X)f(Y)]. \begin{aligned} \mathbb P(\text{cycle}) &=\frac{1+\mathbb E[f(X)f(Y)]+\mathbb E[f(Y)f(Z)]+\mathbb E[f(Z)f(X)]}{4}\\ &=\frac14+\frac34\mathbb E[f(X)f(Y)]. \end{aligned} Our combinatorial design problem has been reduced to a single analytic question: when (X,Y)(X,Y) is a pair of uniform sign vectors whose coordinates are independent with correlation E[XiYi]=−1/3\mathbb E[X_iY_i]=-1/3, how does E[f(X)f(Y)]\mathbb E[f(X)f(Y)] depend on the rule ff?

1.3 Product functions as a basis for collective rules

Why was E[f(X)f(Y)]\mathbb E[f(X)f(Y)] hard to compute directly for a general rule ff, yet effortless if ff were a product of coordinate signs? Because the coordinate pairs (Xi,Yi)(X_i,Y_i) are independent across voters i∈[n]i\in[n], expectation factors over products of distinct coordinates! This observation suggests expanding any function f:{−1,1}n→Rf:\{-1,1\}^n\to\mathbb R as a linear combination of product functions.

Throughout this section, single-cube expectations use a uniform random vector X∈{−1,1}nX\in\{-1,1\}^n (independent fair signs X1,…,XnX_1,\ldots,X_n). For each subset S⊆[n]S\subseteq[n], define the parity function (or character) χS(x)=∏i∈Sxi,χ∅(x)=1. \chi_S(x)=\prod_{i\in S}x_i, \qquad \chi_\varnothing(x)=1. Its value is +1+1 when an even number of coordinates in SS are −1-1, and −1-1 when that number is odd. Because xi2=1x_i^2=1 for every ii, multiplying two characters cancels their shared coordinates and keeps those in the symmetric difference S△T=(S∖T)∪(T∖S)S\mathbin{\triangle}T=(S\setminus T)\cup(T\setminus S): χS(x)χT(x)=χS△T(x). \chi_S(x)\chi_T(x)=\chi_{S\mathbin{\triangle}T}(x).

Recall (Bits vs. signs). Why use {−1,1}\{-1,1\} rather than {0,1}\{0,1\}? Under xi=(−1)ui=1−2uix_i=(-1)^{u_i}=1-2u_i, addition modulo 22 becomes multiplication: (−1)ui⊕vi=xiyi(-1)^{u_i\oplus v_i}=x_i y_i. The parity ⨁i∈Sui mod 2\bigoplus_{i\in S}u_i\bmod 2 thus becomes the monomial χS(x)=∏i∈Sxi\chi_S(x)=\prod_{i\in S}x_i—the group characters of F2n\mathbb F_2^n (or Walsh–Hadamard basis).

Under independent fair signs, the expectation of any nonempty product is ∏i∈SE[Xi]=0\prod_{i\in S}\mathbb E[X_i]=0, whereas E[χ∅]=1\mathbb E[\chi_\varnothing]=1. Hence the 2n2^n characters are orthonormal with respect to the inner product ⟨f,g⟩=E[f(X)g(X)]\langle f,g\rangle=\mathbb E[f(X)g(X)]: E[χS(X)χT(X)]={1,S=T,0,S≠T. \mathbb E[\chi_S(X)\chi_T(X)]= \begin{cases} 1,&S=T,\\ 0,&S\ne T. \end{cases} Since the vector space of real-valued functions on {−1,1}n\{-1,1\}^n has dimension 2n2^n, these 2n2^n orthonormal functions form a complete basis. Better still, following O’Donnell [3, §1.1], we can write down any function’s expansion explicitly by interpolating its truth table row by row:

Combining this explicit truth-table interpolation with the orthonormality of {χS}S⊆[n]\{\chi_S\}_{S\subseteq[n]} yields the fundamental theorem of Boolean Fourier analysis.

Theorem 1 (Fourier expansion and Parseval’s identity) Every function f:{−1,1}n→Rf:\{-1,1\}^n\to\mathbb R has a unique multilinear expansion f(x)=∑S⊆[n]f^(S)χS(x),f^(S)=⟨f,χS⟩=E[f(X)χS(X)]=2−n∑a∈{−1,1}nf(a)χS(a). f(x)=\sum_{S\subseteq[n]}\widehat f(S)\chi_S(x), \qquad \widehat f(S)=\langle f,\chi_S\rangle=\mathbb E[f(X)\chi_S(X)]=2^{-n}\sum_{a\in\{-1,1\}^n}f(a)\chi_S(a). For any f,g:{−1,1}n→Rf,g:\{-1,1\}^n\to\mathbb R, orthogonality gives Plancherel’s identity E[f(X)g(X)]=∑S⊆[n]f^(S)g^(S). \mathbb E[f(X)g(X)]=\sum_{S\subseteq[n]}\widehat f(S)\widehat g(S). In particular, setting g=fg=f yields Parseval’s identity and the spectral variance formula: E[f(X)2]=∑S⊆[n]f^(S)2,Var⁡(f)=E[f(X)2]−(Ef(X))2=∑S≠∅f^(S)2. \mathbb E[f(X)^2]=\sum_{S\subseteq[n]}\widehat f(S)^2, \qquad \operatorname{Var}(f)=\mathbb E[f(X)^2]-(\mathbb E f(X))^2 =\sum_{S\ne\varnothing}\widehat f(S)^2.

Proof. Existence of the expansion f=∑Sf^(S)χSf=\sum_S\widehat f(S)\chi_S was just shown by truth-table interpolation; for uniqueness, if f=∑ScSχSf=\sum_S c_S\chi_S, taking the inner product with χT\chi_T gives ⟨f,χT⟩=∑ScS⟨χS,χT⟩=cT\langle f,\chi_T\rangle=\sum_S c_S\langle\chi_S,\chi_T\rangle=c_T. Expanding both ff and gg in E[fg]\mathbb E[fg] and applying E[χSχT]=1S=T\mathbb E[\chi_S\chi_T]=\mathbf 1_{S=T} eliminates all cross terms S≠TS\ne T. Finally, f^(∅)=E[f⋅1]=Ef\widehat f(\varnothing)=\mathbb E[f\cdot 1]=\mathbb E f, so subtracting f^(∅)2\widehat f(\varnothing)^2 from E[f2]\mathbb E[f^2] leaves the sum over nonempty subsets.

When f:{−1,1}n→{−1,1}f:\{-1,1\}^n\to\{-1,1\} is Boolean-valued, f(x)2=1f(x)^2=1 at every input, so Parseval states that the squared Fourier coefficients form a probability distribution over the 2n2^n subsets of [n][n]: ∑S⊆[n]f^(S)2=1. \sum_{S\subseteq[n]}\widehat f(S)^2=1. The cardinality ∣S∣|S| is called the degree or level of the coefficient f^(S)\widehat f(S), and we group the total energy at degree k∈{0,1,…,n}k\in\{0,1,\ldots,n\} into the level-kk Fourier weight Wk(f)=∑∣S∣=kf^(S)2,∑k=0nWk(f)=E[f(X)2]=1. W_k(f)=\sum_{|S|=k}\widehat f(S)^2, \qquad \sum_{k=0}^n W_k(f)=\mathbb E[f(X)^2]=1. Geometrically, if we measure the relative Hamming distance between two Boolean functions by dist⁡(f,g)=P(f(X)≠g(X))\operatorname{dist}(f,g)=\mathbb P(f(X)\ne g(X)), then f(X)χS(X)f(X)\chi_S(X) equals +1+1 on agreement and −1-1 on disagreement, so f^(S)=E[f(X)χS(X)]=1−2dist⁡(f,χS). \widehat f(S)=\mathbb E[f(X)\chi_S(X)] =1-2\operatorname{dist}(f,\chi_S). Thus f^(S)\widehat f(S) records how strongly ff agrees with the parity χS\chi_S (an interpretation we will use directly for linearity testing in Reading Contents), while the weights W0(f),W1(f),…,Wn(f)W_0(f),W_1(f),\ldots,W_n(f) partition one unit of total energy across interaction levels.

Computing Fourier coefficients: basic examples

Before introducing the noise operator, let us compute the Fourier coefficients of four canonical families of functions that will serve as our running examples throughout the chapter. Together, they illustrate how our two views of the expansion—interpolating the truth table via ∏i1+aixi2\prod_i\frac{1+a_i x_i}{2} and evaluating inner products f^(S)=E[f(X)χS(X)]\widehat f(S)=\mathbb E[f(X)\chi_S(X)] using symmetry and conditioning—work hand in hand.

Example 2 (Dictator and Parity (single-character rules))  

  • Dictator f(x)=xi=χ{i}(x)f(x)=x_i=\chi_{\{i\}}(x): since ff is already a basis character, uniqueness of the Fourier expansion gives f^({i})=1\widehat f(\{i\})=1 and f^(S)=0\widehat f(S)=0 for all S≠{i}S\ne\{i\}. All Fourier weight sits on level 11: W1(f)=1W_1(f)=1 and Wk(f)=0W_k(f)=0 for k≠1k\ne 1.
  • Parity on T⊆[n]T\subseteq[n] f(x)=χT(x)=∏i∈Txif(x)=\chi_T(x)=\prod_{i\in T}x_i: similarly, f^(T)=1\widehat f(T)=1 and f^(S)=0\widehat f(S)=0 for S≠TS\ne T. If ∣T∣=k|T|=k, all Fourier weight sits on level kk: Wk(χT)=1W_k(\chi_T)=1. In particular, the full parity χ[n](x)=x1⋯xn\chi_{[n]}(x)=x_1\cdots x_n concentrates all of its energy at the highest degree Wn=1W_n=1.

Example 3 (Subcube indicator and ANDw\mathrm{AND}_w (single-row interpolation)) Let A={x∈{−1,1}n:x1=⋯=xw=+1}A=\{x\in\{-1,1\}^n:x_1=\cdots=x_w=+1\} be the subcube where the first ww coordinates are +1+1 (this will be the building block of the Tribes function in §1.10 and the rare-event persistence example in §1.12). How do we expand its 0/10/1-valued indicator 1A(x)\mathbf 1_A(x) and the corresponding {−1,1}\{-1,1\}-valued conjunction ANDw(x)=21A(x)−1\mathrm{AND}_w(x)=2\mathbf 1_A(x)-1?

Viewed on the first ww coordinates, 1A\mathbf 1_A has a single nonzero row a=(+1,…,+1)a=(+1,\ldots,+1) in its truth table! Applying our point-indicator formula across i=1,…,wi=1,\ldots,w immediately expands 1A(x)\mathbf 1_A(x): 1A(x)=∏i=1w(1+xi2)=2−w∑S⊆[w]χS(x). \mathbf 1_A(x)=\prod_{i=1}^w\left(\frac{1+x_i}{2}\right) =2^{-w}\sum_{S\subseteq[w]}\chi_S(x). By uniqueness of the Fourier expansion, 1A^(S)=2−w\widehat{\mathbf 1_A}(S)=2^{-w} for every S⊆[w]S\subseteq[w] and 00 otherwise. (Equivalently, from the expectation definition 1A^(S)=E[1A(X)χS(X)]\widehat{\mathbf 1_A}(S)=\mathbb E[\mathbf 1_A(X)\chi_S(X)]: on the event AA of probability 2−w2^{-w}, every xix_i with i∈[w]i\in[w] is +1+1, so χS(x)=+1\chi_S(x)=+1 for all S⊆[w]S\subseteq[w].) For w=2w=2, this gives 1{x1=x2=+1}(x)=1+x1+x2+x1x24\mathbf 1_{\{x_1=x_2=+1\}}(x)=\frac{1+x_1+x_2+x_1x_2}{4}.

Converting from {0,1}\{0,1\} to {−1,1}\{-1,1\} via ANDw(x)=21A(x)−1\mathrm{AND}_w(x)=2\mathbf 1_A(x)-1 yields the Fourier expansion of Boolean conjunction: ANDw(x)=(21−w−1)+∑∅≠S⊆[w]21−wχS(x), \mathrm{AND}_w(x)=(2^{1-w}-1)+\sum_{\varnothing\ne S\subseteq[w]}2^{1-w}\chi_S(x), so ANDw^(∅)=21−w−1\widehat{\mathrm{AND}_w}(\varnothing)=2^{1-w}-1 and ANDw^(S)=21−w\widehat{\mathrm{AND}_w}(S)=2^{1-w} for every nonempty S⊆[w]S\subseteq[w]. Its level weights are W0(ANDw)=(1−21−w)2W_0(\mathrm{AND}_w)=(1-2^{1-w})^2 and Wk(ANDw)=(wk)22−2wW_k(\mathrm{AND}_w)=\binom{w}{k}2^{2-2w} for 1≤k≤w1\le k\le w, which sum to (1−21−w)2+(2w−1)22−2w=1(1-2^{1-w})^2+(2^w-1)2^{2-2w}=1.

Example 4 (Three-voter Majority Maj⁡3\operatorname{Maj}_3 (truth-table interpolation vs. inner products & symmetry)) In §1.1 we used Maj⁡3:{−1,1}3→{−1,1}\operatorname{Maj}_3:\{-1,1\}^3\to\{-1,1\} to exhibit Condorcet’s paradox. Let us compute its Fourier expansion in three ways that highlight our different tools:

  1. By expanding the truth table row by row: Maj⁡3(a)\operatorname{Maj}_3(a) equals +1+1 on four rows—(+1,+1,+1)(+1,+1,+1), (+1,+1,−1)(+1,+1,-1), (+1,−1,+1)(+1,-1,+1), (−1,+1,+1)(-1,+1,+1)—and −1-1 on their four negations −a-a. In the truth-table sum ∑a∈{−1,1}3Maj⁡3(a)∏i=131+aixi2\sum_{a\in\{-1,1\}^3}\operatorname{Maj}_3(a)\prod_{i=1}^3\frac{1+a_i x_i}{2}, pairing each positive row aa with its negation −a-a cancels all even-degree terms: ∏i=13(1+aixi2)−∏i=13(1−aixi2)=a1x1+a2x2+a3x3+a1a2a3 x1x2x34. \prod_{i=1}^3\left(\frac{1+a_i x_i}{2}\right)-\prod_{i=1}^3\left(\frac{1-a_i x_i}{2}\right) =\frac{a_1 x_1+a_2 x_2+a_3 x_3+a_1a_2a_3\,x_1x_2x_3}{4}. Summing this expression over the four positive rows (+1,+1,+1)(+1,+1,+1), (+1,+1,−1)(+1,+1,-1), (+1,−1,+1)(+1,-1,+1), and (−1,+1,+1)(-1,+1,+1) gives Maj⁡3(x)=14[(x1+x2+x3+x1x2x3)+(x1+x2−x3−x1x2x3)+(x1−x2+x3−x1x2x3)+(−x1+x2+x3−x1x2x3)]=12x1+12x2+12x3−12x1x2x3. \begin{aligned} \operatorname{Maj}_3(x) &=\frac14\Big[(x_1+x_2+x_3+x_1x_2x_3)+(x_1+x_2-x_3-x_1x_2x_3)\\ &\qquad\quad+(x_1-x_2+x_3-x_1x_2x_3)+(-x_1+x_2+x_3-x_1x_2x_3)\Big]\\ &=\frac12 x_1+\frac12 x_2+\frac12 x_3-\frac12 x_1x_2x_3. \end{aligned}

  2. Directly from the inner product f^(S)=E[Maj⁡3(X)χS(X)]\widehat f(S)=\mathbb E[\operatorname{Maj}_3(X)\chi_S(X)]:

    • Degrees 00 and 22 (∣S∣|S| even): Because Maj⁡3(−x)=−Maj⁡3(x)\operatorname{Maj}_3(-x)=-\operatorname{Maj}_3(x) is odd, replacing XX by −X-X negates Maj⁡3(X)χS(X)\operatorname{Maj}_3(X)\chi_S(X) whenever ∣S∣|S| is even, so Maj⁡3^(∅)=Maj⁡3^({i,j})=0\widehat{\operatorname{Maj}_3}(\varnothing)=\widehat{\operatorname{Maj}_3}(\{i,j\})=0.
    • Degree 11 (S={i}S=\{i\}): For S={1}S=\{1\}, Maj⁡3(X)\operatorname{Maj}_3(X) agrees with X1X_1 on 33 of the 44 equally likely pairs (X2,X3)∈{−1,1}2(X_2,X_3)\in\{-1,1\}^2 (failing only when X2=X3=−X1X_2=X_3=-X_1), so Maj⁡3^({i})=34(+1)+14(−1)=12\widehat{\operatorname{Maj}_3}(\{i\})=\frac34(+1)+\frac14(-1)=\frac12.
    • Degree 33 (S={1,2,3}S=\{1,2,3\}): On the 22 unanimous inputs, Maj⁡3(x)\operatorname{Maj}_3(x) and x1x2x3x_1 x_2 x_3 agree (product +1+1); on the 66 two-against-one inputs, Maj⁡3(x)\operatorname{Maj}_3(x) equals the majority sign while x1x2x3x_1 x_2 x_3 equals the minority sign (product −1-1). Hence Maj⁡3^({1,2,3})=28(+1)+68(−1)=−12\widehat{\operatorname{Maj}_3}(\{1,2,3\})=\frac28(+1)+\frac68(-1)=-\frac12.
  3. Shortcut via oddness and permutation symmetry: Since Maj⁡3\operatorname{Maj}_3 is odd and symmetric under permuting x1,x2,x3x_1,x_2,x_3, it must have the form Maj⁡3(x)=α(x1+x2+x3)+β x1x2x3\operatorname{Maj}_3(x)=\alpha(x_1+x_2+x_3)+\beta\,x_1x_2x_3. Plugging in x=(+1,+1,+1)x=(+1,+1,+1) and x=(+1,+1,−1)x=(+1,+1,-1) gives 3α+β=13\alpha+\beta=1 and α−β=1\alpha-\beta=1, immediately recovering α=1/2\alpha=1/2 and β=−1/2\beta=-1/2.

In all three ways, Maj⁡3\operatorname{Maj}_3 places three quarters of its spectral weight on level 11 and one quarter on level 33: W1(Maj⁡3)=3(12)2=34,W3(Maj⁡3)=(−12)2=14,W0=W2=0. W_1(\operatorname{Maj}_3)=3\left(\frac12\right)^2=\frac34, \qquad W_3(\operatorname{Maj}_3)=\left(-\frac12\right)^2=\frac14, \qquad W_0=W_2=0.

Example 5 (General odd-nn Majority Maj⁡n\operatorname{Maj}_n: conditioning and level-11 weight) What does the Fourier spectrum of Maj⁡n(x)=sgn⁡(x1+⋯+xn)\operatorname{Maj}_n(x)=\operatorname{sgn}(x_1+\cdots+x_n) look like for an arbitrary odd number of voters n=2m+1n=2m+1?

By oddness (Maj⁡n(−x)=−Maj⁡n(x)\operatorname{Maj}_n(-x)=-\operatorname{Maj}_n(x)), all even levels vanish: W0(Maj⁡n)=W2(Maj⁡n)=⋯=0W_0(\operatorname{Maj}_n)=W_2(\operatorname{Maj}_n)=\cdots=0. To compute the degree-11 coefficient Maj⁡n^({i})=E[Maj⁡n(X)Xi]\widehat{\operatorname{Maj}_n}(\{i\})=\mathbb E[\operatorname{Maj}_n(X)X_i], condition on the other n−1=2mn-1=2m signs X−iX_{-i} and let S−i=∑j≠iXjS_{-i}=\sum_{j\ne i}X_j:

  • When S−i≠0S_{-i}\ne 0: since S−iS_{-i} is a nonzero even integer (∣S−i∣≥2|S_{-i}|\ge 2), adding Xi=±1X_i=\pm1 cannot change its sign. Thus Maj⁡n(X)=sgn⁡(S−i)\operatorname{Maj}_n(X)=\operatorname{sgn}(S_{-i}) is already fixed by X−iX_{-i}, and conditioning on X−iX_{-i} gives E[Maj⁡n(X)Xi∣X−i]=sgn⁡(S−i) E[Xi]=0\mathbb E[\operatorname{Maj}_n(X)X_i\mid X_{-i}]=\operatorname{sgn}(S_{-i})\,\mathbb E[X_i]=0.
  • When S−i=0S_{-i}=0: the other 2m2m voters are tied (mm pluses and mm minuses), so voter ii breaks the tie and Maj⁡n(X)=Xi\operatorname{Maj}_n(X)=X_i. On this tie event, Maj⁡n(X)Xi=Xi2=1\operatorname{Maj}_n(X)X_i=X_i^2=1!

Taking expectations over X−iX_{-i} shows that Maj⁡n^({i})\widehat{\operatorname{Maj}_n}(\{i\}) equals the probability that the other n−1n-1 fair coins tie: Maj⁡n^({i})=P ⁣(∑j≠iXj=0)=2−(n−1)(n−1(n−1)/2)∼2πn, \widehat{\operatorname{Maj}_n}(\{i\}) =\mathbb P\!\left(\sum_{j\ne i}X_j=0\right) =2^{-(n-1)}\binom{n-1}{(n-1)/2} \sim\sqrt{\frac{2}{\pi n}}, where the asymptotic follows from Stirling’s formula (2mm)∼4mπm\binom{2m}{m}\sim\frac{4^m}{\sqrt{\pi m}} with m=(n−1)/2m=(n-1)/2. Summing the squares of these nn singleton coefficients yields the level-11 Fourier weight of Maj⁡n\operatorname{Maj}_n: W1(Maj⁡n)=∑i=1nMaj⁡n^({i})2=n(2−(n−1)(n−1(n−1)/2))2⟶2π≈0.6366(n→∞). W_1(\operatorname{Maj}_n) =\sum_{i=1}^n\widehat{\operatorname{Maj}_n}(\{i\})^2 =n\left(2^{-(n-1)}\binom{n-1}{(n-1)/2}\right)^2 \longrightarrow\frac{2}{\pi}\approx 0.6366 \qquad(n\to\infty). Even as n→∞n\to\infty, nearly two-thirds (2/π≈63.7%2/\pi\approx 63.7\%) of Majority’s total Fourier weight remains on the linear level k=1k=1! (We will see in §1.8 why this tie-breaking probability is also the coordinate influence Inf⁡i(Maj⁡n)\operatorname{Inf}_i(\operatorname{Maj}_n), and in Reading Contents we work out the higher odd levels W3,W5,…W_3,W_5,\ldots as well.)

Why 2/π2/\pi? By the central limit theorem, for large nn the normalized sum 1n∑i=1nXi\frac{1}{\sqrt n}\sum_{i=1}^n X_i approaches a standard Gaussian. Sheppard’s arcsine formula (1899) states that for ρ\rho-correlated standard Gaussians (G1,G2)(G_1,G_2), E[sgn⁡(G1)sgn⁡(G2)]=2πarcsin⁡ρ=2πρ+13πρ3+⋯\mathbb E[\operatorname{sgn}(G_1)\operatorname{sgn}(G_2)]=\frac{2}{\pi}\arcsin\rho=\frac{2}{\pi}\rho+\frac{1}{3\pi}\rho^3+\cdots. Matching the linear term in ρ\rho with Stab⁡ρ(Maj⁡n)=∑kρkWk(Maj⁡n)\operatorname{Stab}_\rho(\operatorname{Maj}_n)=\sum_k\rho^k W_k(\operatorname{Maj}_n) explains why W1(Maj⁡n)→2/πW_1(\operatorname{Maj}_n)\to 2/\pi.

Comparing these four examples already reveals how the Fourier spectrum reflects combinatorial structure: a dictator concentrates all energy on a single degree-11 coordinate (W1=1W_1=1), full parity pushes all energy to the highest degree (Wn=1W_n=1), ANDw\mathrm{AND}_w spreads its variance across degrees 11 through ww, and Maj⁡n\operatorname{Maj}_n places 2/π2/\pi of its weight on level 11 while leaving the remaining 1−2/π1-2/\pi on higher odd levels k≥3k\ge 3.

1.4 The output correlation we need: noise stability

In the ranking problem, the pair (X,Y)(X,Y) has coordinate correlation ρ=−1/3\rho=-1/3. It costs nothing—and pays off in Meeting 2—to analyze the same coordinatewise experiment for any correlation parameter ρ∈[−1,1]\rho\in[-1,1].

Given a uniform sign vector X∈{−1,1}nX\in\{-1,1\}^n, generate Y∈{−1,1}nY\in\{-1,1\}^n by independently flipping each coordinate XiX_i with probability (1−ρ)/2(1-\rho)/2 and keeping Yi=XiY_i=X_i with probability (1+ρ)/2(1+\rho)/2. Then each coordinate satisfies E[Yi∣Xi]=(1+ρ2−1−ρ2)Xi=ρXi,E[XiYi]=ρ. \mathbb E[Y_i\mid X_i] =\left(\frac{1+\rho}{2}-\frac{1-\rho}{2}\right)X_i =\rho X_i, \qquad \mathbb E[X_iY_i]=\rho. When ρ≥0\rho\ge0, flipping with probability (1−ρ)/2(1-\rho)/2 is equivalent to retaining XiX_i with probability ρ\rho and resampling a fresh fair sign with probability 1−ρ1-\rho. The flip parameterization has the advantage of working for negative correlations as well: at ρ=−1/3\rho=-1/3, each sign is flipped with probability 2/32/3, which is precisely the marginal pair law of (X,Y)(X,Y) under uniform random rankings.

Definition 3 (Noise operator) For h:{−1,1}n→Rh:\{-1,1\}^n\to\mathbb R and −1≤ρ≤1-1\le\rho\le1, the noise operator TρT_\rho averages h(Y)h(Y) conditioned on X=xX=x: (Tρh)(x)=E[h(Y)∣X=x]. (T_\rho h)(x)=\mathbb E[h(Y)\mid X=x].

By the law of total expectation, the correlation between f(X)f(X) and f(Y)f(Y) is the inner product of ff with TρfT_\rho f:

Definition 4 (Noise stability) For f:{−1,1}n→Rf:\{-1,1\}^n\to\mathbb R and a ρ\rho-correlated uniform pair (X,Y)(X,Y), the noise stability of ff at correlation ρ\rho is Stab⁡ρ(f)=E[f(X)f(Y)]=E[f(X) E[f(Y)∣X]]=⟨f,Tρf⟩. \operatorname{Stab}_\rho(f)=\mathbb E[f(X)f(Y)] =\mathbb E[f(X)\,\mathbb E[f(Y)\mid X]] =\langle f,T_\rho f\rangle.

Now we see why the character basis was tailor-made for this operator. Because the coordinate flips are independent across i∈[n]i\in[n], conditioning on X=xX=x factors across any product of coordinates: (TρχS)(x)=E ⁣[∏i∈SYi | X=x]=∏i∈SE[Yi∣Xi=xi]=ρ∣S∣χS(x). (T_\rho\chi_S)(x) =\mathbb E\!\left[\prod_{i\in S}Y_i\,\middle|\,X=x\right] =\prod_{i\in S}\mathbb E[Y_i\mid X_i=x_i] =\rho^{|S|}\chi_S(x).

Recall (Chapter 2 semigroup). When ρ=e−t∈(0,1]\rho=e^{-t}\in(0,1], Te−t=etLT_{e^{-t}}=e^{t\mathcal L} is the continuous-time Markov semigroup on {−1,1}n\{-1,1\}^n that refreshes each coordinate independently at rate 11 (generator L=−∑i=1nDi\mathcal L=-\sum_{i=1}^n D_i). The identity TρχS=ρ∣S∣χST_\rho\chi_S=\rho^{|S|}\chi_S is the exact eigenbasis diagonalization of LχS=−∣S∣χS\mathcal L\chi_S=-|S|\chi_S; see §2.6.

Every character χS\chi_S is an eigenfunction of TρT_\rho with eigenvalue ρ∣S∣\rho^{|S|}! Expanding f=∑Sf^(S)χSf=\sum_S\widehat f(S)\chi_S and applying Plancherel’s identity immediately diagonalizes the output correlation: Stab⁡ρ(f)=⟨f,Tρf⟩=∑S⊆[n]ρ∣S∣f^(S)2=∑k=0nρkWk(f). \operatorname{Stab}_\rho(f) =\langle f,T_\rho f\rangle =\sum_{S\subseteq[n]}\rho^{|S|}\widehat f(S)^2 =\sum_{k=0}^n\rho^k W_k(f). For a Boolean rule ff, where ∑k=0nWk(f)=1\sum_{k=0}^n W_k(f)=1, Stab⁡ρ(f)\operatorname{Stab}_\rho(f) is simply a weighted average of the powers 1,ρ,ρ2,…,ρn1,\rho,\rho^2,\ldots,\rho^n. When ρ<0\rho<0, odd and even levels contribute with opposite signs—the exact leverage we need to analyze ρ=−1/3\rho=-1/3.

1.5 The probability of a cyclic outcome

Combining our cycle-probability reduction from §1.2 with the spectral formula for Stab⁡−1/3(f)\operatorname{Stab}_{-1/3}(f) from §1.4 expresses the probability of a Condorcet cycle directly in terms of the rule’s Fourier weights Wk(f)=∑∣S∣=kf^(S)2W_k(f)=\sum_{|S|=k}\widehat f(S)^2.

Theorem 2 (Kalai’s cycle-probability formula) In a three-candidate election using a common pairwise rule f:{−1,1}n→{−1,1}f:\{-1,1\}^n\to\{-1,1\} under independent uniform voter rankings, the probability of a cyclic outcome is P(cycle)=14+34Stab⁡−1/3(f)=14+34∑k=0n(−13)kWk(f). \mathbb P(\text{cycle}) =\frac14+\frac34\operatorname{Stab}_{-1/3}(f) =\frac14+\frac34\sum_{k=0}^n\left(-\frac13\right)^k W_k(f). Equivalently, using ∑k=0nWk(f)=1\sum_{k=0}^n W_k(f)=1, P(cycle)=∑k=0n1+3(−1/3)k4 Wk(f)=W0(f)+13W2(f)+29W3(f)+727W4(f)+⋯ . \mathbb P(\text{cycle}) =\sum_{k=0}^n\frac{1+3(-1/3)^k}{4}\,W_k(f) =W_0(f)+\frac13 W_2(f)+\frac29 W_3(f)+\frac{7}{27}W_4(f)+\cdots.

Proof. In §1.2 we showed that P(cycle)=14+34E[f(X)f(Y)]\mathbb P(\text{cycle})=\frac14+\frac34\mathbb E[f(X)f(Y)], where (X,Y)(X,Y) is a uniform pair with coordinate correlation ρ=−1/3\rho=-1/3. Thus E[f(X)f(Y)]=Stab⁡−1/3(f)=∑k=0n(−1/3)kWk(f)\mathbb E[f(X)f(Y)]=\operatorname{Stab}_{-1/3}(f)=\sum_{k=0}^n(-1/3)^k W_k(f), giving the first formula. Writing 14=14∑k=0nWk(f)\frac14=\frac14\sum_{k=0}^n W_k(f) combines the two terms into the level-by-level sum, where the k=1k=1 coefficient is 1+3(−1/3)4=0\frac{1+3(-1/3)}{4}=0.

This identity is due to Kalai [6] (see also O’Donnell [3], Theorem 2.56). Using the Fourier expansions we computed in §1.3, every special case can be read off at a glance:

1.6 Perfect consistency concentrates power in one voter

We have turned consistency of a collective ranking into an exact restriction on the rule’s spectrum: all Fourier mass must sit on level one. It remains to determine which Boolean rules can satisfy that restriction.

Lemma 1 (A Boolean rule supported on level one) If a Boolean function f:{−1,1}n→{−1,1}f:\{-1,1\}^n\to\{-1,1\} satisfies W1(f)=1W_1(f)=1, then f(x)=xif(x)=x_i or f(x)=−xif(x)=-x_i for some coordinate ii.

Proof. Because W1(f)=1W_1(f)=1 and ∑k=0nWk(f)=1\sum_{k=0}^n W_k(f)=1, all Fourier coefficients outside level one vanish. Thus f(x)=∑i=1nbixif(x)=\sum_{i=1}^n b_i x_i with ∑i=1nbi2=1\sum_{i=1}^n b_i^2=1. The Boolean range f(x)∈{−1,1}f(x)\in\{-1,1\} now imposes a much stronger pointwise constraint: f(x)2=1f(x)^2=1 at every input xx. Expanding the square in the character basis gives 1=f(x)2=(∑i=1nbi2)χ∅(x)+2∑1≤i<j≤nbibj χ{i,j}(x). 1=f(x)^2=\left(\sum_{i=1}^n b_i^2\right)\chi_\varnothing(x)+2\sum_{1\le i<j\le n}b_i b_j\,\chi_{\{i,j\}}(x). Since the characters {χS}\{\chi_S\} form a basis, matching coefficients of χ{i,j}\chi_{\{i,j\}} on both sides forces bibj=0b_i b_j=0 for every distinct pair i<ji<j. Hence at most one coefficient bib_i can be nonzero; since ∑i=1nbi2=1\sum_{i=1}^n b_i^2=1, that single coefficient must be ±1\pm1, so f(x)=±xif(x)=\pm x_i.

Theorem 3 (Arrow’s theorem: the three-candidate common-rule version) Suppose a unanimous Boolean rule ff is used for all three pairwise comparisons. If every profile of strict individual rankings produces a collective ranking without a cycle, then f(x)=xi f(x)=x_i for one fixed voter ii.

Proof. The hypothesis is Hf(π)∈R\mathsf H_f(\boldsymbol\pi)\in\mathcal R for every profile π∈Rn\boldsymbol\pi\in\mathcal R^n. Because every profile has positive weight 6−n>06^{-n}>0 under the uniform-ranking experiment, universal consistency implies P(cycle)=0\mathbb P(\text{cycle})=0. The spectral identity in §1.5 then forces W1(f)=1W_1(f)=1, and Lemma 1 gives f=±xif=\pm x_i. Unanimity (f(1)=+1f(\mathbf 1)=+1) rules out −xi-x_i. Conversely, the dictator rule f(x)=xif(x)=x_i simply copies voter ii’s strict ranking on every pair, which is always transitive.

General Arrow theorem. Classical social choice allows three distinct pairwise rules (fAB,fBC,fCA)(f_{AB},f_{BC},f_{CA}) subject to unrestricted domain, IIA, unanimity, and transitivity. Our symmetric common-rule version follows O’Donnell [3, §2.5]; Exercise 2.51 there and Kalai [6] show how the same Fourier correlation argument extends to general triples of rules.

Notice that the dictator ii is fixed by the rule ff itself—it is the same voter for every profile and every pairwise comparison. Moreover, Theorem 3 is a completely deterministic statement about all preference profiles; the uniform distribution on Rn\mathcal R^n served purely as an analytic probe to turn a combinatorial universal quantifier into a spectral equation.

We began this chapter by asking for a collective decision rule that is both consistent and shared among many participants. For three-candidate rankings, universal consistency collapses all power into a single dictator. That stark conclusion sets up our second classroom story: if we drop three-candidate rankings and ask only for a single binary decision, can control be spread so widely across nn participants that no small coalition can manipulate the outcome?

1.7 Can participants collectively generate an honest random bit?

In a binary decision there is only one pair of alternatives, so Condorcet cycles disappear. We can therefore focus squarely on the dispersion of power.

Consider nn participants who want to generate a common random bit in a single broadcast round. Each participant ii broadcasts a sign xi∈{−1,1}x_i\in\{-1,1\}, and everyone applies a publicly known deterministic Boolean rule f:{−1,1}n→{−1,1}f:\{-1,1\}^n\to\{-1,1\} to the broadcast vector x=(x1,…,xn)x=(x_1,\dots,x_n). When all participants are honest, their broadcasts are independent and uniform on {−1,1}\{-1,1\}. We call ff balanced when E[f(X)]=0\mathbb E[f(X)]=0, meaning that an all-honest execution produces a fair coin (P(f(X)=+1)=1/2\mathbb P(f(X)=+1)=1/2).

Now suppose a coalition S⊆[n]S\subseteq[n] of dishonest participants colludes to bias the outcome toward +1+1. We study a single-round rushing adversary: the coalition SS is fixed before the protocol begins, waits to see the broadcasts X[n]∖SX_{[n]\setminus S} of the honest participants, and then coordinates its own signs yS∈{−1,1}Sy_S\in\{-1,1\}^S to force f(X[n]∖S,yS)=+1f(X_{[n]\setminus S},y_S)=+1 whenever possible. Here xS=(xi)i∈Sx_S=(x_i)_{i\in S}, and f(x[n]∖S,yS)f(x_{[n]\setminus S},y_S) denotes ff evaluated on the combined vector taking xix_i outside SS and yiy_i inside SS. Under uniform honest broadcasts X[n]∖SX_{[n]\setminus S}, the coalition’s maximum achievable probability of forcing +1+1 is p+(S)=PX[n]∖S ⁣(there exists yS∈{−1,1}S such that f(X[n]∖S,yS)=+1). p_+(S)=\mathbb P_{X_{[n]\setminus S}}\!\left( \text{there exists }y_S\in\{-1,1\}^S\text{ such that }f(X_{[n]\setminus S},y_S)=+1\right). Our design question is concrete: can we construct a balanced rule ff for which every small coalition SS keeps p+(S)p_+(S) close to 1/21/2? This collective coin-flipping problem was introduced by Ben-Or and Linial [5] and motivated the foundational influence theorem of Kahn, Kalai, and Linial [2].

Why the obvious rules are vulnerable

Before looking for a general theorem, let us test three natural candidate rules against a rushing coalition:

Dictator and parity fall to a coalition of size 11; majority survives up to coalition size Θ(n)\Theta(\sqrt n). Can a cleverer balanced rule resist coalitions of size Θ(n)\Theta(n)? We will prove that no balanced rule can: every balanced ff admits a coalition of size O(n/log⁡n)O(n/\log n)—a vanishing o(1)o(1) fraction of the participants—that forces +1+1 with probability at least 0.990.99. Our strategy is to measure how much leverage a single coordinate has over the outcome, turn that leverage into one step of a greedy coalition attack, and then iterate.

How close to nn can we get? Ben-Or and Linial [5] showed that iterated majority on ternary trees resists rushing coalitions of size O(nlog⁡32)≈O(n0.631)O(n^{\log_3 2})\approx O(n^{0.631}). Ajtai and Linial (1993) later took an AND\mathrm{AND} of O(log⁡n)O(\log n) random Tribes functions on overlapping coordinate partitions to balance the +1+1 and −1-1 vulnerabilities, resisting coalitions of size Ω(n/log⁡2n)\Omega(n/\log^2 n) in both directions—just one log⁡n\log n factor below the O(n/log⁡n)O(n/\log n) upper bound we prove via KKL [2]. Closing that remaining log⁡n\log n gap between Ω(n/log⁡2n)\Omega(n/\log^2 n) and O(n/log⁡n)O(n/\log n) is still open!

Notice the shift in the role of probability between our two classroom stories. In Arrow’s theorem, random rankings were merely a proof device to test a deterministic condition required at every profile. In collective coin flipping, the uniform distribution on honest inputs is part of the protocol model itself, and the theorem directly measures the probability that a coalition biases the protocol’s random output.

1.8 Encoding a participant’s opportunity to control the answer

When can a single rushing participant ii control the output of ff? Exactly when the other n−1n-1 honest broadcasts fall in a configuration where flipping coordinate ii changes the value of ff. Let x⊕ix^{\oplus i} denote xx with its ii-th coordinate flipped (xi↦−xix_i\mapsto -x_i).

Definition 5 (Influence and total influence) For a Boolean function f:{−1,1}n→{−1,1}f:\{-1,1\}^n\to\{-1,1\}, the influence of coordinate ii is the probability that coordinate ii is pivotal under a uniform random input: Inf⁡i(f)=PX∼Unif({−1,1}n) ⁣(f(X)≠f(X⊕i)). \operatorname{Inf}_i(f)=\mathbb P_{X\sim\mathrm{Unif}(\{-1,1\}^n)}\!\left(f(X)\ne f(X^{\oplus i})\right). The total influence of ff is I(f)=∑i=1nInf⁡i(f)I(f)=\sum_{i=1}^n\operatorname{Inf}_i(f).

To connect this combinatorial pivot probability to the Fourier expansion from Meeting 1, we encode the pivotal event algebraically. Define the discrete derivative of ff along coordinate ii by Dif(x)=f(x)−f(x⊕i)2. D_i f(x)=\frac{f(x)-f(x^{\oplus i})}{2}. Because ff takes values in {−1,1}\{-1,1\}, the difference f(x)−f(x⊕i)f(x)-f(x^{\oplus i}) is 00 when f(x)=f(x⊕i)f(x)=f(x^{\oplus i}) and ±2\pm2 when f(x)≠f(x⊕i)f(x)\ne f(x^{\oplus i}). Therefore hi(x)=Dif(x)h_i(x)=D_i f(x) takes values in {−1,0,1}\{-1,0,1\}, and its square is the exact 0/10/1 indicator of the pivotal event: (Dif(x))2=1{f(x)≠f(x⊕i)},Inf⁡i(f)=E ⁣[(Dif(X))2]=∥Dif∥22. (D_i f(x))^2=\mathbf 1_{\{f(x)\ne f(x^{\oplus i})\}}, \qquad \operatorname{Inf}_i(f)=\mathbb E\!\left[(D_i f(X))^2\right]=\|D_i f\|_2^2. Now evaluate DiD_i on the character basis. Flipping xix_i leaves χS(x)\chi_S(x) unchanged if i∉Si\notin S and negates χS(x)\chi_S(x) if i∈Si\in S. Consequently, χS(x)−χS(x⊕i)2\frac{\chi_S(x)-\chi_S(x^{\oplus i})}{2} equals χS(x)\chi_S(x) when i∈Si\in S and 00 when i∉Si\notin S. By linearity, applying DiD_i simply projects the Fourier expansion onto the characters that contain ii: Dif=∑S∋if^(S)χS. D_i f=\sum_{S\ni i}\widehat f(S)\chi_S. Applying Parseval’s identity to DifD_i f immediately yields the Fourier formulas for individual and total influence: Inf⁡i(f)=∑S∋if^(S)2,I(f)=∑i=1nInf⁡i(f)=∑S⊆[n]∣S∣f^(S)2. \operatorname{Inf}_i(f)=\sum_{S\ni i}\widehat f(S)^2, \qquad I(f)=\sum_{i=1}^n\operatorname{Inf}_i(f)=\sum_{S\subseteq[n]} |S|\widehat f(S)^2.

direction i = 1 (2/4 edges pivotal)

Geometrically, Inf⁡i(f)\operatorname{Inf}_i(f) is the fraction of direction-ii edges whose endpoints get opposite signs (red edges above for Maj⁡3\operatorname{Maj}_3, giving Inf⁡1(Maj⁡3)=1/2\operatorname{Inf}_1(\operatorname{Maj}_3)=1/2). Notation: our DifD_i f keeps xix_i inside χS\chi_S; O’Donnell [3] writes Dibookf=xiDifD_i^{\mathrm{book}}f=x_i D_i f, which has the same ∣Dif∣|D_i f| everywhere.

The first formula says that participant ii’s influence is the total Fourier weight of all interactions involving ii; the second says that total influence weights each Fourier level by its degree ∣S∣|S|, because a coalition SS of size kk is counted once by each of its kk members.

Example 6 (When is a voter pivotal in Majority?) For odd nn, voter ii is pivotal in Maj⁡n\operatorname{Maj}_n precisely when the other n−1n-1 honest signs split into (n−1)/2(n-1)/2 pluses and (n−1)/2(n-1)/2 minuses. By Stirling’s approximation for the central binomial coefficient, Inf⁡i(Maj⁡n)=2−(n−1)(n−1(n−1)/2)∼2πn,I(Maj⁡n)∼2nπ. \operatorname{Inf}_i(\operatorname{Maj}_n) =2^{-(n-1)}\binom{n-1}{(n-1)/2} \sim\sqrt{\frac{2}{\pi n}}, \qquad I(\operatorname{Maj}_n)\sim\sqrt{\frac{2n}{\pi}}. Compare this with Example 5 in §1.3: for Maj⁡n\operatorname{Maj}_n, the influence Inf⁡i(Maj⁡n)\operatorname{Inf}_i(\operatorname{Maj}_n) is equal to the degree-11 Fourier coefficient Maj⁡n^({i})\widehat{\operatorname{Maj}_n}(\{i\})! (In Reading Contents we show that Inf⁡i(f)=f^({i})\operatorname{Inf}_i(f)=\widehat f(\{i\}) holds for every monotone Boolean function.) For n=3n=3, Inf⁡i(Maj⁡3)=2−2(21)=1/2\operatorname{Inf}_i(\operatorname{Maj}_3)=2^{-2}\binom{2}{1}=1/2; the Fourier formula Inf⁡i(f)=∑S∋if^(S)2\operatorname{Inf}_i(f)=\sum_{S\ni i}\widehat f(S)^2 gives the same answer from Example 4, since coordinate ii participates in Maj⁡3^({i})=1/2\widehat{\operatorname{Maj}_3}(\{i\})=1/2 and Maj⁡3^({1,2,3})=−1/2\widehat{\operatorname{Maj}_3}(\{1,2,3\})=-1/2, contributing (1/2)2+(−1/2)2=1/2(1/2)^2+(-1/2)^2=1/2.

1.9 One controlled coordinate and the limits of Poincaré

How does influence translate into progress for a rushing coalition? Suppose the coalition already controls a subset S⊆[n]S\subseteq[n] and wants to force +1+1. Define the residual success function on the remaining honest coordinates [n]∖S[n]\setminus S by fS(x[n]∖S)=max⁡yS∈{−1,1}Sf(x[n]∖S,yS), f_S(x_{[n]\setminus S})=\max_{y_S\in\{-1,1\}^S} f(x_{[n]\setminus S},y_S), where −1<+1-1<+1. By definition, fS(x[n]∖S)=+1f_S(x_{[n]\setminus S})=+1 if and only if the coalition SS can force +1+1 on the honest draw x[n]∖Sx_{[n]\setminus S}, so P(fS=+1)=p+(S)\mathbb P(f_S=+1)=p_+(S). Crucially, fSf_S is again a Boolean function taking values in {−1,1}\{-1,1\} on the (n−∣S∣)(n-|S|)-dimensional cube of untouched honest coordinates—even when the original rule ff is neither monotone nor symmetric.

Now suppose the coalition recruits one additional participant i∉Si\notin S. Since maximizing over S∪{i}S\cup\{i\} is the same as maximizing over yi∈{−1,1}y_i\in\{-1,1\} after maximizing over yS∈{−1,1}Sy_S\in\{-1,1\}^S, we have fS∪{i}(x[n]∖(S∪{i}))=max⁡yi∈{−1,1}fS(x[n]∖(S∪{i}),yi). f_{S\cup\{i\}}(x_{[n]\setminus(S\cup\{i\})}) =\max_{y_i\in\{-1,1\}} f_S(x_{[n]\setminus(S\cup\{i\})},y_i). Fix the remaining honest signs x[n]∖(S∪{i})x_{[n]\setminus(S\cup\{i\})}. There are three possibilities for the pair of values fS(⋅,+1)f_S(\cdot,+1) and fS(⋅,−1)f_S(\cdot,-1):

  1. Both are +1+1: then fS=+1f_S=+1 and fS∪{i}=+1f_{S\cup\{i\}}=+1, so recruiting ii changes nothing.
  2. Both are −1-1: then fS=−1f_S=-1 and fS∪{i}=−1f_{S\cup\{i\}}=-1, so recruiting ii changes nothing.
  3. One is +1+1 and the other is −1-1 (that is, coordinate ii is pivotal for fSf_S): when ii was honest, its random sign Xi∈{−1,1}X_i\in\{-1,1\} gave +1+1 with probability 1/21/2; once ii joins the coalition, it chooses the winning sign and gives +1+1 with probability 11.

Thus recruiting ii increases the success probability from 1/21/2 to 11 on the pivotal event of probability Inf⁡i(fS)\operatorname{Inf}_i(f_S), and changes nothing elsewhere! This gives the exact one-step gain identity: p+(S∪{i})−p+(S)=12Inf⁡i(fS). p_+(S\cup\{i\})-p_+(S)=\frac12\operatorname{Inf}_i(f_S).

How large an influence can our Fourier formulas guarantee on an arbitrary Boolean function hh with r=n−∣S∣r=n-|S| remaining coordinates? Since ∣A∣≥1|A|\ge1 for every nonconstant Fourier term A≠∅A\ne\varnothing, comparing variance with total influence gives the Poincaré inequality on the cube: Var⁡(h)=∑A≠∅h^(A)2≤∑A⊆[r]∣A∣h^(A)2=I(h). \operatorname{Var}(h)=\sum_{A\ne\varnothing}\widehat h(A)^2 \le\sum_{A\subseteq[r]}|A|\widehat h(A)^2=I(h).

Recall (Poincaré from Chapter 2). For any reversible Markov chain with spectral gap γ\gamma, Chapter 2 showed Var⁡(h)≤1γE(h,h)\operatorname{Var}(h)\le\frac{1}{\gamma}\mathcal E(h,h). On the hypercube with rate-11 coordinate refreshes, the Dirichlet form is E(h,h)=∑i=1r∥Dih∥22=I(h)\mathcal E(h,h)=\sum_{i=1}^r\|D_i h\|_2^2=I(h) and γ=1\gamma=1. On the Fourier side, this becomes the one-line inequality 1≤∣A∣1\le |A| for A≠∅A\ne\varnothing.

By averaging over the rr coordinates, some coordinate ii must satisfy Inf⁡i(h)≥I(h)/r≥Var⁡(h)/r\operatorname{Inf}_i(h)\ge I(h)/r\ge\operatorname{Var}(h)/r. Because h∈{−1,1}h\in\{-1,1\} with p=P(h=+1)p=\mathbb P(h=+1), its variance is Var⁡(h)=4p(1−p)\operatorname{Var}(h)=4p(1-p), so greedy recruitment increases p+(S)p_+(S) by at least p+(S∪{i})−p+(S)≥Var⁡(fS)2r=2p(1−p)r. p_+(S\cup\{i\})-p_+(S)\ge\frac{\operatorname{Var}(f_S)}{2r}=\frac{2p(1-p)}{r}.

Caution (Why Poincaré stalls at a trivial Θ(n)\Theta(n) coalition).

Near p=1/2p=1/2, Poincaré guarantees an increment of only Θ(1/n)\Theta(1/n) per recruited coordinate. At rate 1/n1/n, moving p+(S)p_+(S) from 1/21/2 to 0.990.99 requires Θ(n)\Theta(n) coordinates—a constant fraction of all participants, which is trivial. To obtain a sublinear coalition of size o(n)o(n), we need an influence lower bound that beats Var⁡(h)/r\operatorname{Var}(h)/r by an unbounded factor on every intermediate Boolean function h=fSh=f_S.

Before trying to prove such an inequality, let us pause and ask a benchmark question: how small can the maximum individual influence max⁡iInf⁡i(f)\max_i\operatorname{Inf}_i(f) actually be for a balanced Boolean function? For Dictator it is 11; for Majority it is Θ(n−1/2)\Theta(n^{-1/2}). Can a balanced function push every coordinate’s influence all the way down to the Poincaré floor Θ(1/n)\Theta(1/n), or is there a genuine barrier above 1/n1/n?

1.10 Tribes: the Θ((log⁡n)/n)\Theta((\log n)/n) benchmark for individual influence

Ben-Or and Linial’s Tribes function [5] answers this benchmark question completely: a balanced Boolean function can push every individual influence down to Θ((log⁡n)/n)\Theta((\log n)/n), far smaller than Majority’s Θ(n−1/2)\Theta(n^{-1/2}), but stop just Θ(log⁡n)\Theta(\log n) above the Poincaré floor 1/n1/n.

Partition the n=mwn=mw coordinates into mm disjoint blocks (“tribes”) T1,…,TmT_1,\dots,T_m, each of width ww. Define f(x)=+1f(x)=+1 if at least one tribe consists entirely of +1+1 signs, and f(x)=−1f(x)=-1 otherwise (in 0/10/1 logic, an OR\mathrm{OR} of mm disjoint AND\mathrm{AND}s of width ww).

Because the tribes are disjoint, each tribe is unanimously +1+1 with probability 2−w2^{-w} independently of the others. Hence the probability that no tribe succeeds is P(f(X)=−1)=(1−2−w)m. \mathbb P(f(X)=-1)=(1-2^{-w})^m.

OR (m tribes) T₁ (AND) w bits T₂ (AND) w bits ⋯ Tₘ (AND) w bits n = m · w disjoint coordinates

Because (1−2−w)2w→e−1(1-2^{-w})^{2^w}\to e^{-1}, choosing m≈(ln⁡2)2wm\approx(\ln 2)2^w makes the all-fail probability (1−2−w)m≈e−ln⁡2=1/2(1-2^{-w})^m\approx e^{-\ln 2}=1/2.

To make ff approximately balanced (P(f(X)=−1)≈1/2\mathbb P(f(X)=-1)\approx 1/2), we choose mm and ww so that (1−2−w)m≈1/2(1-2^{-w})^m\approx 1/2. Taking logarithms gives mlog⁡(1−2−w)≈−log⁡2m\log(1-2^{-w})\approx -\log 2, so m∼(log⁡2)2wm\sim(\log 2)2^w and therefore n=mw∼(log⁡2)w2wn=mw\sim(\log 2)w2^w, which means w=log⁡2n−log⁡2log⁡2n+O(1)=Θ(log⁡n)w=\log_2 n - \log_2\log_2 n + O(1)=\Theta(\log n).

When is one coordinate pivotal in Tribes?

Participant ii (belonging to tribe T1T_1, say) can change the outcome of ff if and only if two independent events hold simultaneously:

Multiplying these two independent probabilities gives the exact influence of every coordinate: Inf⁡i(f)=2−(w−1)(1−2−w)m−1. \operatorname{Inf}_i(f)=2^{-(w-1)}(1-2^{-w})^{m-1}. Since 2−w∼log⁡2m=wlog⁡2n2^{-w}\sim\frac{\log 2}{m}=\frac{w\log 2}{n}, the detailed asymptotic check in Reading Contents gives Inf⁡i(f)=Θ ⁣(log⁡nn),I(f)=nInf⁡i(f)=Θ(log⁡n),Var⁡(f)→1. \operatorname{Inf}_i(f)=\Theta\!\left(\frac{\log n}{n}\right), \qquad I(f)=n\operatorname{Inf}_i(f)=\Theta(\log n), \qquad \operatorname{Var}(f)\to 1.

This single example teaches us two essential lessons for the rest of the chapter:

  1. Why Θ((log⁡n)/n)\Theta((\log n)/n) is the exact target for KKL: Because Tribes has Var⁡(f)≈1\operatorname{Var}(f)\approx 1 and max⁡iInf⁡i(f)=Θ((log⁡n)/n)\max_i\operatorname{Inf}_i(f)=\Theta((\log n)/n), no universal theorem can improve on Poincaré’s Var⁡(f)/n\operatorname{Var}(f)/n bound by more than a Θ(log⁡n)\Theta(\log n) factor. At the same time, a log⁡n\log n improvement is just enough to break the 1/n1/n barrier in §1.9 and reduce the greedy coalition size from O(n)O(n) to O(n/log⁡n)O(n/\log n)!
  2. Small individual influence does not mean coalition resilience: Even though every single coordinate in Tribes has tiny influence Θ((log⁡n)/n)\Theta((\log n)/n), a rushing coalition that corrupts just one full tribe of w=Θ(log⁡n)w=\Theta(\log n) participants can set all ww signs to +1+1 and force f=+1f=+1 with probability 11. Thus Tribes shows that KKL’s individual-influence bound Ω((log⁡n)/n)\Omega((\log n)/n) is tight, while leaving open whether even smaller coalitions might exist for specific rules.

1.11 Why Poincaré is never tight when influence is dispersed

Tribes shows that the best universal improvement we can hope for over Poincaré is a log⁡n\log n factor—or more precisely, a factor of log⁡(1/max⁡iInf⁡i(f))\log(1/\max_i\operatorname{Inf}_i(f)). Why did our baseline bound in §1.9 miss this logarithmic factor, and why must such a factor exist for every Boolean function?

Look back at the two-step chain that gave the 1/n1/n bound in §1.9: max⁡1≤i≤nInf⁡i(f)  ≥(1) averaging  I(f)n  ≥(2) Poincareˊ  Var⁡(f)n. \max_{1\le i\le n}\operatorname{Inf}_i(f) \;\underset{\text{(1) averaging}}{\ge}\; \frac{I(f)}{n} \;\underset{\text{(2) Poincar\'e}}{\ge}\; \frac{\operatorname{Var}(f)}{n}. For the combined bound max⁡iInf⁡i(f)≥Var⁡(f)/n\max_i\operatorname{Inf}_i(f)\ge\operatorname{Var}(f)/n to be tight (or even within a constant factor of tight), both inequalities (1) and (2) would have to be nearly tight at the same time:

Now compare what happens for arbitrary real-valued functions versus Boolean-valued functions:

  1. For real-valued functions g:{−1,1}n→Rg:\{-1,1\}^n\to\mathbb R, (1) and (2) can be tight simultaneously. Take the normalized linear sum g(x)=1n∑i=1nxi,Dig(x)=xin,∥Dig∥22=1n. g(x)=\frac{1}{\sqrt n}\sum_{i=1}^n x_i, \qquad D_i g(x)=\frac{x_i}{\sqrt n}, \qquad \|D_i g\|_2^2=\frac1n. Here gg is purely linear (W1(g)=Var⁡(g)=I(g)=1W_1(g)=\operatorname{Var}(g)=I(g)=1, so (2) holds with equality) and symmetric (max⁡i∥Dig∥22=1/n\max_i\|D_i g\|_2^2=1/n, so (1) holds with equality). This example delivers a crucial lesson: no purely L2L^2 argument can improve on Var⁡(f)/n\operatorname{Var}(f)/n. Any stronger inequality must exploit the Boolean range f(x)∈{−1,1}f(x)\in\{-1,1\}.
  2. For Boolean functions f:{−1,1}n→{−1,1}f:\{-1,1\}^n\to\{-1,1\}, (1) and (2) are in head-on conflict. Suppose a balanced Boolean function makes Poincaré (step (2)) tight: I(f)=Var⁡(f)=1I(f)=\operatorname{Var}(f)=1, so W1(f)=1W_1(f)=1. By our level-one rigidity lemma from Meeting 1 (Lemma 1 in §1.6; see also Theorem 9 in §2.4), the only Boolean functions with W1(f)=1W_1(f)=1 are the dictators f(x)=±xjf(x)=\pm x_j. And for a dictator, step (1) fails as badly as possible: one coordinate has influence 11 and the other n−1n-1 have influence 00, giving max⁡iInf⁡i(f)=1≫1/n\max_i\operatorname{Inf}_i(f)=1\gg 1/n! Conversely, if every coordinate has small influence (a=max⁡iInf⁡i(f)≪1a=\max_i\operatorname{Inf}_i(f)\ll 1, so we are in the dispersed regime where (1) is tight), ff is far from a dictator, so its variance cannot live on level 11—it is forced up to higher degrees k≥2k\ge 2, where I(f)=∑kk Wk(f)I(f)=\sum_k k\,W_k(f) pays a factor of kk and makes Poincaré (step (2)) loose.

Real vs. Boolean derivatives. For the real-valued counterexample g(x)=n−1/2∑ixig(x)=n^{-1/2}\sum_i x_i, the derivative Dig(x)=n−1/2xiD_i g(x)=n^{-1/2}x_i has tiny nonzero magnitude ∣Dig(x)∣=n−1/2|D_i g(x)|=n^{-1/2} at every vertex of the cube. For a Boolean function f:{−1,1}n→{−1,1}f:\{-1,1\}^n\to\{-1,1\}, Dif(x)=f(x)−f(x⊕i)2D_i f(x)=\frac{f(x)-f(x^{\oplus i})}{2} can only take values in {−1,0,1}\{-1,0,1\}: small influence ai=∥Dif∥22≪1a_i=\|D_i f\|_2^2\ll 1 therefore means DifD_i f is sparse (zero almost everywhere, and ±1\pm 1 on a rare pivotal set of measure aia_i).

Warm-up at degree 11: why small influences expel Fourier mass from W1(f)W_1(f)

Before tackling general degrees, let us see by a two-line calculation how the Boolean range f(x)∈{−1,1}f(x)\in\{-1,1\} prevents variance from sitting on level 11 whenever a=max⁡iInf⁡i(f)a=\max_i\operatorname{Inf}_i(f) is small. Write V=Var⁡(f),ai=Inf⁡i(f),a=max⁡1≤i≤nai. V=\operatorname{Var}(f),\qquad a_i=\operatorname{Inf}_i(f),\qquad a=\max_{1\le i\le n} a_i. Because ff is {−1,1}\{-1,1\}-valued, its discrete derivative hi=Dif∈{−1,0,1}h_i=D_i f\in\{-1,0,1\} is nonzero on a fraction ai=P(hi(X)≠0)a_i=\mathbb P(h_i(X)\ne 0) of the cube, so its L1L^1 norm is E∣hi(X)∣=ai\mathbb E|h_i(X)|=a_i (unlike the real-valued derivative ∣Dig∣≡n−1/2|D_i g|\equiv n^{-1/2}, whose L1L^1 norm is ai=n−1/2\sqrt{a_i}=n^{-1/2}). Recall from §1.8 that Dif(x)=∑S∋if^(S)χS(x)D_i f(x)=\sum_{S\ni i}\widehat f(S)\chi_S(x). Multiplying by xi∈{−1,1}x_i\in\{-1,1\} and taking expectations extracts the singleton coefficient S={i}S=\{i\}, so ∣f^({i})∣=∣E[XiDif(X)]∣≤E∣Dif(X)∣=ai. |\widehat f(\{i\})| =|\mathbb E[X_i D_i f(X)]| \le\mathbb E|D_i f(X)| =a_i. Squaring and summing over i=1,…,ni=1,\dots,n immediately bounds the entire level-11 Fourier weight: W1(f)=∑i=1nf^({i})2≤∑i=1nai2≤a∑i=1nai=a I(f). W_1(f) =\sum_{i=1}^n\widehat f(\{i\})^2 \le\sum_{i=1}^n a_i^2 \le a\sum_{i=1}^n a_i =a\,I(f). Look at what happened: because the {−1,0,1}\{-1,0,1\} sparsity of DifD_i f gave f^({i})2≤ai2\widehat f(\{i\})^2\le a_i^2 (an extra power of aia_i beyond the L2L^2 bound ∥Dif∥22=ai\|D_i f\|_2^2=a_i), the level-11 weight W1(f)W_1(f) is at most a I(f)a\,I(f). When every coordinate has tiny influence (a≪1a\ll 1), W1(f)≤a I(f)≪I(f)W_1(f)\le a\,I(f)\ll I(f) is only a tiny fraction of total influence I(f)I(f): thus if a noticeable fraction of the variance VV sits on level 11 (as in Majority, where W1≈2/πW_1\approx 2/\pi and I(Maj⁡n)≥W1/a∼2n/πI(\operatorname{Maj}_n)\ge W_1/a\sim\sqrt{2n/\pi}), I(f)I(f) is already huge of order V/aV/a; and if I(f)I(f) is not that large (as in Tribes, where I(f)∼log⁡nI(f)\sim\log n), then small individual influences force variance out of degree 11 and up to higher Fourier degrees.

From degree 11 to degree dd: pushing variance above degree dd

Being pushed from degree 11 to degree 22 would only double I(f)=∑kk Wk(f)I(f)=\sum_k k\,W_k(f). To prove I(f)≳log⁡(1/a) VI(f)\gtrsim\log(1/a)\,V, we want to generalize W1(f)≤a I(f)W_1(f)\le a\,I(f) to all degrees ≤d≍log⁡(1/a)\le d\asymp\log(1/a), showing that when a=max⁡iai≪1a=\max_i a_i\ll 1, the low-degree weight up to degree dd is a tiny fraction of I(f)I(f)!

Fix a degree cutoff d≥1d\ge 1 and split the variance V=∑S≠∅f^(S)2V=\sum_{S\ne\varnothing}\widehat f(S)^2 into its low-degree part (≤d\le d) and high-degree part (>d>d): V=Mlow+Mhigh,Mlow=∑1≤∣S∣≤df^(S)2,Mhigh=∑∣S∣>df^(S)2. V=M_{\mathrm{low}}+M_{\mathrm{high}}, \qquad M_{\mathrm{low}}=\sum_{1\le|S|\le d}\widehat f(S)^2, \qquad M_{\mathrm{high}}=\sum_{|S|>d}\widehat f(S)^2. Why do we win as soon as most of the variance sits in MhighM_{\mathrm{high}}? Because I(f)=∑S∣S∣f^(S)2I(f)=\sum_{S}|S|\widehat f(S)^2 multiplies every Fourier coefficient of degree ∣S∣>d|S|>d by at least d+1d+1: I(f)≥(d+1)∑∣S∣>df^(S)2=(d+1)Mhigh,or equivalently,Mhigh≤I(f)d+1. I(f)\ge(d+1)\sum_{|S|>d}\widehat f(S)^2=(d+1)M_{\mathrm{high}}, \qquad\text{or equivalently,}\qquad M_{\mathrm{high}}\le\frac{I(f)}{d+1}. Notice what this says: Mhigh/VM_{\mathrm{high}}/V is not small in the bottleneck regime—on the contrary, whenever Mhigh≥12VM_{\mathrm{high}}\ge\frac12 V, we immediately get I(f)≥d+12VI(f)\ge\frac{d+1}{2}V! The only way I(f)I(f) could avoid paying this (d+1)(d+1) factor is if the variance stayed at low degrees 1≤∣S∣≤d1\le |S|\le d (inside MlowM_{\mathrm{low}}). Therefore, our entire task is to generalize our degree-11 bound W1(f)≤a I(f)W_1(f)\le a\,I(f) to all degrees ≤d\le d, showing that when a≪1a\ll 1, Mlow≪I(f)M_{\mathrm{low}}\ll I(f) up to d≍log⁡(1/a)d\asymp\log(1/a)—so that combining Mlow≪I(f)M_{\mathrm{low}}\ll I(f) with Mhigh≤I(f)/(d+1)M_{\mathrm{high}}\le I(f)/(d+1) forces V=Mlow+Mhigh≲I(f)/(d+1)V=M_{\mathrm{low}}+M_{\mathrm{high}}\lesssim I(f)/(d+1).

What does small MlowM_{\mathrm{low}} mean combinatorially? For any subset S⊆[n]S\subseteq[n] of ∣S∣≤d|S|\le d coordinates, revealing XSX_S gives conditional mean E[f(X)∣XS]=∑T⊆Sf^(T)χT(XS)\mathbb E[f(X)\mid X_S]=\sum_{T\subseteq S}\widehat f(T)\chi_T(X_S), whose variance is Var⁡XS(E[f(X)∣XS])=∑∅≠T⊆Sf^(T)2≤Mlow\operatorname{Var}_{X_S}(\mathbb E[f(X)\mid X_S])=\sum_{\varnothing\ne T\subseteq S}\widehat f(T)^2\le M_{\mathrm{low}}. Thus Mlow≪VM_{\mathrm{low}}\ll V means that for any chosen set of at most dd coordinates, the conditional mean stays close to the global mean E[f]\mathbb E[f] for most revealed assignments XSX_S—revealing ≤d\le d bits gives almost no local hint about f(X)f(X), forcing its +1/−1+1/-1 regions to interweave finely with boundary I(f)≥(d+1)MhighI(f)\ge(d+1)M_{\mathrm{high}}.

To bound MlowM_{\mathrm{low}} in terms of the sparse derivatives hi=Dif∈{−1,0,1}h_i=D_i f\in\{-1,0,1\}, define the low-degree Fourier energy of hih_i by Bd(hi)=∑∣S∣≤dh^i(S)2=∑S∋i, ∣S∣≤df^(S)2. B_d(h_i)=\sum_{|S|\le d}\widehat h_i(S)^2=\sum_{S\ni i,\,|S|\le d}\widehat f(S)^2. Since every nonempty set SS has ∣S∣≥1|S|\ge 1 and is counted ∣S∣|S| times across ∑i=1nBd(hi)\sum_{i=1}^n B_d(h_i), we have Mlow=∑1≤∣S∣≤df^(S)2≤∑1≤∣S∣≤d∣S∣f^(S)2=∑i=1nBd(hi). M_{\mathrm{low}}=\sum_{1\le|S|\le d}\widehat f(S)^2 \le\sum_{1\le|S|\le d}|S|\widehat f(S)^2 =\sum_{i=1}^n B_d(h_i). Our entire problem now reduces to a single question about a sparse function hi∈{−1,0,1}h_i\in\{-1,0,1\} supported on a fraction ai=P(hi≠0)a_i=\mathbb P(h_i\ne 0) of the cube: how well can we bound its low-degree energy Bd(hi)B_d(h_i) in terms of aia_i?

Caution (Why naive L2L^2 and L1L^1 bounds on Bd(hi)B_d(h_i) fail for opposite reasons).

At degree 11, hih_i contributed only a single singleton coefficient f^({i})2≤ai2\widehat f(\{i\})^2\le a_i^2. Up to degree dd, however, Bd(hi)B_d(h_i) sums over all subsets ∣S∣≤d|S|\le d, and the two elementary bounds fail for opposite reasons:

  • Attempt 1 (Parseval / L2L^2 orthogonality): Dropping the degree restriction ∣S∣≤d|S|\le d gives Bd(hi)≤∑Sh^i(S)2=∥hi∥22=ai. B_d(h_i)\le\sum_S\widehat h_i(S)^2=\|h_i\|_2^2=a_i. This bound is dimension-free because it uses the orthogonality of the characters χS\chi_S, but it is completely blind to the Boolean sparsity hi∈{−1,0,1}h_i\in\{-1,0,1\}: the real-valued counterexample ∣Dig∣≡n−1/2|D_i g|\equiv n^{-1/2} has the exact same ∥Dig∥22=ai\|D_i g\|_2^2=a_i! Summing Bd(hi)≤aiB_d(h_i)\le a_i over ii just gives Mlow≤I(f)M_{\mathrm{low}}\le I(f), returning us to Poincaré.
  • Attempt 2 (Pointwise L1L^1 bound on each coefficient, as in our degree-11 warm-up): Because ∣hi∣∈{0,1}|h_i|\in\{0,1\}, we have E∣hi∣=ai\mathbb E|h_i|=a_i, so every individual Fourier coefficient satisfies ∣h^i(S)∣=∣E[hiχS]∣≤E∣hi∣=ai|\widehat h_i(S)|=|\mathbb E[h_i\chi_S]|\le\mathbb E|h_i|=a_i and hence h^i(S)2≤ai2\widehat h_i(S)^2\le a_i^2. At d=1d=1 there was only one term per ii; up to degree dd, summing h^i(S)2≤ai2\widehat h_i(S)^2\le a_i^2 over all sets of size at most dd yields Bd(hi)≤(∑k=0d(nk))ai2. B_d(h_i)\le\left(\sum_{k=0}^d\binom{n}{k}\right)a_i^2. This bound uses the {−1,0,1}\{-1,0,1\} sparsity of hih_i to gain an extra power of aia_i, but bounding each coefficient separately destroys orthogonality and pays the combinatorial factor ∑k=0d(nk)∼nd/d!\sum_{k=0}^d\binom{n}{k}\sim n^d/d!, which blows up with the ambient dimension nn.

Look at the complementary virtues of the two attempts. Attempt 1 is dimension-free because it keeps all characters together in an L2L^2 sum, while Attempt 2 extracts an extra power of aia_i because it measures the {−1,0,1}\{-1,0,1\}-valued derivative hih_i in a norm (L1L^1) with exponent smaller than 22. What we need is a special inequality for sparse functions on the hypercube that combines both virtues—an estimate of the form Bd(hi)≤Cdai1+η B_d(h_i)\le C^d a_i^{1+\eta} for universal constants C>1C>1 and η>0\eta>0 independent of nn (paying an exponential factor CdC^d in the degree dd rather than ndn^d in the dimension nn). Indeed, if such an estimate holds, summing over i=1,…,ni=1,\dots,n gives Mlow≤∑i=1nCdai1+η≤Cdaη∑i=1nai=CdaηI(f), M_{\mathrm{low}}\le\sum_{i=1}^n C^d a_i^{1+\eta}\le C^d a^\eta\sum_{i=1}^n a_i=C^d a^\eta I(f), and combining with Mhigh≤I(f)/(d+1)M_{\mathrm{high}}\le I(f)/(d+1) immediately yields V=Mlow+Mhigh≤I(f)(Cdaη+1d+1). V=M_{\mathrm{low}}+M_{\mathrm{high}}\le I(f)\left(C^d a^\eta+\frac{1}{d+1}\right). Choosing d≍log⁡(1/a)d\asymp\log(1/a) so that Cdaη≪1C^d a^\eta\ll 1 will then force V≲I(f)log⁡(1/a)V\lesssim \frac{I(f)}{\log(1/a)}, giving the exact logarithmic gain over Poincaré!

Our analytic goal is now crystal clear: for a sparse function h∈{−1,0,1}h\in\{-1,0,1\} with P(h≠0)=α\mathbb P(h\ne 0)=\alpha, prove a dimension-free low-degree bound Bd(h)≤Cdα1+ηB_d(h)\le C^d\alpha^{1+\eta}. The next two sections show how the noise operator TρT_\rho and Bonami’s hypercontractive inequality are built to do precisely this.

1.12 Using noise to extract a dimension-free rarity bound

Let h:{−1,1}n→{−1,0,1}h:\{-1,1\}^n\to\{-1,0,1\} be supported on a set A={x:h(x)≠0}A=\{x:h(x)\ne 0\} of probability α=P(A)\alpha=\mathbb P(A). Why was it hard to bound the low-degree sum Bd(h)=∑∣S∣≤dh^(S)2B_d(h)=\sum_{|S|\le d}\widehat h(S)^2 directly? Because the sharp degree cutoff 1{∣S∣≤d}\mathbf 1_{\{|S|\le d\}} has no clean representation in the physical domain {−1,1}n\{-1,1\}^n where the sparsity ∣h∣=1A|h|=\mathbf 1_A lives.

Step 1: Smooth degree damping replaces the sharp cutoff

In §1.4 we saw that the noise operator TρT_\rho multiplies each character χS\chi_S by ρ∣S∣\rho^{|S|}. In Meeting 1 we took ρ=−1/3\rho=-1/3 to encode ranking correlations; now take a fixed positive parameter ρ∈(0,1)\rho\in(0,1). By Parseval, the L2L^2 energy of the smoothed function TρhT_\rho h is ∥Tρh∥22=∑S⊆[n]ρ2∣S∣h^(S)2. \|T_\rho h\|_2^2=\sum_{S\subseteq[n]}\rho^{2|S|}\widehat h(S)^2. Instead of cutting off abruptly at degree dd, TρT_\rho damps every Fourier level smoothly by ρ2∣S∣\rho^{2|S|}. Because ρ2∣S∣≥ρ2d\rho^{2|S|}\ge\rho^{2d} whenever ∣S∣≤d|S|\le d, every low-degree term retains at least a ρ2d\rho^{2d} fraction of its squared mass in ∥Tρh∥22\|T_\rho h\|_2^2: Bd(h)=∑∣S∣≤dh^(S)2≤ρ−2d∑∣S∣≤dρ2∣S∣h^(S)2≤ρ−2d∥Tρh∥22. B_d(h)=\sum_{|S|\le d}\widehat h(S)^2 \le\rho^{-2d}\sum_{|S|\le d}\rho^{2|S|}\widehat h(S)^2 \le\rho^{-2d}\|T_\rho h\|_2^2. Look at what this simple inequality achieves: it produces the exact exponential degree factor Cd=ρ−2dC^d=\rho^{-2d} we anticipated in §1.11, with no dependence on the dimension nn! Moreover, unlike the sharp cutoff Bd(h)B_d(h), the operator TρT_\rho is a genuine averaging operator on the cube: for 0≤ρ≤10\le\rho\le 1, generating YY from xx by keeping each coordinate xix_i with probability ρ\rho and resampling it uniformly with probability 1−ρ1-\rho gives E[Yi∣x]=ρxi\mathbb E[Y_i\mid x]=\rho x_i, so Tρh(x)=E[h(Y)∣X=x]T_\rho h(x)=\mathbb E[h(Y)\mid X=x].

Step 2: What ∥Tρh∥22\|T_\rho h\|_2^2 means probabilistically

Before bounding ∥Tρh∥22\|T_\rho h\|_2^2 analytically, let us see what it measures in the physical domain. Because TρT_\rho is diagonal in the orthonormal basis {χS}\{\chi_S\}, it is self-adjoint and satisfies the semigroup law: ⟨g,Tρh⟩=∑Sρ∣S∣g^(S)h^(S)=⟨Tρg,h⟩,TρTσ=Tρσ. \langle g,T_\rho h\rangle=\sum_S\rho^{|S|}\widehat g(S)\widehat h(S)=\langle T_\rho g,h\rangle, \qquad T_\rho T_\sigma=T_{\rho\sigma}. Therefore, if (X,Z)(X,Z) is a pair of uniform inputs with coordinate correlation ρ2\rho^2, ∥Tρh∥22=⟨Tρh,Tρh⟩=⟨h,Tρ2h⟩=E[h(X)h(Z)]. \|T_\rho h\|_2^2=\langle T_\rho h,T_\rho h\rangle=\langle h,T_{\rho^2}h\rangle=\mathbb E[h(X)h(Z)]. When h=1Ah=\mathbf 1_A is the indicator of a rare event AA of measure α\alpha, ∥Tρ1A∥22=P(X∈A, Z∈A)\|T_\rho\mathbf 1_A\|_2^2=\mathbb P(X\in A,\,Z\in A) is simply the probability that AA persists after random noise! If we applied no noise (ρ=1\rho=1, so Z=XZ=X), this probability would be α\alpha; if we resampled completely (ρ=0\rho=0, so Z⊥XZ\perp X), it would be α2\alpha^2. For a fixed noise rate 0<ρ<10<\rho<1, a rare event should still struggle to survive resampling, so we expect P(X∈A,Z∈A)≤α1+η\mathbb P(X\in A,Z\in A)\le\alpha^{1+\eta} for some η>0\eta>0.

For example, take the two-coordinate subcube A={x1=x2=+1}A=\{x_1=x_2=+1\} of probability α=1/4\alpha=1/4, and set ρ2=1/3\rho^2=1/3. For each coordinate i∈{1,2}i\in\{1,2\}, keeping the sign with probability 1/31/3 and resampling otherwise gives P(Xi=Zi=+1)=12(13+23⋅12)=13. \mathbb P(X_i=Z_i=+1)=\frac12\left(\frac13+\frac23\cdot\frac12\right)=\frac13. By independence across coordinates, P(X∈A,Z∈A)=(1/3)2=1/9\mathbb P(X\in A,Z\in A)=(1/3)^2=1/9, which is already strictly smaller than α=1/4\alpha=1/4 (in fact, 1/9<1/8=α3/21/9 < 1/8 = \alpha^{3/2}). We can verify the exact same number from the Fourier expansion 1A(x)=1+x1+x2+x1x24\mathbf 1_A(x)=\frac{1+x_1+x_2+x_1x_2}{4} in Example 3: ∥Tρ1A∥22=∑S⊆{1,2}ρ2∣S∣(14)2=1+2ρ2+ρ416=(1+ρ24)2=19. \|T_\rho\mathbf 1_A\|_2^2=\sum_{S\subseteq\{1,2\}}\rho^{2|S|}\left(\frac14\right)^2=\frac{1+2\rho^2+\rho^4}{16}=\left(\frac{1+\rho^2}{4}\right)^2=\frac19. Our goal is to prove this α3/2\alpha^{3/2} persistence bound for every set AA and every {−1,0,1}\{-1,0,1\}-valued function hh of support α\alpha, regardless of the dimension nn or the geometry of AA.

Step 3: Recording rarity in L4/3L^{4/3} and dualizing to a 2→42\to 4 norm bound

How do we express the rarity of h∈{−1,0,1}h\in\{-1,0,1\} analytically without losing orthogonality? For any p≥1p\ge 1, define the LpL^p norm on the cube by ∥h∥p=(E∣h(X)∣p)1/p\|h\|_p=(\mathbb E|h(X)|^p)^{1/p}. Because ∣h(x)∣∈{0,1}|h(x)|\in\{0,1\} at every point, ∣h(x)∣p=1A(x)|h(x)|^p=\mathbf 1_A(x) for every exponent pp, so E∣h∣p=α\mathbb E|h|^p=\alpha and hence ∥h∥p=α1/p. \|h\|_p=\alpha^{1/p}. When p=2p=2, ∥h∥2=α1/2\|h\|_2=\alpha^{1/2}, so ∥h∥22=α\|h\|_2^2=\alpha gives no extra power of α\alpha. But whenever we choose an exponent p<2p<2, the 1/p1/p power is strictly larger than 1/21/2! In particular, for p=4/3p=4/3, ∥h∥4/3=α3/4,∥h∥4/32=α3/2. \|h\|_{4/3}=\alpha^{3/4}, \qquad \|h\|_{4/3}^2=\alpha^{3/2}. Thus, if we can prove the norm inequality ∥Tρh∥2≤∥h∥4/3\|T_\rho h\|_2\le\|h\|_{4/3} for some fixed ρ>0\rho>0, squaring both sides will immediately give ∥Tρh∥22≤∥h∥4/32=α3/2\|T_\rho h\|_2^2\le\|h\|_{4/3}^2=\alpha^{3/2} (so η=1/2\eta=1/2).

Why is p=4/3p=4/3 so convenient? Because its Hölder conjugate exponent qq (satisfying 1p+1q=1\frac{1}{p}+\frac{1}{q}=1, i.e., 34+14=1\frac34+\frac14=1) is q=4q=4, an even integer! Using the L2L^2 duality identity ∥u∥2=sup⁡∥g∥2=1∣⟨g,u⟩∣\|u\|_2=\sup_{\|g\|_2=1}|\langle g,u\rangle|, the self-adjointness of TρT_\rho, and Hölder’s inequality ∣E[hk]∣≤∥h∥4/3∥k∥4|\mathbb E[hk]|\le\|h\|_{4/3}\|k\|_4 (proved in Reading Contents), we can pass the operator TρT_\rho from the sparse function hh to a test function g∈L2g\in L^2: ∥Tρh∥2=sup⁡∥g∥2=1∣⟨g,Tρh⟩∣=sup⁡∥g∥2=1∣⟨Tρg,h⟩∣≤∥h∥4/3sup⁡∥g∥2=1∥Tρg∥4. \begin{aligned} \|T_\rho h\|_2 &=\sup_{\|g\|_2=1}|\langle g,T_\rho h\rangle|\\ &=\sup_{\|g\|_2=1}|\langle T_\rho g,h\rangle|\\ &\le\|h\|_{4/3}\sup_{\|g\|_2=1}\|T_\rho g\|_4. \end{aligned}

Recall (Hölder via double Cauchy–Schwarz). For conjugate exponents 1p+1q=1\frac1p+\frac1q=1, ∣E[hk]∣≤∥h∥p∥k∥q|\mathbb E[hk]|\le\|h\|_p\|k\|_q (proved for p=4/3,q=4p=4/3,q=4 in §2.11). When ∣h∣=1A|h|=\mathbf 1_A with P(A)=α\mathbb P(A)=\alpha, you can also see (4/3,4)(4/3,4) directly by applying ordinary Cauchy–Schwarz twice: ∣E[hk]∣≤E[1A∣k∣]≤α1/2(E[1Ak2])1/2≤α3/4∥k∥4=∥h∥4/3∥k∥4|\mathbb E[hk]|\le\mathbb E[\mathbf 1_A|k|]\le\alpha^{1/2}(\mathbb E[\mathbf 1_A k^2])^{1/2}\le\alpha^{3/4}\|k\|_4=\|h\|_{4/3}\|k\|_4.

Every step of our reduction is now complete. To establish ∥Tρh∥2≤∥h∥4/3\|T_\rho h\|_2\le\|h\|_{4/3}—and with it the rare-event bound Bd(hi)≤ρ−2dai3/2B_d(h_i)\le\rho^{-2d}a_i^{3/2} and KKL—we only need to find a fixed noise rate ρ>0\rho>0 such that ∥Tρg∥4≤∥g∥2\|T_\rho g\|_4\le\|g\|_2 for every function gg on {−1,1}n\{-1,1\}^n. And because q=4q=4 is an even integer, ∥Tρg∥44=E[(Tρg)4]\|T_\rho g\|_4^4=\mathbb E[(T_\rho g)^4] is an ordinary polynomial expectation that we can compute by hand!

1.13 How much averaging controls a fourth moment?

Let us appreciate what the inequality ∥Tρg∥4≤∥g∥2\|T_\rho g\|_4\le\|g\|_2 is asking for. On any probability space, Cauchy–Schwarz implies ∥g∥2=(E[g2])1/2≤(E[g4])1/4=∥g∥4\|g\|_2=(\mathbb E[g^2])^{1/2}\le(\mathbb E[g^4])^{1/4}=\|g\|_4: the L4L^4 norm is always larger than the L2L^2 norm because fourth powers heavily penalize sharp spikes. For instance, if g=α−1/21Ag=\alpha^{-1/2}\mathbf 1_A on a small set AA of measure α\alpha, then ∥g∥2=1\|g\|_2=1 whereas ∥g∥4=α−1/4→∞\|g\|_4=\alpha^{-1/4}\to\infty as α→0\alpha\to 0. For ∥Tρg∥4≤∥g∥2\|T_\rho g\|_4\le\|g\|_2 to hold, the noise averaging TρT_\rho must smooth out those spikes enough to reverse the natural ordering between L4L^4 and L2L^2. An operator that maps LpL^p into a higher norm LqL^q (q>pq>p) is called hypercontractive.

spike g: ‖g‖₄ ≫ ‖g‖₂ T_ρ g smoothing: ‖T_{1/√3} g‖₄ ≤ ‖g‖₂

A sparse spike g=α−1/21Ag=\alpha^{-1/2}\mathbf 1_A has ∥g∥2=1\|g\|_2=1 but huge fourth moment ∥g∥4=α−1/4\|g\|_4=\alpha^{-1/4}. Noise averaging T1/3T_{1/\sqrt 3} spreads the spike across neighbors so that its L4L^4 norm drops below ∥g∥2\|g\|_2.

How small must ρ\rho be for TρT_\rho to reverse the L4L^4-versus-L2L^2 comparison? We can discover the exact threshold ρ\rho on a single bit (n=1n=1), and then verify that the same threshold works in every dimension nn by induction.

On a single uniform sign x∈{−1,1}x\in\{-1,1\}, every real-valued function has the linear Fourier expansion g(x)=u+vxg(x)=u+vx, where u=g^(∅)u=\widehat g(\varnothing) and v=g^({1})v=\widehat g(\{1\}). Its second moment is ∥g∥22=u2+v2\|g\|_2^2=u^2+v^2, so ∥g∥24=(u2+v2)2=u4+2u2v2+v4. \|g\|_2^4=(u^2+v^2)^2=u^4+2u^2v^2+v^4. Now apply TρT_\rho: we get Tρg(x)=u+ρvxT_\rho g(x)=u+\rho vx. Averaging its fourth power over x=±1x=\pm1 kills the odd powers of xx and gives ∥Tρg∥44=(u+ρv)4+(u−ρv)42=u4+6ρ2u2v2+ρ4v4. \|T_\rho g\|_4^4 =\frac{(u+\rho v)^4+(u-\rho v)^4}{2} =u^4+6\rho^2u^2v^2+\rho^4v^4. Compare the two polynomials term by term: ∥g∥24\|g\|_2^4 has middle coefficient 2u2v22u^2v^2, whereas ∥Tρg∥44\|T_\rho g\|_4^4 has middle coefficient 6ρ2u2v26\rho^2 u^2v^2! Without noise (ρ=1\rho=1), the coefficient 66 exceeds 22. To make u4+6ρ2u2v2+ρ4v4≤u4+2u2v2+v4u^4+6\rho^2u^2v^2+\rho^4v^4\le u^4+2u^2v^2+v^4 hold for all u,v∈Ru,v\in\mathbb R, we need 6ρ2≤26\rho^2\le 2 (and ρ4≤1\rho^4\le 1), which forces ρ≤13. \rho\le\frac{1}{\sqrt 3}. At the exact threshold ρ=1/3\rho=1/\sqrt 3, we have 6ρ2=26\rho^2=2 and ρ4=1/9≤1\rho^4=1/9\le 1, so ∥T1/3g∥44=u4+2u2v2+19v4≤u4+2u2v2+v4=(u2+v2)2=∥g∥24. \|T_{1/\sqrt 3}g\|_4^4 =u^4+2u^2v^2+\frac19v^4 \le u^4+2u^2v^2+v^4 =(u^2+v^2)^2=\|g\|_2^4. Taking fourth roots proves ∥T1/3g∥4≤∥g∥2\|T_{1/\sqrt 3}g\|_4\le\|g\|_2 on a single bit! Remarkably, this one-bit calculation already gives the exact constant for the nn-dimensional hypercube.

Theorem 4 (The dimension-free 2→42\to4 hypercontractive inequality (Bonami)) For every n≥1n\ge 1 and every real-valued function gg on {−1,1}n\{-1,1\}^n, ∥T1/3g∥4≤∥g∥2. \|T_{1/\sqrt3}g\|_4\le\|g\|_2.

Historical note. The general hypercontractive inequality ∥Tρg∥q≤∥g∥p\|T_\rho g\|_q\le\|g\|_p for ρ2≤p−1q−1\rho^2\le\frac{p-1}{q-1} (1<p≤q<∞1<p\le q<\infty) was proved by Aline Bonami (1970) in harmonic analysis on Z2n\mathbb Z_2^n, and independently by Nelson (1973), Gross (1975, via logarithmic Sobolev inequalities), and Beckner (1975). Kahn, Kalai, and Linial [2] brought it into combinatorics and theoretical computer science in 1988.

Proof. We induct on nn. The base case n=1n=1 was just proved above.

For the induction step, split the input into the first n−1n-1 coordinates x′∈{−1,1}n−1x'\in\{-1,1\}^{n-1} and the last coordinate xn∈{−1,1}x_n\in\{-1,1\}, and decompose gg accordingly: g(x′,xn)=g0(x′)+xng1(x′),x′∈{−1,1}n−1. g(x',x_n)=g_0(x')+x_n g_1(x'), \qquad x'\in\{-1,1\}^{n-1}. Set ρ=1/3\rho=1/\sqrt3. Because TρT_\rho acts independently on each coordinate, applying TρT_\rho on {−1,1}n\{-1,1\}^n applies the (n−1)(n-1)-coordinate operator TρT_\rho to g0g_0 and g1g_1 and multiplies xnx_n by ρ\rho: Tρg(x′,xn)=A(x′)+ρxnB(x′),where A=Tρg0,  B=Tρg1. T_\rho g(x',x_n)=A(x')+\rho x_n B(x'), \qquad\text{where }A=T_\rho g_0,\; B=T_\rho g_1. Notice that A(x′)A(x') and B(x′)B(x') depend only on x′x' and are independent of the uniform sign xnx_n. Taking the expectation of (A(x′)+ρxnB(x′))4(A(x')+\rho x_n B(x'))^4 first over xn∈{−1,1}x_n\in\{-1,1\} (using E[xn]=E[xn3]=0\mathbb E[x_n]=\mathbb E[x_n^3]=0 and xn2=xn4=1x_n^2=x_n^4=1) and then over x′x' gives ∥Tρg∥44=E[A4]+6ρ2E[A2B2]+ρ4E[B4]. \|T_\rho g\|_4^4 =\mathbb E[A^4]+6\rho^2\mathbb E[A^2B^2]+\rho^4\mathbb E[B^4]. Write u=∥g0∥2u=\|g_0\|_2 and v=∥g1∥2v=\|g_1\|_2. By the induction hypothesis on {−1,1}n−1\{-1,1\}^{n-1}, we have ∥A∥4≤∥g0∥2=u\|A\|_4\le\|g_0\|_2=u and ∥B∥4≤∥g1∥2=v\|B\|_4\le\|g_1\|_2=v, so E[A4]≤u4,E[B4]≤v4. \mathbb E[A^4]\le u^4,\qquad \mathbb E[B^4]\le v^4. To bound the cross term E[A2B2]\mathbb E[A^2B^2], apply the Cauchy–Schwarz inequality to the random variables A2A^2 and B2B^2: E[A2B2]≤(E[A4])1/2(E[B4])1/2≤u2v2. \mathbb E[A^2B^2]\le(\mathbb E[A^4])^{1/2}(\mathbb E[B^4])^{1/2}\le u^2v^2. Substituting ρ=1/3\rho=1/\sqrt 3 (6ρ2=26\rho^2=2 and ρ4=1/9\rho^4=1/9) therefore yields ∥Tρg∥44≤u4+2u2v2+19v4≤(u2+v2)2. \|T_\rho g\|_4^4 \le u^4+2u^2v^2+\frac19v^4 \le(u^2+v^2)^2. Finally, averaging g(x′,xn)2=g0(x′)2+2xng0(x′)g1(x′)+g1(x′)2g(x',x_n)^2=g_0(x')^2+2x_n g_0(x')g_1(x')+g_1(x')^2 over xnx_n and x′x' gives ∥g∥22=∥g0∥22+∥g1∥22=u2+v2\|g\|_2^2=\|g_0\|_2^2+\|g_1\|_2^2=u^2+v^2. Thus ∥Tρg∥44≤∥g∥24\|T_\rho g\|_4^4\le\|g\|_2^4, and taking fourth roots completes the induction.

Look at why the induction succeeded without losing any constant factor across nn steps: Cauchy–Schwarz on E[A2B2]\mathbb E[A^2B^2] reduced the (n−1)(n-1)-dimensional cross term to the exact same algebraic inequality u4+2u2v2+19v4≤(u2+v2)2u^4+2u^2v^2+\frac19 v^4\le(u^2+v^2)^2 that we solved on a single bit. We now plug this dimension-free theorem directly into our rarity reduction from §1.12.

1.14 The rare-event estimate we needed

Set T=T1/3T=T_{1/\sqrt3}. With Theorem 4 in hand, the duality chain from §1.12 immediately yields our target bound on the smoothed energy of any rare function.

Corollary 1 (Smoothed energy of a rare response) For every real-valued function hh on {−1,1}n\{-1,1\}^n, ∥Th∥2≤∥h∥4/3. \|Th\|_2\le\|h\|_{4/3}. In particular, if h∈{−1,0,1}h\in\{-1,0,1\} is supported on a set A={x:h(x)≠0}A=\{x:h(x)\ne 0\} of probability α=P(A)\alpha=\mathbb P(A), then ∥Th∥22≤α3/2. \|Th\|_2^2\le\alpha^{3/2}.

Proof. In the duality inequality ∥Th∥2≤∥h∥4/3sup⁡∥g∥2=1∥Tg∥4\|Th\|_2\le\|h\|_{4/3}\sup_{\|g\|_2=1}\|Tg\|_4 from §1.12, Theorem 4 shows that sup⁡∥g∥2=1∥Tg∥4≤1\sup_{\|g\|_2=1}\|Tg\|_4\le 1, proving ∥Th∥2≤∥h∥4/3\|Th\|_2\le\|h\|_{4/3}. When ∣h∣=1A|h|=\mathbf 1_A with P(A)=α\mathbb P(A)=\alpha, we have ∥h∥4/3=α3/4\|h\|_{4/3}=\alpha^{3/4}, so squaring gives ∥Th∥22≤α3/2\|Th\|_2^2\le\alpha^{3/2}.

Before returning to derivatives, notice what Corollary 1 says about any event A⊆{−1,1}nA\subseteq\{-1,1\}^n of probability α>0\alpha>0: for a pair of inputs (X,Z)(X,Z) with coordinate correlation ρ2=1/3\rho^2=1/3, P(X∈A, Z∈A)=∥T1A∥22≤α3/2⟹P(Z∈A∣X∈A)≤α. \mathbb P(X\in A,\,Z\in A)=\|T\mathbf 1_A\|_2^2\le\alpha^{3/2} \quad\Longrightarrow\quad \mathbb P(Z\in A\mid X\in A)\le\sqrt\alpha. No matter how AA is embedded in the nn-dimensional cube, resampling two-thirds of the coordinates forces the conditional survival probability of AA below α\sqrt\alpha.

Now apply Corollary 1 to the pivotal derivative hi=Dif∈{−1,0,1}h_i=D_i f\in\{-1,0,1\}, which is nonzero on a set of probability ai=Inf⁡i(f)a_i=\operatorname{Inf}_i(f). Combining the degree-damping bound Bd(Dif)≤ρ−2d∥TDif∥22B_d(D_i f)\le\rho^{-2d}\|TD_i f\|_2^2 (where ρ−2=(1/3)−2=3\rho^{-2}=(1/\sqrt 3)^{-2}=3) with ∥TDif∥22≤ai3/2\|TD_i f\|_2^2\le a_i^{3/2} gives Bd(Dif)≤3d∥TDif∥22≤3dai3/2. B_d(D_i f)\le 3^d\|TD_i f\|_2^2\le 3^d a_i^{3/2}. This is precisely the missing low-degree rarity estimate Bd(hi)≤Cdai1+ηB_d(h_i)\le C^d a_i^{1+\eta} from §1.11, with concrete constants C=3C=3 and η=1/2\eta=1/2! Whenever 3dai<13^d\sqrt{a_i}<1, it beats the naive Parseval bound Bd(Dif)≤aiB_d(D_i f)\le a_i by a factor of 3dai3^d\sqrt{a_i}, paying 3d3^d for the degree dd rather than ndn^d for the dimension nn.

1.15 Closing the variance argument and obtaining KKL

We can now close the variance-splitting argument from §1.11 in three lines. Summing Bd(Dif)≤3dai3/2B_d(D_i f)\le 3^d a_i^{3/2} over all i=1,…,ni=1,\dots,n and using ai≤a=max⁡iaia_i\le a=\max_i a_i bounds the low-degree variance of ff: Mlow≤∑i=1nBd(Dif)≤3d∑i=1nai3/2≤3da∑i=1nai=3da I(f). \begin{aligned} M_{\mathrm{low}} &\le\sum_{i=1}^n B_d(D_i f)\\ &\le 3^d\sum_{i=1}^n a_i^{3/2} \le 3^d\sqrt a\sum_{i=1}^n a_i =3^d\sqrt a\,I(f). \end{aligned} Adding the high-degree bound Mhigh≤I(f)/(d+1)M_{\mathrm{high}}\le I(f)/(d+1) gives, for every integer d≥0d\ge0, V=Mlow+Mhigh≤I(f)(3da+1d+1). V=M_{\mathrm{low}}+M_{\mathrm{high}}\le I(f)\left(3^d\sqrt a+\frac1{d+1}\right). Look at the trade-off inside the parentheses as we vary the cutoff degree dd: raising dd pushes down the high-degree factor 1d+1\frac{1}{d+1} at the cost of growing 3da3^d\sqrt a. When aa is small, we can choose dd on the scale of log⁡(1/a)\log(1/a) so that 3d3^d absorbs only part of a\sqrt a (say 3d≤a−1/43^d\le a^{-1/4}, making 3da≤a1/43^d\sqrt a\le a^{1/4}), while 1d+1\frac{1}{d+1} shrinks to O(1/log⁡(1/a))O(1/\log(1/a)).

Lemma 2 (Maximal-influence trade-off) For every nonconstant Boolean function f:{−1,1}n→{−1,1}f:\{-1,1\}^n\to\{-1,1\} with 0<a=max⁡iInf⁡i(f)<10<a=\max_i\operatorname{Inf}_i(f)<1, I(f)≥Var⁡(f)6log⁡1a. I(f)\ge\frac{\operatorname{Var}(f)}6\log\frac1a.

Specifically, choosing d=⌊log⁡(1/a)4log⁡3⌋d=\lfloor\frac{\log(1/a)}{4\log 3}\rfloor ensures 3d≤a−1/43^d\le a^{-1/4}, so 3da+1d+1≤a1/4+1d+1≤6log⁡(1/a)3^d\sqrt a + \frac{1}{d+1}\le a^{1/4}+\frac{1}{d+1}\le\frac{6}{\log(1/a)} for all a∈(0,1)a\in(0,1) (the elementary calculus check is given in Reading Contents).

Theorem 5 (Kahn–Kalai–Linial (KKL) theorem) For every Boolean function f:{−1,1}n→{−1,1}f:\{-1,1\}^n\to\{-1,1\} on n≥2n\ge2 uniform bits, some coordinate i∈[n]i\in[n] satisfies Inf⁡i(f)≥112Var⁡(f)log⁡nn. \operatorname{Inf}_i(f)\ge\frac1{12}\operatorname{Var}(f)\frac{\log n}{n}. (The constant 1/121/12 is chosen for clean arithmetic rather than optimality.)

Proof. If ff is constant, Var⁡(f)=0\operatorname{Var}(f)=0 and the inequality is trivial. Otherwise let a=max⁡iInf⁡i(f)>0a=\max_i\operatorname{Inf}_i(f)>0. We distinguish two cases:

  • Case 1 (a≥n−1/2a\ge n^{-1/2}): Since Var⁡(f)≤1\operatorname{Var}(f)\le 1 and log⁡nn≤n−1/2\frac{\log n}{n}\le n^{-1/2} for all n≥2n\ge 2, we already have a≥n−1/2≥Var⁡(f)log⁡nn>112Var⁡(f)log⁡nna\ge n^{-1/2}\ge\operatorname{Var}(f)\frac{\log n}{n}>\frac{1}{12}\operatorname{Var}(f)\frac{\log n}{n}.
  • Case 2 (0<a<n−1/20<a<n^{-1/2}): Here log⁡1a>12log⁡n\log\frac1a>\frac12\log n. Using na≥∑i=1nai=I(f)n a\ge\sum_{i=1}^n a_i=I(f) together with Lemma 2 gives na≥I(f)≥Var⁡(f)6log⁡1a>Var⁡(f)12log⁡n, na\ge I(f)\ge\frac{\operatorname{Var}(f)}6\log\frac1a >\frac{\operatorname{Var}(f)}{12}\log n, and dividing by nn yields a>112Var⁡(f)log⁡nna>\frac{1}{12}\operatorname{Var}(f)\frac{\log n}{n}.

For any balanced Boolean function (Var⁡(f)=1\operatorname{Var}(f)=1), Theorem 5 guarantees a coordinate of influence at least log⁡n12n\frac{\log n}{12n}—matching the Θ((log⁡n)/n)\Theta((\log n)/n) benchmark of Tribes from §1.10! Even better, Theorem 5 requires no symmetry and no monotonicity, and its dependence on the bias of ff enters solely through the variance Var⁡(f)=4p(1−p)\operatorname{Var}(f)=4p(1-p). That universality is exactly what allows us to apply Theorem 5 repeatedly to every residual function fSf_S during our greedy coalition attack.

1.16 Completing the attack: a coalition of sublinear size

We now assemble the exact one-coordinate gain from §1.9 and the KKL theorem from §1.15 to solve the collective coin-flipping problem from §1.7.

Theorem 6 (A sublinear coalition can bias any balanced rule (Ben-Or–Linial / KKL)) There is a universal constant C>0C>0 such that, for every balanced function f:{−1,1}n→{−1,1}f:\{-1,1\}^n\to\{-1,1\} with n≥2n\ge2 and every 0<ε<1/20<\varepsilon<1/2, there exists a coalition S⊆[n]S\subseteq[n] of size ∣S∣≤C nlog⁡nlog⁡1ε |S|\le C\,\frac{n}{\log n}\log\frac1\varepsilon with a rushing strategy that forces +1+1 with probability p+(S)≥1−ε. p_+(S)\ge1-\varepsilon. In particular, for ε=0.01\varepsilon=0.01, a coalition of O(n/log⁡n)=o(n)O(n/\log n)=o(n) participants can force the output +1+1 with probability at least 0.990.99.

Proof. We build the coalition greedily starting from S0=∅S_0=\varnothing. At step t≥0t\ge 0, let ht=fSth_t=f_{S_t} be the residual Boolean function on the rt=n−tr_t=n-t untouched coordinates, and write pt=P(ht=+1)=p+(St),et=1−pt=P(ht=−1). p_t=\mathbb P(h_t=+1)=p_+(S_t), \qquad e_t=1-p_t=\mathbb P(h_t=-1). Since ff is balanced, p0=1/2p_0=1/2 (e0=1/2e_0=1/2), and adding coordinates to the coalition can only increase ptp_t (and decrease the failure probability ete_t). As long as et>εe_t>\varepsilon and t<n/2t<n/2, apply Theorem 5 to hth_t on its rtr_t-dimensional cube to choose a remaining coordinate it∉Sti_t\notin S_t of maximum influence, and set St+1=St∪{it}S_{t+1}=S_t\cup\{i_t\}.

By the one-step gain identity (§1.9), Theorem 5, and the variance formula Var⁡(ht)=4pt(1−pt)=4ptet\operatorname{Var}(h_t)=4p_t(1-p_t)=4p_t e_t, et−et+1=pt+1−pt=12Inf⁡it(ht)≥124Var⁡(ht)log⁡rtrt=16ptetlog⁡rtrt. \begin{aligned} e_t-e_{t+1}=p_{t+1}-p_t &=\frac12\operatorname{Inf}_{i_t}(h_t)\\ &\ge\frac1{24}\operatorname{Var}(h_t)\frac{\log r_t}{r_t}\\ &=\frac16 p_t e_t\frac{\log r_t}{r_t}. \end{aligned} Suppose n≥4n\ge4 and t≤n/2t\le n/2. Then n/2≤rt≤nn/2\le r_t\le n and pt≥p0=1/2p_t\ge p_0=1/2, so log⁡rtrt≥log⁡(n/2)n≥log⁡n2n\frac{\log r_t}{r_t}\ge\frac{\log(n/2)}{n}\ge\frac{\log n}{2n}. Substituting pt≥1/2p_t\ge 1/2 and log⁡rtrt≥log⁡n2n\frac{\log r_t}{r_t}\ge\frac{\log n}{2n} into the display above gives et−et+1≥16⋅12 et⋅log⁡n2n=log⁡n24net, e_t-e_{t+1}\ge\frac16\cdot\frac12\,e_t\cdot\frac{\log n}{2n}=\frac{\log n}{24n}e_t, and rearranging for et+1e_{t+1} yields the geometric decay et+1≤(1−log⁡n24n)et⟹et≤12exp⁡ ⁣(−tlog⁡n24n). e_{t+1}\le\left(1-\frac{\log n}{24n}\right)e_t \quad\Longrightarrow\quad e_t\le\frac12\exp\!\left(-\frac{t\log n}{24n}\right). Now set K=⌈24nlog⁡nlog⁡12ε⌉K=\lceil\frac{24n}{\log n}\log\frac{1}{2\varepsilon}\rceil. If K≤n/2K\le n/2, then after at most KK greedy steps the right-hand side is at most ε\varepsilon, so et≤εe_t\le\varepsilon and p+(St)≥1−εp_+(S_t)\ge 1-\varepsilon. If instead K>n/2K>n/2 (or if n∈{2,3}n\in\{2,3\}), the coalition can simply corrupt all nn coordinates and achieve p+([n])=1p_+([n])=1. In both cases, ∣S∣≤Cnlog⁡nlog⁡1ε|S|\le C\frac{n}{\log n}\log\frac1\varepsilon for a universal constant CC.

Finally, notice that the sequence of residual functions hth_t and chosen coordinates iti_t depends only on the truth table of ff, so the coalition SS is completely determined before any honest bits are broadcast. Once the honest participants broadcast x[n]∖Sx_{[n]\setminus S} with fS(x[n]∖S)=+1f_S(x_{[n]\setminus S})=+1, the coalition simply outputs a witnessing assignment yS∈{−1,1}Sy_S\in\{-1,1\}^S with f(x[n]∖S,yS)=+1f(x_{[n]\setminus S},y_S)=+1.

Applying Theorem 6 to −f-f produces a coalition of the same size that forces −1-1 with probability at least 1−ε1-\varepsilon (and taking the union of the two coalitions gives a single coalition that can steer the bit in either direction on almost all honest inputs; see Reading Contents).

Caution (Single-step KKL tightness vs. multi-step coalition resilience).

The Tribes function (§1.10) makes the single-step KKL bound max⁡iInf⁡i(f)=Ω((log⁡n)/n)\max_i\operatorname{Inf}_i(f)=\Omega((\log n)/n) tight, and forcing Tribes⁡m,w=−1\operatorname{Tribes}_{m,w}=-1 indeed requires corrupting one coordinate in almost every tribe (Ω(n/log⁡n)\Omega(n/\log n) coordinates). Yet to force Tribes⁡m,w=+1\operatorname{Tribes}_{m,w}=+1, a coalition only needs a single tribe of size w=Θ(log⁡n)w=\Theta(\log n)! Why does greedy recruitment finish in ww steps instead of Θ(n/log⁡n)\Theta(n/\log n)? Because as soon as the coalition sets one coordinate i1∈T1i_1\in T_1 to +1+1, the remaining w−1w-1 members of T1T_1 see their influence on the residual function f{i1}f_{\{i_1\}} double from 2−w2^{-w} to 2−(w−1)2^{-(w-1)}, then 2−(w−2)2^{-(w-2)}, and so on—snowballing inside T1T_1 rather than staying at Θ((log⁡n)/n)\Theta((\log n)/n).

Let us step back and take stock of both classroom stories and the scope of what we proved:

II. Reading Contents

The first two sections retain black-box testing as another complete application of Fourier analysis. The remaining sections supply calculations and extensions supporting the two classroom stories.

2.1 Another use of the basis: BLR linearity testing

Suppose a program answers queries to an unknown function g:{0,1}n→{0,1}g:\{0,1\}^n\to\{0,1\}. A linear function has the form ℓa(u)=a1u1+⋯+anun(mod2),a∈{0,1}n. \ell_a(u)=a_1u_1+\cdots+a_nu_n\pmod2, \qquad a\in\{0,1\}^n. The vector aa is not supplied to us. We want to test whether the program follows some linear rule, without reading all 2n2^n entries of its truth table.

Exact certification is too strong for a small number of queries. A program could disagree with a linear function at a single unqueried input. We therefore measure distance by the fraction of inputs on which two functions disagree: dist⁡(g,h)=PU uniform(g(U)≠h(U)). \operatorname{dist}(g,h)=\mathbb P_{U\text{ uniform}}(g(U)\ne h(U)). Our task is a promise problem: accept every linear function, and reject with high probability whenever gg has distance at least ε\varepsilon from every linear function. Functions between these two cases need not receive a prescribed answer.

The desired query bound should depend on ε\varepsilon and the failure probability, but not on 2n2^n. Query complexity counts calls to the program; generating an nn-bit query still takes time proportional to nn.

A local identity and its three-query test

Every linear function obeys g(u⊕v)=g(u)+g(v)(mod2), g(u\oplus v)=g(u)+g(v)\pmod2, where u⊕vu\oplus v is coordinatewise addition modulo two. Conversely, if the identity holds for every pair, then g(0)=0g(0)=0, and repeatedly applying it to the standard basis vectors gives g(u)=∑iuig(ei)(mod2)g(u)=\sum_i u_i g(e_i)\pmod2. The identity characterizes linearity exactly.

This suggests a randomized test: choose independent uniform U,VU,V, query g(U)g(U), g(V)g(V), and g(U⊕V)g(U\oplus V), and check the identity. Linear functions always pass. The difficulty is the converse under errors: why should satisfying most local identities force agreement with one global linear function?

To analyze the test, set xi=(−1)uix_i=(-1)^{u_i} and f(x)=(−1)g(u). f(x)=(-1)^{g(u)}. Addition of bits becomes multiplication of signs. Write xy=(x1y1,…,xnyn)xy=(x_1y_1,\ldots,x_ny_n). A linear rule becomes a parity χS(x)=∏i∈Sxi,S={i:ai=1}, \chi_S(x)=\prod_{i\in S}x_i, \qquad S=\{i:a_i=1\}, and its local identity becomes χS(xy)=χS(x)χS(y)\chi_S(xy)=\chi_S(x)\chi_S(y).

Definition 6 (The BLR test) Choose independent uniform X,Y∈{−1,1}nX,Y\in\{-1,1\}^n. Query f(X)f(X), f(Y)f(Y), and f(XY)f(XY), and accept if f(X)f(Y)=f(XY). f(X)f(Y)=f(XY). The test makes at most three queries. The empty parity χ∅=1\chi_\varnothing=1 is included among the linear rules.

Recall (Pairwise vs. mutual independence). The three query points (X,Y,XY)(X,Y,XY) are pairwise independent—any two are uniform and independent—but completely dependent as a triple since X⋅Y⋅(XY)=1X\cdot Y\cdot(XY)=\mathbf 1. This is the nn-dimensional cousin of the classic three-coin-event counterexample from introductory probability.

The local test statistic involves three values of ff. The character basis simplifies the products and averages needed to analyze it. The character basis developed in Course Content provides that representation.

The cubic statistic

The three query locations are related: XYXY is determined by X,YX,Y. Each pair among X,Y,XYX,Y,XY is independent, but the three are not jointly independent. We must retain this relation in the calculation.

Since the queried values are signs, the acceptance indicator is 1accept=1+f(X)f(Y)f(XY)2. \mathbf 1_{\mathrm{accept}}=\frac{1+f(X)f(Y)f(XY)}2. Expand the three factors in the character basis. A term with indices A,B,CA,B,C has expectation E[χA(X)χB(Y)χC(XY)]=E[χA(X)χC(X)] E[χB(Y)χC(Y)]=1A=C1B=C. \begin{aligned} \mathbb E[\chi_A(X)\chi_B(Y)\chi_C(XY)] &=\mathbb E[\chi_A(X)\chi_C(X)]\, \mathbb E[\chi_B(Y)\chi_C(Y)]\\ &=\mathbf 1_{A=C}\mathbf 1_{B=C}. \end{aligned} The first step uses the character identity and independence of X,YX,Y; the second uses orthogonality. All terms disappear except those with A=B=CA=B=C. Therefore E[f(X)f(Y)f(XY)]=∑Sf^(S)3, \mathbb E[f(X)f(Y)f(XY)]=\sum_S\widehat f(S)^3, and the rejection probability δ\delta satisfies 1−2δ=∑Sf^(S)3. 1-2\delta=\sum_S\widehat f(S)^3. We have not computed the exponentially many coefficients. We have used them to express a statistic that the tester can sample using three queries.

Soundness and query amplification

Theorem 7 (Soundness of the BLR test) If the BLR test rejects ff with probability δ\delta, then some parity χS\chi_S satisfies dist⁡(f,χS)≤δ. \operatorname{dist}(f,\chi_S)\le\delta. Equivalently, if ff is at distance at least ε\varepsilon from every parity, then a single test rejects with probability at least ε\varepsilon.

Proof. Let M=max⁡Sf^(S)M=\max_S\widehat f(S). Squared coefficients are nonnegative, so Parseval gives 1−2δ=∑Sf^(S)3≤M∑Sf^(S)2=M. 1-2\delta=\sum_S\widehat f(S)^3 \le M\sum_S\widehat f(S)^2=M. Choose SS attaining MM. The correlation–distance identity yields dist⁡(f,χS)=1−M2≤δ. \operatorname{dist}(f,\chi_S)=\frac{1-M}{2}\le\delta.

The maximum here is of the coefficients themselves, not their absolute values. A negative parity −χS-\chi_S corresponds to an affine rule with constant term one, rather than a linear map, and fails every BLR test. This distinction is essential to completeness.

To finish the opening algorithmic problem, repeat the test independently rr times and accept only if every trial accepts. A linear function is always accepted. If ff is ε\varepsilon-far, then P(all trials accept)≤(1−ε)r≤e−εr. \mathbb P(\text{all trials accept})\le(1-\varepsilon)^r\le e^{-\varepsilon r}. For failure probability 0<η<10<\eta<1, take r=⌈log⁡(1/η)ε⌉. r=\left\lceil\frac{\log(1/\eta)}{\varepsilon}\right\rceil. The algorithm uses at most 3r3r queries, independent of nn. It certifies closeness through a probabilistic promise guarantee; passing finitely many tests does not certify exact correctness at every input. This is the BLR theorem of Blum, Luby, and Rubinfeld [1], with the Fourier proof presented in O’Donnell’s course and textbook [3,4].

2.2 Recovering one corrupted parity entry

The local identity does more than test a table. Suppose ff is promised to be within distance α<1/4\alpha<1/4 of an unknown parity χS\chi_S. We are given a particular input xx, and want the correct value χS(x)\chi_S(x) even if the table entry f(x)f(x) is corrupted.

Querying f(x)f(x) may repeatedly read the same corrupted entry. The parity identity gives another route: for any yy, the correct value at xx equals χS(y)χS(xy)\chi_S(y)\chi_S(xy). We can choose yy uniformly so that both queried locations are uniform, spreading the risk over the whole table. This suggests returning f(Y)f(xY). f(Y)f(xY). For the uncorrupted parity, χS(Y)χS(xY)=χS(x)\chi_S(Y)\chi_S(xY)=\chi_S(x). Both YY and xYxY are uniform, so each queried entry is corrupted with probability at most α\alpha. They need not be independent; a union bound is sufficient.

Proposition 1 (Two-query local correction) If dist⁡(f,χS)≤α\operatorname{dist}(f,\chi_S)\le\alpha, then for every fixed xx, PY(f(Y)f(xY)=χS(x))≥1−2α. \mathbb P_Y(f(Y)f(xY)=\chi_S(x))\ge1-2\alpha.

The guarantee holds at every target input, with probability over the decoder’s randomness. If α<1/4\alpha<1/4, independent repetitions and a majority vote amplify it. Writing γ=1/2−2α>0\gamma=1/2-2\alpha>0, rr repetitions have failure probability at most e−2rγ2e^{-2r\gamma^2} by Hoeffding. For a fixed error bound below 1/41/4, the number of queries needed for failure probability η\eta is O(log⁡(1/η))O(\log(1/\eta)).

The table (χS(x))x∈{−1,1}n(\chi_S(x))_{x\in\{-1,1\}^n} is a Hadamard codeword encoding SS. Testing asks whether a received table is close to a codeword; local correction recovers one specified codeword entry. Neither procedure needs to read the whole table or identify SS. The BLR test is therefore a concrete example of local information controlling global algebraic structure.

2.3 Further Fourier expansions: monotone rules, Maj⁡n\operatorname{Maj}_n, and Tribes

In §1.3 (Example 4 and Example 5) we computed the full expansion of Maj⁡3\operatorname{Maj}_3 and the degree-11 coefficients of Maj⁡n\operatorname{Maj}_n. Here we develop three extensions that connect those calculations to the rest of the chapter: why degree-11 Fourier coefficients equal influences for all monotone functions, how the higher odd levels of Maj⁡n\operatorname{Maj}_n behave, and how to write down the exact Fourier expansion of Tribes.

Degree-one coefficients of monotone functions

A Boolean function f:{−1,1}n→{−1,1}f:\{-1,1\}^n\to\{-1,1\} is monotone (or increasing) if flipping any coordinate from −1-1 to +1+1 can never decrease the output: f(x−i,+1)≥f(x−i,−1)f(x_{-i},+1)\ge f(x_{-i},-1) for every i∈[n]i\in[n] and x−i∈{−1,1}[n]∖{i}x_{-i}\in\{-1,1\}^{[n]\setminus\{i\}}. Majority, ANDw\mathrm{AND}_w, ORw\mathrm{OR}_w, and Tribes are all monotone, whereas Parity (for n≥2n\ge 2) is not.

Proposition 2 (Linear Fourier coefficients of a monotone Boolean function) If f:{−1,1}n→{−1,1}f:\{-1,1\}^n\to\{-1,1\} is monotone, then for every coordinate i∈[n]i\in[n], f^({i})=Inf⁡i(f),∑i=1nf^({i})=I(f). \widehat f(\{i\})=\operatorname{Inf}_i(f), \qquad \sum_{i=1}^n\widehat f(\{i\})=I(f).

Proof. Recall from §1.8 that Dif(x)=f(x)−f(x⊕i)2=∑S∋if^(S)χS(x)D_i f(x)=\frac{f(x)-f(x^{\oplus i})}{2}=\sum_{S\ni i}\widehat f(S)\chi_S(x). Multiplying by xi∈{−1,1}x_i\in\{-1,1\} gives xiDif(x)=f(x−i,+1)−f(x−i,−1)2=∑S∋if^(S)χS∖{i}(x). x_i D_i f(x)=\frac{f(x_{-i},+1)-f(x_{-i},-1)}{2}=\sum_{S\ni i}\widehat f(S)\chi_{S\setminus\{i\}}(x). Taking expectations over XX extracts the S={i}S=\{i\} term on the right: E[XiDif(X)]=f^({i})\mathbb E[X_i D_i f(X)]=\widehat f(\{i\}). On the other hand, because ff is monotone and {−1,1}\{-1,1\}-valued, the difference f(x−i,+1)−f(x−i,−1)2\frac{f(x_{-i},+1)-f(x_{-i},-1)}{2} is +1+1 whenever coordinate ii is pivotal and 00 otherwise—it is never −1-1! Hence xiDif(x)=1{f(x)≠f(x⊕i)}=(Dif(x))2x_i D_i f(x)=\mathbf 1_{\{f(x)\ne f(x^{\oplus i})\}}=(D_i f(x))^2 at every input xx. Taking expectations gives f^({i})=E[(Dif(X))2]=Inf⁡i(f)\widehat f(\{i\})=\mathbb E[(D_i f(X))^2]=\operatorname{Inf}_i(f).

Notice the sharp contrast between the two formulas for Inf⁡i(f)\operatorname{Inf}_i(f) when ff is monotone: Inf⁡i(f)=f^({i})=∑S∋if^(S)2. \operatorname{Inf}_i(f)=\widehat f(\{i\})=\sum_{S\ni i}\widehat f(S)^2. For Maj⁡n\operatorname{Maj}_n, where Inf⁡i(Maj⁡n)∼2πn\operatorname{Inf}_i(\operatorname{Maj}_n)\sim\sqrt{\frac{2}{\pi n}}, summing the unsquared degree-11 coefficients gives I(Maj⁡n)=∑i=1nMaj⁡n^({i})∼2nπI(\operatorname{Maj}_n)=\sum_{i=1}^n\widehat{\operatorname{Maj}_n}(\{i\})\sim\sqrt{\frac{2n}{\pi}}, whereas summing their squares gives the level-11 Fourier weight W1(Maj⁡n)=∑i=1nMaj⁡n^({i})2∼2πW_1(\operatorname{Maj}_n)=\sum_{i=1}^n\widehat{\operatorname{Maj}_n}(\{i\})^2\sim\frac{2}{\pi}.

Higher Fourier levels of Maj⁡n\operatorname{Maj}_n and five-voter Majority

How do we compute Maj⁡n^(S)\widehat{\operatorname{Maj}_n}(S) for larger odd subsets ∣S∣=k|S|=k? Let us first work out n=5n=5 explicitly, and then record the general formula.

Example 7 (Five-voter Majority Maj⁡5\operatorname{Maj}_5 and higher odd degrees) By oddness (Maj⁡5(−x)=−Maj⁡5(x)\operatorname{Maj}_5(-x)=-\operatorname{Maj}_5(x)) and permutation symmetry, only odd levels k∈{1,3,5}k\in\{1,3,5\} appear, and all coefficients within a level are equal: Maj⁡5(x)=a∑∣S∣=1χS(x)+b∑∣S∣=3χS(x)+c χ[5](x). \operatorname{Maj}_5(x) =a\sum_{|S|=1}\chi_S(x)+b\sum_{|S|=3}\chi_S(x)+c\,\chi_{[5]}(x). From Example 5, we already know a=Maj⁡5^({1})=2−4(42)=616=38a=\widehat{\operatorname{Maj}_5}(\{1\})=2^{-4}\binom{4}{2}=\frac{6}{16}=\frac38. To find bb and cc, evaluate Maj⁡5(x)=+1\operatorname{Maj}_5(x)=+1 on two inputs with different numbers of +1+1’s:

  • At x=(+1,+1,+1,+1,+1)x=(+1,+1,+1,+1,+1), every character χS(x)\chi_S(x) equals +1+1, so 5a+10b+c=15a+10b+c=1.
  • At x=(+1,+1,+1,+1,−1)x=(+1,+1,+1,+1,-1) (four +1+1’s and one −1-1), the singleton sum is 4−1=34-1=3, the (53)=10\binom{5}{3}=10 triples split into (43)=4\binom{4}{3}=4 with product +1+1 and (42)=6\binom{4}{2}=6 with product −1-1 (sum 4−6=−24-6=-2), and χ[5](x)=−1\chi_{[5]}(x)=-1, giving 3a−2b−c=13a-2b-c=1.

Adding the two equations gives 8a+8b=28a+8b=2, so b=14−a=−18b=\frac14-a=-\frac18, and then c=1−5a−10b=38c=1-5a-10b=\frac38: Maj⁡5(x)=38∑∣S∣=1χS(x)−18∑∣S∣=3χS(x)+38χ[5](x). \operatorname{Maj}_5(x) =\frac38\sum_{|S|=1}\chi_S(x)-\frac18\sum_{|S|=3}\chi_S(x)+\frac38\chi_{[5]}(x). Its spectral weights across levels k=1,3,5k=1,3,5 are W1(Maj⁡5)=5(38)2=4564,W3(Maj⁡5)=10(−18)2=1064,W5(Maj⁡5)=(38)2=964, W_1(\operatorname{Maj}_5)=5\left(\frac38\right)^2=\frac{45}{64}, \qquad W_3(\operatorname{Maj}_5)=10\left(-\frac18\right)^2=\frac{10}{64}, \qquad W_5(\operatorname{Maj}_5)=\left(\frac38\right)^2=\frac{9}{64}, which sum to 45+10+964=1\frac{45+10+9}{64}=1.

More generally, for any odd n=2m+1n=2m+1 and any subset S⊆[n]S\subseteq[n] of odd size k=2r+1k=2r+1, picking an element i∈Si\in S and using xiDiMaj⁡n(x)=1{∑j≠ixj=0}x_i D_i\operatorname{Maj}_n(x)=\mathbf 1_{\{\sum_{j\ne i}x_j=0\}} expresses Maj⁡n^(S)\widehat{\operatorname{Maj}_n}(S) as the expectation of ∏j∈S∖{i}Xj\prod_{j\in S\setminus\{i\}}X_j on the tie event {∑j≠iXj=0}\{\sum_{j\ne i}X_j=0\}, yielding the closed form (see O’Donnell [3], Theorem 5.19): Maj⁡n^(S)=(−1)r(mr)(2m2r) 2−2m(2mm)(∣S∣=2r+1,  n=2m+1). \widehat{\operatorname{Maj}_n}(S) =(-1)^r\frac{\binom{m}{r}}{\binom{2m}{2r}}\,2^{-2m}\binom{2m}{m} \qquad(|S|=2r+1,\;n=2m+1). As n→∞n\to\infty, the level-kk weight Wk(Maj⁡n)W_k(\operatorname{Maj}_n) converges for each odd k≥1k\ge 1 to 2πk(k−1(k−1)/2)2−(k−1)\frac{2}{\pi k}\binom{k-1}{(k-1)/2} 2^{-(k-1)} (W1→2π≈0.6366W_1\to\frac{2}{\pi}\approx 0.6366, W3→13π≈0.1061W_3\to\frac{1}{3\pi}\approx 0.1061, W5→320π≈0.0477W_5\to\frac{3}{20\pi}\approx 0.0477, …\ldots). Under Kalai’s cycle formula (§1.5), the central limit theorem implies that as n→∞n\to\infty, the correlated pair (1n∑iXi,1n∑iYi)(\frac{1}{\sqrt n}\sum_i X_i,\frac{1}{\sqrt n}\sum_i Y_i) converges to a bivariate normal pair with correlation ρ=−1/3\rho=-1/3, so Stab⁡−1/3(Maj⁡n)→−2πarcsin⁡(1/3)\operatorname{Stab}_{-1/3}(\operatorname{Maj}_n)\to-\frac{2}{\pi}\arcsin(1/3) and lim⁡n→∞P(cycle under Maj⁡n)=14−32πarcsin⁡ ⁣(13)≈0.0877. \lim_{n\to\infty}\mathbb P(\text{cycle under }\operatorname{Maj}_n) =\frac14-\frac{3}{2\pi}\arcsin\!\left(\frac13\right)\approx 0.0877.

Exact Fourier expansion of Tribes

Finally, we can combine the subcube expansion from Example 3 with independence across disjoint blocks to write down every Fourier coefficient of the Tribes function from §1.10.

Example 8 (Fourier coefficients of Tribes⁡m,w\operatorname{Tribes}_{m,w}) Recall that Tribes⁡m,w\operatorname{Tribes}_{m,w} partitions [n][n] into mm disjoint tribes T1,…,TmT_1,\ldots,T_m of width ww, outputting −1-1 if and only if every tribe TjT_j fails to be unanimously +1+1. Let 1Aj(x)=∏i∈Tj1+xi2=2−w∑S⊆TjχS(x)\mathbf 1_{A_j}(x)=\prod_{i\in T_j}\frac{1+x_i}{2}=2^{-w}\sum_{S\subseteq T_j}\chi_S(x) be the 0/10/1 indicator that tribe TjT_j is unanimously +1+1. Then the 0/10/1 indicator of {Tribes⁡m,w(x)=−1}\{\operatorname{Tribes}_{m,w}(x)=-1\} is the product ∏j=1m(1−1Aj(x))\prod_{j=1}^m(1-\mathbf 1_{A_j}(x)), so in {−1,1}\{-1,1\} sign notation: Tribes⁡m,w(x)=1−2∏j=1m(1−1Aj(x)). \operatorname{Tribes}_{m,w}(x) =1-2\prod_{j=1}^m\left(1-\mathbf 1_{A_j}(x)\right). Now substitute 1−1Aj(x)=(1−2−w)−2−w∑∅≠Sj⊆TjχSj(x)1-\mathbf 1_{A_j}(x)=(1-2^{-w})-2^{-w}\sum_{\varnothing\ne S_j\subseteq T_j}\chi_{S_j}(x) into the product. Because the tribes T1,…,TmT_1,\ldots,T_m are pairwise disjoint, any subset S⊆[n]S\subseteq[n] decomposes uniquely into its intersections Sj=S∩TjS_j=S\cap T_j with the mm tribes! Expanding the product across j=1,…,mj=1,\ldots,m immediately gives every Fourier coefficient:

  • For S=∅S=\varnothing, picking the constant term 1−2−w1-2^{-w} from all mm factors gives Tribes⁡m,w^(∅)=1−2(1−2−w)m. \widehat{\operatorname{Tribes}_{m,w}}(\varnothing)=1-2(1-2^{-w})^m.
  • For any nonempty S⊆[n]S\subseteq[n], let r(S)=∣{j∈[m]:S∩Tj≠∅}∣≥1r(S)=\bigl|\{j\in[m]:S\cap T_j\ne\varnothing\}\bigr|\ge 1 be the number of tribes that SS intersects. Picking −2−wχS∩Tj(x)-2^{-w}\chi_{S\cap T_j}(x) from those r(S)r(S) tribes and the constant 1−2−w1-2^{-w} from the remaining m−r(S)m-r(S) tribes gives Tribes⁡m,w^(S)=−2(−2−w)r(S)(1−2−w)m−r(S). \widehat{\operatorname{Tribes}_{m,w}}(S) =-2\left(-2^{-w}\right)^{r(S)}(1-2^{-w})^{m-r(S)}.

Notice how this formula checks against our earlier calculations: when S={i}S=\{i\} is a singleton, it intersects r({i})=1r(\{i\})=1 tribe, so Tribes⁡m,w^({i})=2−(w−1)(1−2−w)m−1\widehat{\operatorname{Tribes}_{m,w}}(\{i\})=2^{-(w-1)}(1-2^{-w})^{m-1}—which matches the exact influence Inf⁡i(Tribes⁡m,w)\operatorname{Inf}_i(\operatorname{Tribes}_{m,w}) from §1.10, just as Proposition 2 requires! Moreover, unlike Majority (where 2/π2/\pi of the Fourier weight sits on level 11), a singleton in Tribes has squared coefficient f^({i})2=Θ(log⁡2nn2)\widehat f(\{i\})^2=\Theta(\frac{\log^2 n}{n^2}), so the total level-11 weight of Tribes is W1(Tribes⁡m,w)=n f^({1})2=Θ(log⁡2nn)→0W_1(\operatorname{Tribes}_{m,w})=n\,\widehat f(\{1\})^2=\Theta(\frac{\log^2 n}{n})\to 0. Almost all of Tribes’ variance lives around degree k=Θ(log⁡n)k=\Theta(\log n), which is why its total influence I(f)=∑kkWk(f)=Θ(log⁡n)I(f)=\sum_k k W_k(f)=\Theta(\log n) grows logarithmically with nn.

2.4 Poincaré equality and the smallest edge boundary

The attack argument used Poincaré as its first bound on available influence. Here we record its spectral interpretation and equality case.

Theorem 8 (Poincaré inequality on the cube) Every real-valued ff satisfies Var⁡(f)≤∑i=1n∥Dif∥22. \operatorname{Var}(f)\le\sum_{i=1}^n\|D_i f\|_2^2. For Boolean-valued ff, this says Var⁡(f)≤I(f)\operatorname{Var}(f)\le I(f). In particular, a balanced Boolean function has I(f)≥1I(f)\ge1.

Proof. Every nonempty set has size at least one. Therefore Var⁡(f)=∑S≠∅f^(S)2≤∑S∣S∣ f^(S)2=∑i∥Dif∥22. \operatorname{Var}(f) =\sum_{S\ne\varnothing}\widehat f(S)^2 \le\sum_S|S|\,\widehat f(S)^2 =\sum_i\|D_i f\|_2^2. If ff is balanced and Boolean, Var⁡(f)=Ef2=1\operatorname{Var}(f)=\mathbb E f^2=1.

For the rate-one-per-coordinate generator, −⟨f,Lf⟩=∑i∥Dif∥22-\langle f,\mathcal Lf\rangle=\sum_i\|D_i f\|_2^2 and the spectral gap is one. For the discrete random-coordinate kernel, the energy is smaller by nn and the gap is 1/n1/n. Both versions give the same displayed inequality.

Theorem 9 (Equality forces a dictator) If f:{−1,1}n→{−1,1}f:\{-1,1\}^n\to\{-1,1\} is balanced, then I(f)=1I(f)=1 if and only if f=±xif=\pm x_i for some ii.

Proof. Equality in Poincaré removes every Fourier level of degree at least two. Balance removes level zero, so W1(f)=1W_1(f)=1. The level-one rigidity proved in Lemma 1 gives f=±xif=\pm x_i. Conversely, a signed dictator has exactly one coordinate of influence one.

In geometric terms, let A={x:f(x)=1}A=\{x:f(x)=1\}. The number of undirected cube edges crossing from AA to its complement is ∣∂A∣=2n−1I(f). |\partial A|=2^{n-1}I(f). Each crossing edge is counted twice by the directed pivotal events. Among sets of size 2n−12^{n-1}, the smallest possible edge boundary is 2n−12^{n-1}, attained exactly by the coordinate half-cubes.

Minimal total influence forces dependence on one bit. The coalition problem requires a stronger statement about a large individual influence on every intermediate success function, which is why the classroom argument proceeds to hypercontractivity and KKL.

2.5 Conditional variance and resampling

The spectral inequality above is also the variance tensorization discussed in last year’s notes. The term “tensorization” refers to the ability to decompose a high-dimensional value into a sum of one-dimensional values. Here the tensorized quantity is the upper bound on variance. For fixed values of all coordinates except ii, write u=f(x−i,1),v=f(x−i,−1). u=f(x_{-i},1),\qquad v=f(x_{-i},-1). The conditional mean is (u+v)/2(u+v)/2, and the conditional variance is (u−v)2/4(u-v)^2/4. Both values of DifD_i f have squared magnitude (u−v)2/4(u-v)^2/4. Thus ∥Dif∥22=E[Var⁡(f(X)∣X−i)]. \|D_i f\|_2^2 =\mathbb E[\operatorname{Var}(f(X)\mid X_{-i})]. Poincaré becomes Var⁡(f(X))≤∑i=1nE[Var⁡(f(X)∣X−i)]. \operatorname{Var}(f(X)) \le\sum_{i=1}^n\mathbb E[\operatorname{Var}(f(X)\mid X_{-i})].

There is an equivalent resampling form. Let X(i)X^{(i)} be obtained by replacing XiX_i by an independent uniform sign, with all other coordinates unchanged. Conditional on X−iX_{-i}, the two function values are independent and identically distributed. Therefore E[(f(X)−f(X(i)))2]=2E[Var⁡(f(X)∣X−i)]. \mathbb E[(f(X)-f(X^{(i)}))^2] =2\mathbb E[\operatorname{Var}(f(X)\mid X_{-i})]. We obtain the Efron–Stein form on the cube: Var⁡(f(X))≤12∑iE[(f(X)−f(X(i)))2]. \operatorname{Var}(f(X)) \le\frac12\sum_i\mathbb E[(f(X)-f(X^{(i)}))^2]. For Boolean-valued ff, a resampling changes the output with probability Inf⁡i(f)/2\operatorname{Inf}_i(f)/2. Flipping and refreshing a bit must therefore be distinguished even though both yield the same energy identity after the appropriate factors are included.

2.6 Noise stability and the reversible-chain clock

Let XX be uniform and YY its ρ\rho-correlated noisy version. Both marginals are uniform. We use 0≤ρ≤10\le\rho\le1 for the random-perturbation comparisons below.

Recall the noise stability Stab⁡ρ(f)=E[f(X)f(Y)]\operatorname{Stab}_\rho(f)=\mathbb E[f(X)f(Y)] defined in the Arrow argument.

By orthogonality, Stab⁡ρ(f)=∑Sρ∣S∣f^(S)2. \operatorname{Stab}_\rho(f)=\sum_S\rho^{|S|}\widehat f(S)^2. For a Boolean-valued function, f(X)f(Y)f(X)f(Y) is 11 when the two outputs agree and −1-1 otherwise, so P(f(X)≠f(Y))=1−Stab⁡ρ(f)2. \mathbb P(f(X)\ne f(Y)) =\frac{1-\operatorname{Stab}_\rho(f)}2. Thus the spectral formula gives an exact measure of robustness under a specified random perturbation.

Example 9 (Dictator, parity, and three-bit majority) Their stabilities are Stab⁡ρ(x1)=ρ,Stab⁡ρ(χ[n])=ρn,Stab⁡ρ(Maj⁡3)=34ρ+14ρ3. \begin{aligned} \operatorname{Stab}_\rho(x_1)&=\rho,\\ \operatorname{Stab}_\rho(\chi_{[n]})&=\rho^n,\\ \operatorname{Stab}_\rho(\operatorname{Maj}_3)&=\frac34\rho+\frac14\rho^3. \end{aligned} For fixed ρ<1\rho<1, parity loses its correlation as nn grows. For a dictator, the output remains correlated with the original value exactly to the extent that its one relevant input does.

Noise stability and influence are connected, but they are not the same statistic: ddρStab⁡ρ(f)∣ρ=1=I(f). \left.\frac{\mathrm d}{\mathrm d\rho}\operatorname{Stab}_\rho(f)\right|_{\rho=1} =I(f). Influence controls the initial response to small noise; stability at a fixed ρ\rho depends on the whole distribution of Fourier degrees.

Connecting the clocks

Let PP be the discrete chain that refreshes one uniformly chosen coordinate. A character χS\chi_S survives an update if the selected coordinate is outside SS, and otherwise its conditional mean becomes zero. Hence PχS=(1−∣S∣n)χS. P\chi_S=\left(1-\frac{|S|}{n}\right)\chi_S. For the continuous process that refreshes each coordinate at rate one, use L=n(P−I)=−∑i=1nDi. \mathcal L=n(P-I)=-\sum_{i=1}^n D_i. Then LχS=−∣S∣χS\mathcal L\chi_S=-|S|\chi_S, and etL=Te−t. e^{t\mathcal L}=T_{e^{-t}}. This clock is nn times faster than the total-rate-one Poissonization et(P−I)e^{t(P-I)} in Chapter 2. The latter is Te−t/nT_{e^{-t/n}}. Stating the clock avoids losing a factor of nn in the spectral gap or Dirichlet form.

2.7 Spectral tails and small-noise bounds

For a Boolean function, define its degree-dd truncation by f≤d=∑∣S∣≤df^(S)χS. f^{\le d}=\sum_{|S|\le d}\widehat f(S)\chi_S. Since every discarded level has degree at least d+1d+1, ∥f−f≤d∥22=∑∣S∣>df^(S)2≤I(f)d+1. \|f-f^{\le d}\|_2^2 =\sum_{|S|>d}\widehat f(S)^2 \le\frac{I(f)}{d+1}. Low total influence therefore implies approximation by a low-degree real polynomial. If we round that polynomial to a sign, using any fixed tie rule at zero, every misclassified point contributes at least one to (f−f≤d)2(f-f^{\le d})^2. Thus the same upper bound controls the classification error. Low degree alone does not imply dependence on few coordinates; many singleton terms may still appear.

For random noise, the exact stability formula gives P(f(X)≠f(Y))=12∑S(1−ρ∣S∣)f^(S)2≤1−ρ2I(f), \mathbb P(f(X)\ne f(Y)) =\frac12\sum_S(1-\rho^{|S|})\widehat f(S)^2 \le\frac{1-\rho}{2}I(f), where 1−ρk≤k(1−ρ)1-\rho^k\le k(1-\rho). If δ=(1−ρ)/2\delta=(1-\rho)/2 is the bit-flip probability, the bound is δI(f)\delta I(f).

The first derivative at zero noise is determined by influence: lim⁡δ↓0P(f(X)≠f(Y))δ=I(f). \lim_{\delta\downarrow0} \frac{\mathbb P(f(X)\ne f(Y))}{\delta}=I(f). This is a finite-sum calculation, so no interchange-of-limit issue arises. For a sequence of functions on larger cubes, the small-noise approximation need not be uniform in nn.

2.8 The degree cutoff and its constants

Proof: choosing the degree cutoff

Put L=log⁡(1/a)>0L=\log(1/a)>0 and take d=⌊L/(4log⁡3)⌋d=\lfloor L/(4\log3)\rfloor. Then 3da≤e−L/4,1d+1≤4log⁡3L. 3^d\sqrt a\le e^{-L/4}, \qquad \frac1{d+1}\le\frac{4\log3}{L}. Since sup⁡L>0Le−L/4=4/e\sup_{L>0}Le^{-L/4}=4/e, e−L/4+4log⁡3L≤4/e+4log⁡3L<6L. e^{-L/4}+\frac{4\log3}{L} \le\frac{4/e+4\log3}{L}<\frac6L. Substitute into the low/high-degree estimate and rearrange. The argument includes d=0d=0; there is no need to assume aa is very small.

2.9 Tribes parameters and exact finite examples

In §1.10 we defined the Tribes function on n=mwn=mw coordinates (mm disjoint tribes of width ww) and found the exact formulas P(f(X)=−1)=(1−2−w)m,Inf⁡i(f)=2−(w−1)(1−2−w)m−1. \mathbb P(f(X)=-1)=(1-2^{-w})^m, \qquad \operatorname{Inf}_i(f)=2^{-(w-1)}(1-2^{-w})^{m-1}. Here we record the exact asymptotic calculation showing that Inf⁡i(f)=Θ((log⁡n)/n)\operatorname{Inf}_i(f)=\Theta((\log n)/n) along a sequence of nearly balanced Tribes functions, together with a concrete finite example.

For each integer width w≥1w\ge 1, choose the number of tribes to be mw=⌊(log⁡2) 2w⌋m_w=\lfloor(\log 2)\,2^w\rfloor, and set nw=mwwn_w=m_w w. Because log⁡(1−2−w)=−2−w+O(4−w)\log(1-2^{-w})=-2^{-w}+O(4^{-w}) and mw=(log⁡2) 2w+O(1)m_w=(\log 2)\,2^w+O(1), we have mwlog⁡(1−2−w)=−log⁡2+O(2−w),(1−2−w)mw=12+O(2−w). m_w\log(1-2^{-w})=-\log 2 + O(2^{-w}), \qquad (1-2^{-w})^{m_w}=\frac12+O(2^{-w}). Consequently E[f(X)]=O(2−w)→0\mathbb E[f(X)]=O(2^{-w})\to 0 and Var⁡(f)=1−(Ef)2=1−O(4−w)→1\operatorname{Var}(f)=1-(\mathbb E f)^2=1-O(4^{-w})\to 1 as w→∞w\to\infty. (Because mwm_w and ww must be integers, (1−2−w)mw(1-2^{-w})^{m_w} cannot equal 1/21/2 exactly, which is why Theorem 5’s dependence on Var⁡(f)\operatorname{Var}(f) rather than exact balance Ef=0\mathbb E f=0 is so convenient.)

Now express the individual influence in terms of n=nwn=n_w. Since (1−2−w)mw−1→1/2(1-2^{-w})^{m_w-1}\to 1/2, we have Inf⁡i(f)=2−(w−1)(1−2−w)mw−1∼2−w. \operatorname{Inf}_i(f)=2^{-(w-1)}(1-2^{-w})^{m_w-1}\sim 2^{-w}. Meanwhile, n=mww∼(log⁡2) w2wn=m_w w\sim(\log 2)\,w 2^w. Taking base-22 logarithms gives log⁡2n=w+log⁡2w+log⁡2log⁡2+o(1)⟹w=log⁡2n−log⁡2log⁡2n−log⁡2log⁡2+o(1), \log_2 n = w + \log_2 w + \log_2\log 2 + o(1) \quad\Longrightarrow\quad w = \log_2 n - \log_2\log_2 n - \log_2\log 2 + o(1), and substituting 2−w∼(log⁡2)wn∼log⁡nn2^{-w}\sim\frac{(\log 2)w}{n}\sim\frac{\log n}{n} yields the sharp asymptotics Inf⁡i(f)∼log⁡nn,I(f)=nInf⁡i(f)∼log⁡n. \operatorname{Inf}_i(f)\sim\frac{\log n}{n}, \qquad I(f)=n\operatorname{Inf}_i(f)\sim\log n.

Example 10 (A concrete 8-coordinate example) Take m=4m=4 tribes of width w=2w=2, so n=8n=8. Each tribe of size 22 is unanimously +1+1 with probability 1/41/4, so P(f(X)=+1)=1−(34)4=175256≈0.684, \mathbb P(f(X)=+1)=1-\left(\frac34\right)^4=\frac{175}{256}\approx 0.684, and every coordinate has influence Inf⁡i(f)=2−1(34)3=27128≈0.211. \operatorname{Inf}_i(f)=2^{-1}\left(\frac34\right)^3=\frac{27}{128}\approx 0.211. This small example illustrates that the exact probability and influence formulas hold for every (m,w)(m,w), while approximate balance only emerges once mm is chosen near (log⁡2)2w(\log 2)2^w.

Finally, notice the striking asymmetry between forcing +1+1 and forcing −1-1 on Tribes:

2.10 One coalition that can choose either outcome

For a coalition SS, define its probability of full control by q(S)=PX[n]∖S ⁣(f(X[n]∖S,⋅) takes both signs). q(S)=\mathbb P_{X_{[n]\setminus S}}\!\left( f(X_{[n]\setminus S},\cdot)\text{ takes both signs}\right). This is not the sum of the members’ individual influences. It asks whether, after seeing one honest assignment, the coalition can choose either outcome.

Let S+S_+ and S−S_- be the coalitions obtained by applying the classroom theorem to ff and −f-f, with error ε/2\varepsilon/2. Put S=S+∪S−S=S_+\cup S_-. Adding controlled coordinates cannot reduce the probability of achieving a given target: every honest assignment allowing success before the addition still supplies a witness after the addition. Formally, p+(S)≥p+(S+)p_+(S)\ge p_+(S_+) and p−(S)≥p−(S−)p_-(S)\ge p_-(S_-), with each probability evaluated on its corresponding remaining uniform cube.

On the remaining cube for SS, let E+E_+ and E−E_- be the events that the two target signs are achievable. Their probabilities are each at least 1−ε/21-\varepsilon/2. A union bound gives q(S)=P(E+∩E−)≥1−ε,∣S∣=O ⁣(nlog⁡nlog⁡1ε). q(S)=\mathbb P(E_+\cap E_-)\ge1-\varepsilon, \qquad |S|=O\!\left(\frac n{\log n}\log\frac1\varepsilon\right). Thus a single fixed coalition can choose either output on most honest assignments. The input witnesses for the two outputs generally differ. For a monotone rule, the witnesses are simply all +1+1 for the positive target and all −1-1 for the negative target; for a general rule they may depend on the honest assignment.

2.11 General hypercontractivity and polynomial moments

One useful consequence is the degree form of Bonami. If gg has degree at most dd, define hh by h^(S)=3∣S∣/2g^(S)\widehat h(S)=3^{|S|/2}\widehat g(S). Then T1/3h=gT_{1/\sqrt3}h=g, so ∥g∥4≤∥h∥2≤3d/2∥g∥2. \|g\|_4\le\|h\|_2\le3^{d/2}\|g\|_2. Low-degree polynomials cannot concentrate their entire L2L^2 mass on arbitrarily rare, arbitrarily large spikes.

The Hölder step used in class

On our finite probability space, the needed inequality is ∣E[hk]∣≤∥h∥4/3∥k∥4|\mathbb E[hk]|\le\|h\|_{4/3}\|k\|_4.

Proof of the finite-space Hölder inequality

If either norm is zero, the inequality is immediate. Otherwise set a=∣h∣/∥h∥4/3a=|h|/\|h\|_{4/3} and b=∣k∣/∥k∥4b=|k|/\|k\|_4, so Ea4/3=Eb4=1\mathbb E a^{4/3}=\mathbb E b^4=1. For nonnegative a,ba,b, Young’s inequality in the required exponents is ab≤34a4/3+14b4. ab\le\frac34a^{4/3}+\frac14b^4. To verify it directly, put t=a1/3t=a^{1/3}. The difference between right and left sides factors as b4+3t4−4t3b4=(b−t)2(b2+2bt+3t2)4≥0. \frac{b^4+3t^4-4t^3b}{4} =\frac{(b-t)^2(b^2+2bt+3t^2)}4\ge0. Averaging gives E[ab]≤1\mathbb E[ab]\le1, and undoing the normalization gives the stated Hölder bound.

The classroom proof established exactly the exponents needed for KKL. The general theorem has the same shape:

Theorem 10 (Hypercontractivity on the uniform cube — cited theorem) For 1<p≤q<∞1<p\le q<\infty and 0≤ρ≤p−1q−1, 0\le\rho\le\sqrt{\frac{p-1}{q-1}}, every real-valued function satisfies ∥Tρg∥q≤∥g∥p. \|T_\rho g\|_q\le\|g\|_p.

We quote this full statement; the special p=2,q=4p=2,q=4 case was proved in Part I. A proof of the general two-point inequality and its tensorization is in O’Donnell’s Analysis of Boolean Functions, Chapters 9–10. We prove the following consequences from the cited theorem.

First, if gg has degree at most dd and q≥2q\ge2, rescale its Fourier coefficients as in the degree-form Bonami argument to get ∥g∥q≤(q−1)d/2∥g∥2. \|g\|_q\le(q-1)^{d/2}\|g\|_2. This converts a degree bound into moment bounds. For d≥1d\ge1, normalize ∥g∥2=1\|g\|_2=1, take t≥2d/2t\ge2^{d/2}, and choose the real exponent q=1+t2/d/2≥2q=1+t^{2/d}/2\ge2. Markov’s inequality gives P(∣g∣≥t)≤((q−1)d/2t)q=2−dq/2≤exp⁡ ⁣(−dlog⁡24t2/d). \mathbb P(|g|\ge t) \le\left(\frac{(q-1)^{d/2}}{t}\right)^q =2^{-dq/2} \le \exp\!\left(-\frac{d\log2}{4}t^{2/d}\right). For degree one this is Gaussian-type decay; higher-degree polynomials have a correspondingly weaker tail scale.

Second, consider a small set AA of uniform mass μ\mu. Let X,YX,Y be ρ\rho-correlated. Self-adjointness and the semigroup identity give P(X∈A,Y∈A)=⟨1A,Tρ1A⟩=∥Tρ1A∥22. \mathbb P(X\in A,Y\in A) =\langle\mathbf1_A,T_\rho\mathbf1_A\rangle =\|T_{\sqrt\rho}\mathbf1_A\|_2^2. For 0<ρ<10<\rho<1, use hypercontractivity with input exponent 1+ρ1+\rho and output exponent 22: P(X∈A,Y∈A)≤μ2/(1+ρ). \mathbb P(X\in A,Y\in A)\le\mu^{2/(1+\rho)}. Thus P(Y∈A∣X∈A)≤μ(1−ρ)/(1+ρ)\mathbb P(Y\in A\mid X\in A)\le\mu^{(1-\rho)/(1+\rho)} when μ>0\mu>0. A rare set cannot retain most of its conditional mass under a fixed positive amount of noise. The endpoint cases follow directly or by continuity.

2.12 A semigroup route to the influence trade-off

There is another way to combine smoothing across different scales. Let Pt=Te−tP_t=T_{e^{-t}} use the rate-one-per-coordinate clock. From Chapter 2, ddtVar⁡(Ptf)=−2∑i∥DiPtf∥22. \frac{\mathrm d}{\mathrm dt}\operatorname{Var}(P_t f) =-2\sum_i\|D_i P_t f\|_2^2. Because Ptf→EfP_t f\to\mathbb E f, integration gives the exact identity Var⁡(f)=2∫0∞∑i∥DiPtf∥22 dt. \operatorname{Var}(f)=2\int_0^\infty\sum_i\|D_iP_t f\|_2^2\,\mathrm dt. To expose the decay associated with coordinate ii, write hi(x−i)=xiDif(x)=∑S∋if^(S)χS∖{i}(x). h_i(x_{-i})=x_iD_i f(x)=\sum_{S\ni i}\widehat f(S)\chi_{S\setminus\{i\}}(x). This function does not depend on xix_i. We obtain DiTρf=ρxiTρhi, D_iT_\rho f=\rho x_i T_\rho h_i, where the operator on the right acts on the other n−1n-1 coordinates. Changing variables ρ=e−t\rho=e^{-t} yields Var⁡(f)=∑i∫012ρ ∥Tρhi∥22 dρ. \operatorname{Var}(f)=\sum_i\int_0^1 2\rho\,\|T_\rho h_i\|_2^2\,\mathrm d\rho. The factor ρ\rho must not be lost when switching derivative conventions.

The general hypercontractive theorem gives ∥Tρhi∥2≤∥hi∥1+ρ2\|T_\rho h_i\|_2\le\|h_i\|_{1+\rho^2}. Since ∣hi∣|h_i| is an indicator of mass aia_i, Var⁡(f)≤∑i∫01ai2/(1+u) du,u=ρ2. \operatorname{Var}(f) \le\sum_i\int_0^1 a_i^{2/(1+u)}\,\mathrm du, \qquad u=\rho^2. If 0<ai<10<a_i<1, let Li=log⁡(1/ai)L_i=\log(1/a_i). Because 2/(1+u)≥1+(1−u)/22/(1+u)\ge1+(1-u)/2 on [0,1][0,1], ∫01ai2/(1+u) du≤ai∫01e−Li(1−u)/2 du≤aimin⁡{1,2/Li}. \int_0^1 a_i^{2/(1+u)}\,\mathrm du \le a_i\int_0^1e^{-L_i(1-u)/2}\,\mathrm du \le a_i\min\{1,2/L_i\}. The elementary comparison min⁡{1,2/L}≤3/(1+L)\min\{1,2/L\}\le3/(1+L) now gives Var⁡(f)≤3∑iai1+log⁡(1/ai). \operatorname{Var}(f)\le3\sum_i\frac{a_i}{1+\log(1/a_i)}. The summand is defined as zero at ai=0a_i=0; at ai=1a_i=1 the preceding integral is one and the displayed bound still holds. This refined estimate uses the individual sizes of all influences. Bounding them by their maximum recovers the logarithmic improvement underlying KKL.

2.13 From hypercontractivity to the logarithmic Sobolev inequality and small-set expansion

Why does the hypercontractive inequality ∥Tρg∥q≤∥g∥p\|T_\rho g\|_q\le\|g\|_p have the same structural consequences as a strengthened Poincaré inequality, and what is its geometric meaning on the hypercube? In 1975, Leonard Gross discovered that hypercontractivity along the continuous-time noise semigroup Pt=Te−tP_t=T_{e^{-t}} is equivalent to a single differential inequality at t=0t=0—the logarithmic Sobolev inequality—which replaces the variance Var⁡(f)\operatorname{Var}(f) in Poincaré’s inequality by an entropy functional Ent⁡(f2)\operatorname{Ent}(f^2) that is far more sensitive to rare spikes and small sets.

Definition 7 (Entropy functional on the cube) For any nonnegative function g:{−1,1}n→[0,∞)g:\{-1,1\}^n\to[0,\infty) (with the convention 0log⁡0=00\log 0=0 and 0log⁡(0/0)=00\log(0/0)=0), the entropy of gg with respect to the uniform measure on {−1,1}n\{-1,1\}^n is Ent⁡(g):=E ⁣[g(X)log⁡g(X)E[g(X)]]=E[g(X)log⁡g(X)]−E[g(X)]log⁡E[g(X)]. \operatorname{Ent}(g) :=\mathbb E\!\left[g(X)\log\frac{g(X)}{\mathbb E[g(X)]}\right] =\mathbb E[g(X)\log g(X)]-\mathbb E[g(X)]\log\mathbb E[g(X)].

By Jensen’s inequality applied to the strictly convex function ϕ(u)=ulog⁡u\phi(u)=u\log u on [0,∞)[0,\infty), we always have Ent⁡(g)≥0\operatorname{Ent}(g)\ge 0, with equality if and only if gg is constant. Moreover, when E[g(X)]=1\mathbb E[g(X)]=1, gg is the likelihood ratio (density) of a probability distribution ν(x)=2−ng(x)\nu(x)=2^{-n}g(x) relative to the uniform distribution μ(x)=2−n\mu(x)=2^{-n}, and Ent⁡(g)=DKL(ν ∥ μ)\operatorname{Ent}(g)=D_{\mathrm{KL}}(\nu\,\|\,\mu) is simply the Kullback–Leibler divergence from μ\mu to ν\nu (a central object in Chapters 4 and 5). For general g≥0g\ge 0 with E[g]>0\mathbb E[g]>0, homogeneity Ent⁡(cg)=cEnt⁡(g)\operatorname{Ent}(cg)=c\operatorname{Ent}(g) means that Ent⁡(g)=E[g] DKL(νg ∥ μ)\operatorname{Ent}(g)=\mathbb E[g]\,D_{\mathrm{KL}}(\nu_g\,\|\,\mu) measures how nonuniformly the mass of gg is concentrated across the cube.

Gross’s differentiation bridge: hypercontractivity at t=0t=0

Recall from §2.6 and §2.12 that Pt=Te−t=etLP_t=T_{e^{-t}}=e^{t\mathcal L} is the continuous-time Markov semigroup with generator L=−∑i=1nDi\mathcal L=-\sum_{i=1}^n D_i. In the general hypercontractive theorem (Theorem 10), fix the input exponent p=2p=2 and set ρ=e−t=1q(t)−1\rho=e^{-t}=\sqrt{\frac{1}{q(t)-1}}, so that the output exponent grows with time t≥0t\ge 0 according to q(t)=1+e2t(q(0)=2,q′(0)=2). q(t)=1+e^{2t} \qquad\bigl(q(0)=2,\quad q'(0)=2\bigr). Then hypercontractivity asserts that for every real-valued function ff, the log-norm curve Λ(t):=log⁡∥Ptf∥q(t)=1q(t)log⁡E ⁣[∣Ptf(X)∣q(t)] \Lambda(t):=\log\|P_t f\|_{q(t)}=\frac{1}{q(t)}\log\mathbb E\!\left[|P_t f(X)|^{q(t)}\right] satisfies Λ(t)≤Λ(0)=log⁡∥f∥2\Lambda(t)\le\Lambda(0)=\log\|f\|_2 for all t≥0t\ge 0. Because Λ(t)\Lambda(t) attains its maximum at t=0t=0, its right derivative at t=0t=0 must satisfy Λ′(0)≤0\Lambda'(0)\le 0!

Let us compute Λ′(0)\Lambda'(0) for a strictly positive function f>0f>0 (the general case f:{−1,1}n→Rf:\{-1,1\}^n\to\mathbb R follows by applying the result to f2+ε\sqrt{f^2+\varepsilon} and letting ε↓0\varepsilon\downarrow 0). Write ut=Ptf>0u_t=P_t f>0 and M(t)=E[utq(t)]M(t)=\mathbb E[u_t^{q(t)}]. Differentiating Λ(t)=1q(t)log⁡M(t)\Lambda(t)=\frac{1}{q(t)}\log M(t) by the product and chain rules gives Λ′(t)=−q′(t)q(t)2log⁡M(t)+1q(t)M(t)E ⁣[utq(t)(q′(t)log⁡ut+q(t)∂tutut)]. \Lambda'(t) =-\frac{q'(t)}{q(t)^2}\log M(t) +\frac{1}{q(t)M(t)}\mathbb E\!\left[u_t^{q(t)}\left(q'(t)\log u_t+q(t)\frac{\partial_t u_t}{u_t}\right)\right]. At t=0t=0, we have q(0)=2q(0)=2, q′(0)=2q'(0)=2, u0=fu_0=f, M(0)=E[f2]M(0)=\mathbb E[f^2], and ∂tut∣t=0=Lf=−∑i=1nDif\partial_t u_t|_{t=0}=\mathcal L f=-\sum_{i=1}^n D_i f. Using 2log⁡f=log⁡(f2)2\log f=\log(f^2) and the Dirichlet-form identity −⟨f,Lf⟩=∑i=1n⟨f,Dif⟩=∑i=1n∥Dif∥22=E(f,f)-\langle f,\mathcal L f\rangle=\sum_{i=1}^n\langle f,D_i f\rangle=\sum_{i=1}^n\|D_i f\|_2^2=\mathcal E(f,f), we obtain Λ′(0)=12E[f2](E[f2log⁡(f2)]−E[f2]log⁡E[f2]+2⟨f,Lf⟩)=12E[f2](Ent⁡(f2)−2∑i=1n∥Dif∥22). \begin{aligned} \Lambda'(0) &=\frac{1}{2\mathbb E[f^2]}\Bigl(\mathbb E[f^2\log(f^2)]-\mathbb E[f^2]\log\mathbb E[f^2]+2\langle f,\mathcal L f\rangle\Bigr)\\ &=\frac{1}{2\mathbb E[f^2]}\Bigl(\operatorname{Ent}(f^2)-2\sum_{i=1}^n\|D_i f\|_2^2\Bigr). \end{aligned} Therefore, the infinitesimal condition Λ′(0)≤0\Lambda'(0)\le 0 is identical to the inequality Ent⁡(f2)≤2∑i=1n∥Dif∥22\operatorname{Ent}(f^2)\le 2\sum_{i=1}^n\|D_i f\|_2^2! Conversely, Gross (1975) showed that applying this single inequality at every time t≥0t\ge 0 to the rescaled function ft=∣Ptf∣q(t)/2f_t=|P_t f|^{q(t)/2} implies Λ′(t)≤0\Lambda'(t)\le 0 for all t≥0t\ge 0, integrating back to the full (p,q)(p,q)-hypercontractive theorem (Theorem 10).

From Stam and Federbush to Gross. An information-theoretic precursor for Gaussian convolutions was proved by A. J. Stam (1959), and Paul Federbush (1969) derived a logarithmic Sobolev inequality in quantum field theory. Leonard Gross (1975) coined the name logarithmic Sobolev inequality and established its exact equivalence with hypercontractivity of Markov semigroups.

Direct proof via entropy tensorization and the one-bit inequality

Just as §§2.4–2.5 proved the Poincaré inequality Var⁡(f)≤∑i=1n∥Dif∥22\operatorname{Var}(f)\le\sum_{i=1}^n\|D_i f\|_2^2 by combining variance tensorization with a one-bit identity, we can prove the logarithmic Sobolev inequality directly on the cube by combining entropy tensorization with an elementary one-bit calculus inequality.

For each coordinate i∈[n]i\in[n], let Ei[g](x)=g(x−i,+1)+g(x−i,−1)2\mathbb E_i[g](x)=\frac{g(x_{-i},+1)+g(x_{-i},-1)}{2} denote the expectation over XiX_i with the remaining n−1n-1 coordinates X−iX_{-i} held fixed, and define the conditional one-coordinate entropy Ent⁡i(g):=Ei ⁣[glog⁡gEi[g]]. \operatorname{Ent}_i(g) :=\mathbb E_i\!\left[g\log\frac{g}{\mathbb E_i[g]}\right].

Lemma 3 (Entropy tensorization (subadditivity) on the cube) For every nonnegative function g:{−1,1}n→[0,∞)g:\{-1,1\}^n\to[0,\infty), Ent⁡(g)≤∑i=1nE[Ent⁡i(g)]. \operatorname{Ent}(g)\le\sum_{i=1}^n\mathbb E[\operatorname{Ent}_i(g)].

Proof: telescoping via the chain rule and joint convexity

By approximation (g+εg+\varepsilon with ε↓0\varepsilon\downarrow 0), we may assume g>0g>0. For 0≤k≤n0\le k\le n, let gk=E[g∣X1,…,Xk]g_k=\mathbb E[g\mid X_1,\ldots,X_k] be the conditional expectation given the first kk coordinates, so gn=gg_n=g and g0=E[g]g_0=\mathbb E[g]. Writing log⁡gE[g]=∑i=1nlog⁡gigi−1\log\frac{g}{\mathbb E[g]}=\sum_{i=1}^n\log\frac{g_i}{g_{i-1}} and using the tower property E[glog⁡gigi−1]=E[gilog⁡gigi−1]\mathbb E[g\log\frac{g_i}{g_{i-1}}]=\mathbb E[g_i\log\frac{g_i}{g_{i-1}}] decomposes the total entropy into a sum of one-step increments: Ent⁡(g)=∑i=1nE ⁣[gilog⁡gigi−1]. \operatorname{Ent}(g)=\sum_{i=1}^n\mathbb E\!\left[g_i\log\frac{g_i}{g_{i-1}}\right]. Now fix i∈[n]i\in[n] and let E>i\mathbb E_{>i} denote averaging over (Xi+1,…,Xn)(X_{i+1},\ldots,X_n) conditioned on (X1,…,Xi)(X_1,\ldots,X_i). By independence of the coordinates, gi=E>i[g]g_i=\mathbb E_{>i}[g] and gi−1=Ei[gi]=E>i[Eig]g_{i-1}=\mathbb E_i[g_i]=\mathbb E_{>i}[\mathbb E_i g]. Because the two-variable relative-entropy function Ψ(u,v)=ulog⁡(u/v)\Psi(u,v)=u\log(u/v) is jointly convex on (0,∞)2(0,\infty)^2 (its Hessian (1/u−1/v−1/vu/v2)\bigl(\begin{smallmatrix}1/u & -1/v\\ -1/v & u/v^2\end{smallmatrix}\bigr) is positive semidefinite), Jensen’s inequality applied to the average E>i\mathbb E_{>i} gives gilog⁡gigi−1=Ψ(E>i[g], E>i[Eig])≤E>i[Ψ(g,Eig)]. g_i\log\frac{g_i}{g_{i-1}} =\Psi\bigl(\mathbb E_{>i}[g],\,\mathbb E_{>i}[\mathbb E_i g]\bigr) \le\mathbb E_{>i}\bigl[\Psi(g,\mathbb E_i g)\bigr]. Taking full expectations and using the tower property E[Ψ(g,Eig)]=E[Ei[Ψ(g,Eig)]]=E[Ent⁡i(g)]\mathbb E[\Psi(g,\mathbb E_i g)]=\mathbb E[\mathbb E_i[\Psi(g,\mathbb E_i g)]]=\mathbb E[\operatorname{Ent}_i(g)] across i=1,…,ni=1,\ldots,n yields the claim.

Theorem 11 (Logarithmic Sobolev inequality on the uniform cube (Gross)) For every real-valued function f:{−1,1}n→Rf:\{-1,1\}^n\to\mathbb R, Ent⁡(f2)≤2∑i=1n∥Dif∥22=2 E(f,f). \operatorname{Ent}(f^2)\le 2\sum_{i=1}^n\|D_i f\|_2^2=2\,\mathcal E(f,f).

Proof. By Lemma 3 applied to g=f2g=f^2, it suffices to prove the one-bit inequality Ent⁡i(f2)≤2Ei[(Dif)2]\operatorname{Ent}_i(f^2)\le 2\mathbb E_i[(D_i f)^2] for each fixed assignment of x−ix_{-i}. Write a=f(x−i,+1)a=f(x_{-i},+1) and b=f(x−i,−1)b=f(x_{-i},-1); then Ei[(Dif)2]=(a−b2)2\mathbb E_i[(D_i f)^2]=(\frac{a-b}{2})^2, so we must show that for all a,b∈Ra,b\in\mathbb R, a22log⁡2a2a2+b2+b22log⁡2b2a2+b2≤(a−b)22. \frac{a^2}{2}\log\frac{2a^2}{a^2+b^2}+\frac{b^2}{2}\log\frac{2b^2}{a^2+b^2}\le\frac{(a-b)^2}{2}. Since ∣∣a∣−∣b∣∣≤∣a−b∣||a|-|b||\le|a-b|, replacing (a,b)(a,b) with (∣a∣,∣b∣)(|a|,|b|) leaves the left-hand side unchanged and can only decrease the right-hand side, so we may assume a,b≥0a,b\ge 0. If a=b=0a=b=0 both sides vanish; otherwise, by symmetry and homogeneity of degree 22, we may assume a≥b≥0a\ge b\ge 0 and normalize a+b=2a+b=2, writing a=1+ta=1+t and b=1−tb=1-t with t∈[0,1]t\in[0,1]. Then (a−b)22=2t2\frac{(a-b)^2}{2}=2t^2 and a2+b22=1+t2\frac{a^2+b^2}{2}=1+t^2, so the left-hand side becomes φ(t):=(1+t)2log⁡(1+t)+(1−t)2log⁡(1−t)−(1+t2)log⁡(1+t2). \varphi(t) :=(1+t)^2\log(1+t)+(1-t)^2\log(1-t)-(1+t^2)\log(1+t^2). For t∈(0,1)t\in(0,1), differentiating twice with respect to tt gives φ′(t)=2(1+t)log⁡(1+t)−2(1−t)log⁡(1−t)−2tlog⁡(1+t2), \varphi'(t)=2(1+t)\log(1+t)-2(1-t)\log(1-t)-2t\log(1+t^2), φ′′(t)=2log⁡1−t21+t2+41+t2≤4, \varphi''(t)=2\log\frac{1-t^2}{1+t^2}+\frac{4}{1+t^2}\le 4, since log⁡1−t21+t2≤0\log\frac{1-t^2}{1+t^2}\le 0 and 41+t2≤4\frac{4}{1+t^2}\le 4. Because φ(0)=φ′(0)=0\varphi(0)=\varphi'(0)=0, integrating φ′′(t)≤4\varphi''(t)\le 4 twice on [0,t][0,t] (and extending to t=1t=1 by continuity) yields φ(t)≤2t2\varphi(t)\le 2t^2 for all t∈[0,1]t\in[0,1]. Averaging over x−ix_{-i} and summing over i=1,…,ni=1,\ldots,n completes the proof.

Notice that in the one-bit proof above, φ′′(0)=4=(2t2)′′∣t=0\varphi''(0)=4=(2t^2)''|_{t=0}: for small perturbations f=1+εhf=1+\varepsilon h around a constant function (with E[h]=0\mathbb E[h]=0), expanding Ent⁡((1+εh)2)=2ε2Var⁡(h)+O(ε3)\operatorname{Ent}((1+\varepsilon h)^2)=2\varepsilon^2\operatorname{Var}(h)+O(\varepsilon^3) and 2E(1+εh,1+εh)=2ε2E(h,h)2\mathcal E(1+\varepsilon h,1+\varepsilon h)=2\varepsilon^2\mathcal E(h,h) shows that the Poincaré inequality Var⁡(h)≤E(h,h)\operatorname{Var}(h)\le\mathcal E(h,h) is simply the second-order linearization of the logarithmic Sobolev inequality around constant functions!

Feature Poincaré Inequality (§§2.4–2.5) Logarithmic Sobolev Inequality (§2.13)
Global inequality Var⁡(f)≤∑i=1n∥Dif∥22=E(f,f)\operatorname{Var}(f)\le\sum_{i=1}^n\|D_i f\|_2^2=\mathcal E(f,f) Ent⁡(f2)≤2∑i=1n∥Dif∥22=2 E(f,f)\operatorname{Ent}(f^2)\le 2\sum_{i=1}^n\|D_i f\|_2^2=2\,\mathcal E(f,f)
Semigroup form (Pt=Te−tP_t=T_{e^{-t}}) ∥Ptf−Ef∥2≤e−t∥f−Ef∥2\|P_t f-\mathbb E f\|_2\le e^{-t}\|f-\mathbb E f\|_2 (L2→L2L^2\to L^2) ∥Ptf∥1+e2t≤∥f∥2\|P_t f\|_{1+e^{2t}}\le\|f\|_2 (L2→Lq(t)L^2\to L^{q(t)} hypercontractivity)
Product-space decomposition Variance tensorization: Var⁡(f)≤∑iE[Var⁡i(f)]\operatorname{Var}(f)\le\sum_i\mathbb E[\operatorname{Var}_i(f)] Entropy tensorization: Ent⁡(f2)≤∑iE[Ent⁡i(f2)]\operatorname{Ent}(f^2)\le\sum_i\mathbb E[\operatorname{Ent}_i(f^2)]
One-coordinate step (n=1n=1) Exact identity Var⁡1(f)=∥D1f∥22=(a−b)24\operatorname{Var}_1(f)=\|D_1 f\|_2^2=\frac{(a-b)^2}{4} Tight two-point bound Ent⁡1(f2)≤2∥D1f∥22=(a−b)22\operatorname{Ent}_1(f^2)\le 2\|D_1 f\|_2^2=\frac{(a-b)^2}{2}
Small set AA of volume α=2−k\alpha=2^{-k} Var⁡(1A)=α(1−α)≈α\operatorname{Var}(\mathbf 1_A)=\alpha(1-\alpha)\approx\alpha   ⟹  \implies out-degree ≥2(1−α)\ge 2(1-\alpha) Ent⁡(1A2)=αlog⁡1α\operatorname{Ent}(\mathbf 1_A^2)=\alpha\log\frac1\alpha   ⟹  \implies out-degree ≥log⁡1α=(ln⁡2)k\ge\log\frac1\alpha=(\ln 2)k
Extremal sets on {−1,1}n\{-1,1\}^n Coordinate half-cubes (α=1/2\alpha=1/2, k=1k=1, exact equality) Small subcubes of codimension kk (α=2−k\alpha=2^{-k}, out-degree kk vs. bound (ln⁡2)k(\ln 2)k)

Geometric intuition: small-set expansion and the log⁡(1/a)\log(1/a) factor in KKL

The table above highlights the exact geometric difference between Poincaré and Log-Sobolev when we test them on the 0/10/1 indicator f=1Af=\mathbf 1_A of a small set A⊆{−1,1}nA\subseteq\{-1,1\}^n of volume α=P(X∈A)=2−k\alpha=\mathbb P(X\in A)=2^{-k}:

  1. Boundary out-degree of small sets: Along coordinate ii, the derivative Di1A(x)=1A(x)−1A(x⊕i)2D_i\mathbf 1_A(x)=\frac{\mathbf 1_A(x)-\mathbf 1_A(x^{\oplus i})}{2} equals ±12\pm\frac12 on boundary edges crossing between AA and AcA^c, and 00 elsewhere. Hence ∥Di1A∥22=14P(1A(X)≠1A(X⊕i))\|D_i\mathbf 1_A\|_2^2=\frac14\mathbb P(\mathbf 1_A(X)\ne\mathbf 1_A(X^{\oplus i})), and summing over i=1,…,ni=1,\ldots,n relates the Dirichlet energy directly to the size ∣∂A∣|\partial A| of the undirected edge boundary between AA and AcA^c: E(1A,1A)=∑i=1n∥Di1A∥22=12⋅∣∂A∣2n=α2⋅∣∂A∣∣A∣, \mathcal E(\mathbf 1_A,\mathbf 1_A)=\sum_{i=1}^n\|D_i\mathbf 1_A\|_2^2=\frac12\cdot\frac{|\partial A|}{2^n}=\frac{\alpha}{2}\cdot\frac{|\partial A|}{|A|}, where ∣∂A∣∣A∣\frac{|\partial A|}{|A|} is the average boundary out-degree of a vertex in AA (the average number of coordinates i∈[n]i\in[n] whose flip exits AA).
    • What Poincaré sees: Because Var⁡(1A)=α(1−α)\operatorname{Var}(\mathbf 1_A)=\alpha(1-\alpha), Poincaré gives α(1−α)≤α2∣∂A∣∣A∣\alpha(1-\alpha)\le\frac{\alpha}{2}\frac{|\partial A|}{|A|}, or ∣∂A∣∣A∣≥2(1−α)\frac{|\partial A|}{|A|}\ge 2(1-\alpha). Even when α=2−k\alpha=2^{-k} is exponentially small, Poincaré only guarantees a constant O(1)O(1) average out-degree, because it is calibrated to half-cubes (α=1/2\alpha=1/2).
    • What Log-Sobolev sees: Evaluating the entropy of 1A2=1A\mathbf 1_A^2=\mathbf 1_A gives Ent⁡(1A2)=E ⁣[1Alog⁡1Aα]=αlog⁡1α. \operatorname{Ent}(\mathbf 1_A^2)=\mathbb E\!\left[\mathbf 1_A\log\frac{\mathbf 1_A}{\alpha}\right]=\alpha\log\frac1\alpha. Plugging this into Theorem 11 yields the small-set isoperimetric inequality on the cube: αlog⁡1α≤2 E(1A,1A)=α⋅∣∂A∣∣A∣⟹∣∂A∣∣A∣≥log⁡1α=(ln⁡2) k. \alpha\log\frac1\alpha\le 2\,\mathcal E(\mathbf 1_A,\mathbf 1_A)=\alpha\cdot\frac{|\partial A|}{|A|} \quad\Longrightarrow\quad \frac{|\partial A|}{|A|}\ge\log\frac1\alpha=(\ln 2)\,k. Every subset A⊆{−1,1}nA\subseteq\{-1,1\}^n of volume α=2−k\alpha=2^{-k} has average boundary out-degree at least (ln⁡2)k=Ω(log⁡(1/α))(\ln 2)k=\Omega(\log(1/\alpha))! Geometrically, the most compact set of volume 2−k2^{-k} in the hypercube is a codimension-kk subcube (fixing kk coordinates), where every vertex in AA has exactly k=log⁡2(1/α)k=\log_2(1/\alpha) outgoing edges to AcA^c.
  2. Where the log⁡(1/a)\log(1/a) factor in KKL comes from: Return to a Boolean function f:{−1,1}n→{−1,1}f:\{-1,1\}^n\to\{-1,1\} whose discrete derivatives hi=Dif∈{−1,0,1}h_i=D_i f\in\{-1,0,1\} have small influences ai=Inf⁡i(f)=P(hi≠0)≤a≪1a_i=\operatorname{Inf}_i(f)=\mathbb P(h_i\ne 0)\le a\ll 1. Because hi2=1Aih_i^2=\mathbf 1_{A_i} is the 0/10/1 indicator of the rare pivotal set AiA_i of volume aia_i, its entropy is Ent⁡((Dif)2)=ailog⁡1ai. \operatorname{Ent}\bigl((D_i f)^2\bigr)=a_i\log\frac{1}{a_i}. Applying Theorem 11 to DifD_i f (where DjDif=∑S⊇{i,j}f^(S)χSD_j D_i f=\sum_{S\supseteq\{i,j\}}\widehat f(S)\chi_S for all i,j∈[n]i,j\in[n], including j=ij=i) and summing over i=1,…,ni=1,\ldots,n gives log⁡1a⋅I(f)≤∑i=1nailog⁡1ai=∑i=1nEnt⁡((Dif)2)≤2∑i=1n∑j=1n∥DjDif∥22=2∑S⊆[n]∣S∣2f^(S)2 \log\frac1a\cdot I(f) \le\sum_{i=1}^n a_i\log\frac{1}{a_i} =\sum_{i=1}^n\operatorname{Ent}\bigl((D_i f)^2\bigr) \le 2\sum_{i=1}^n\sum_{j=1}^n\|D_j D_i f\|_2^2 =2\sum_{S\subseteq[n]}|S|^2\widehat f(S)^2 (and applying Theorem 11 to the xix_i-free derivative xiDif=∑S∋if^(S)χS∖{i}x_i D_i f=\sum_{S\ni i}\widehat f(S)\chi_{S\setminus\{i\}} drops the j=ij=i term, tightening ∣S∣2|S|^2 to ∣S∣(∣S∣−1)|S|(|S|-1)). This immediately shows that small individual influences force the quadratic degree moment ∑S∣S∣2f^(S)2\sum_S|S|^2\widehat f(S)^2 to exceed 12log⁡(1/a) I(f)\frac12\log(1/a)\,I(f)! In §2.12, integrating hypercontractivity along the semigroup Pt=Te−tP_t=T_{e^{-t}} converted this exact log⁡(1/ai)\log(1/a_i) small-set entropy gain on hih_i into the variance bound Var⁡(f)≤3∑iai1+log⁡(1/ai)\operatorname{Var}(f)\le 3\sum_i\frac{a_i}{1+\log(1/a_i)}.

2.14 Changing the bias: the Russo–Margulis formula

The preceding sections used uniform signs. For threshold questions it is convenient to use bits Xi∈{0,1}X_i\in\{0,1\} with independent Ber⁡(p)\operatorname{Ber}(p) laws. Let g:{0,1}n→{0,1}g:\{0,1\}^n\to\{0,1\} be increasing in every coordinate, and put θ(p)=Epg. \theta(p)=\mathbb E_p g. This is the probability of the increasing event {g=1}\{g=1\} as the input density varies.

Theorem 12 (Russo–Margulis formula) For 0<p<10<p<1, θ′(p)=∑i=1nInf⁡i,p(g), \theta'(p)=\sum_{i=1}^n\operatorname{Inf}_{i,p}(g), where Inf⁡i,p(g)\operatorname{Inf}_{i,p}(g) is the probability, under the other n−1n-1 biased bits, that g(X−i,1)≠g(X−i,0). g(X_{-i},1)\ne g(X_{-i},0).

Proof. Give the coordinates separate biases p1,…,pnp_1,\ldots,p_n and let F(p1,…,pn)F(p_1,\ldots,p_n) be the resulting expectation of gg. Conditional on all bits except ii, E[g∣X−i]=pig(X−i,1)+(1−pi)g(X−i,0). \mathbb E[g\mid X_{-i}] =p_i g(X_{-i},1)+(1-p_i)g(X_{-i},0). Differentiating with respect to pip_i gives ∂F∂pi=E[g(X−i,1)−g(X−i,0)]. \frac{\partial F}{\partial p_i} =\mathbb E[g(X_{-i},1)-g(X_{-i},0)]. For increasing Boolean gg, the difference is either zero or one, and equals the pivotal indicator. Apply the chain rule to θ(p)=F(p,…,p)\theta(p)=F(p,\ldots,p).

This identity turns an influence bound into a lower bound on the slope of a probability curve. To control a whole threshold window, one needs an influence bound uniformly throughout that window under the biased product measures. The uniform-cube KKL theorem alone only applies at p=1/2p=1/2; it does not by itself prove a general sharp-threshold theorem.

For example, if gg is the AND of nn bits, then θ(p)=pn\theta(p)=p^n and every biased influence is pn−1p^{n-1}. The identity gives θ′(p)=npn−1\theta'(p)=np^{n-1} directly.

III. Further Topics

The following are established results to study and explain. Their proofs are not included here.

3.1 Stability of the dictator conclusion: the FKN theorem

Background. Arrow forced exact level-one structure and hence a dictator. A quantitative question asks whether almost avoiding cyclic outcomes already forces proximity to dictatorship. The FKN theorem supplies the needed robust structural statement.

Main result. There is a universal constant CC such that if a Boolean function satisfies ∑∣S∣≥2f^(S)2≤ε, \sum_{|S|\ge2}\widehat f(S)^2\le\varepsilon, then it differs from a constant function or a signed dictator on at most CεC\varepsilon of the cube. This turns exact low-degree structure into an approximation theorem.

Arrow continuation. In the common-rule model, if the cyclic-outcome probability is ε\varepsilon, the stability formula implies W1(f)≥1−9ε/2W_1(f)\ge1-9\varepsilon/2. The degree-one form of FKN then gives O(ε)O(\varepsilon) distance from a signed dictator. This is the robust Arrow consequence in O’Donnell §2.5, Corollary 2.60. The approximation is to a signed dictator; unanimity at two extreme inputs alone should not be used to infer the sign of an approximate dictator.

References. Ehud Friedgut, Gil Kalai, and Assaf Naor, Boolean Functions Whose Fourier Transform Is Concentrated on the First Two Levels, Advances in Applied Mathematics 29 (2002), 427–437. An author-written revision of the first proof is FKN, first proof, rewritten.

3.2 Finding a hidden large Fourier coefficient

Background. The BLR proof showed that a highly consistent table has a large positive Fourier coefficient, but did not locate it. The formula f^(S)=E[fχS]\widehat f(S)=\mathbb E[f\chi_S] lets us estimate a specified coefficient by sampling. There are exponentially many possible subsets SS. How can we find the significant coefficients without enumerating them all?

Main result. Given query access to a Boolean function and a threshold τ>0\tau>0, the Goldreich–Levin algorithm can, with high probability, produce a polynomial-size list containing every SS with ∣f^(S)∣≥τ|\widehat f(S)|\ge\tau, in time polynomial in nn, 1/τ1/\tau, and the logarithm of the inverse failure probability. Parseval already bounds the number of such coefficients by 1/τ21/\tau^2; the algorithmic task is to locate them.

References. Oded Goldreich and Leonid Levin, A Hard-Core Predicate for All One-Way Functions, STOC 1989, 25–32; see also O’Donnell’s textbook treatment of learning and large Fourier coefficients. The suggested reading target is the Fourier-search algorithm, with the cryptographic application as optional context.

The results below extend the influence story. Each entry supplies background, a conclusion to understand, and a reading source; proofs are left to the references.

3.3 Friedgut’s junta theorem

Background. The spectral-tail estimate in Reading Contents shows that small total influence gives a low-degree approximation. A stronger structural claim would say that the function is approximately determined by a small set of coordinates.

Main result. If ff is Boolean and 0<ε<10<\varepsilon<1, there is a Boolean function gg depending on at most exp⁡ ⁣(O(I(f)/ε)) \exp\!\left(O(I(f)/\varepsilon)\right) coordinates such that P(f≠g)≤ε\mathbb P(f\ne g)\le\varepsilon. Such a function gg is called a junta. The number of relevant coordinates depends on the total influence and desired accuracy, rather than directly on the ambient dimension.

References. Ehud Friedgut, Boolean Functions with Low Average Sensitivity Depend on Few Coordinates, Combinatorica 18 (1998), 27–35, journal article. See also O’Donnell, Analysis of Boolean Functions, Chapter 9.

3.4 Symmetry and sharp thresholds

Background. If an increasing event treats all coordinates symmetrically, none can absorb a disproportionate share of the influence. A biased version of KKL, combined with Russo–Margulis, then forces the event probability to change rapidly as pp varies.

Main result. Fix 0<δ<1/20<\delta<1/2 and 0<ε<1/20<\varepsilon<1/2. For nonconstant increasing Boolean functions invariant under a transitive permutation group on nn coordinates, any transition from probability ε\varepsilon to 1−ε1-\varepsilon lying in p∈[δ,1−δ]p\in[\delta,1-\delta] has width Oδ ⁣(log⁡(1/ε)log⁡n). O_\delta\!\left(\frac{\log(1/\varepsilon)}{\log n}\right). This is an interior-parameter consequence of the Friedgut–Kalai sharp-threshold framework. Thresholds approaching p=0p=0 or 11 require the corresponding biased estimates with their explicit dependence on pp.

Reference. Ehud Friedgut and Gil Kalai, Every Monotone Graph Property Has a Sharp Threshold, Proceedings of the American Mathematical Society 124 (1996), 2993–3002.

3.5 Many tiny influences and noise sensitivity

Background. The coalition example showed that small individual influences do not guarantee resistance to adversarial manipulation. For independent random perturbations, a different question is whether the original and perturbed outputs become asymptotically uncorrelated. A stronger condition is that the sum of squared influences tends to zero. Tribes satisfies this condition, since its nn equal influences have order (log⁡n)/n(\log n)/n.

Main result. Let fnf_n be a sequence of Boolean functions on uniform cubes. If ∑iInf⁡i(fn)2⟶0, \sum_i\operatorname{Inf}_i(f_n)^2\longrightarrow0, then for every fixed 0≤ρ<10\le\rho<1, with X,YX,Y a ρ\rho-correlated pair, E[fn(X)fn(Y)]−(Efn)2⟶0. \mathbb E[f_n(X)f_n(Y)]-(\mathbb E f_n)^2\longrightarrow0. This is a sufficient condition for noise sensitivity. For the nearly balanced tribes sequence, it implies that the two outputs disagree with probability tending to 1/21/2 under any fixed positive noise level.

Reference. Itai Benjamini, Gil Kalai, and Oded Schramm, Noise Sensitivity of Boolean Functions and Applications to Percolation, Publications Mathématiques de l’IHÉS 90 (1999), 5–43.

3.6 Logarithmic Sobolev inequalities and Markov chain mixing

Background. In Chapter 2 (§1.13–1.14), the spectral gap γ\gamma (or Poincaré inequality Var⁡π(f)≤1γE(f,f)\operatorname{Var}_\pi(f)\le\frac{1}{\gamma}\mathcal E(f,f)) of a lazy, irreducible reversible Markov chain yielded the general L2(π)L^2(\pi) mixing-time upper bound τmix(ε)≤⌈1γ(12log⁡1πmin⁡+log⁡12ε)⌉,πmin⁡=min⁡x∈Ωπ(x). \tau_{\mathrm{mix}}(\varepsilon) \le\left\lceil\frac{1}{\gamma}\left(\frac12\log\frac{1}{\pi_{\min}}+\log\frac{1}{2\varepsilon}\right)\right\rceil, \qquad \pi_{\min}=\min_{x\in\Omega}\pi(x). On the nn-dimensional hypercube under the single-coordinate-per-step walk (γ=1/n\gamma=1/n and πmin⁡=2−n\pi_{\min}=2^{-n}), this generic spectral-gap bound gave O(n2)O(n^2), missing the true Θ(nlog⁡n)\Theta(n\log n) mixing time by a factor of Θ(n/log⁡n)\Theta(n/\log n). As we saw in §2.13, the reason for that loss is geometric: at t=0t=0, a walk started at a deterministic vertex x0x_0 has relative density h0(x)=1{x=x0}/π(x0)h_0(x)=\mathbf 1_{\{x=x_0\}}/\pi(x_0) supported on a tiny set of stationary measure π(x0)=2−n\pi(x_0)=2^{-n}, and Poincaré fails to capture the rapid small-set expansion of rare spikes.

Main result. Let PP be an irreducible reversible transition matrix on a finite state space Ω\Omega with stationary distribution π\pi, continuous-time generator L=P−I\mathcal L=P-I, and Dirichlet form E(f,f)=⟨f,(I−P)f⟩π\mathcal E(f,f)=\langle f,(I-P)f\rangle_\pi. Define the logarithmic Sobolev constant αLSI\alpha_{\mathrm{LSI}} as the largest constant α>0\alpha>0 such that every f:Ω→Rf:\Omega\to\mathbb R satisfies Ent⁡π(f2):=Eπ ⁣[f2log⁡f2Eπ[f2]]≤2α E(f,f). \operatorname{Ent}_\pi(f^2) :=\mathbb E_\pi\!\left[f^2\log\frac{f^2}{\mathbb E_\pi[f^2]}\right] \le\frac{2}{\alpha}\,\mathcal E(f,f). (Linearizing around f=1+δgf=1+\delta g gives αLSI≤γ\alpha_{\mathrm{LSI}}\le\gamma; for the single-coordinate-per-step walk on {−1,1}n\{-1,1\}^n, rescaling Theorem 11 by 1/n1/n gives the matching lower bound αLSI≥1/n=γ\alpha_{\mathrm{LSI}}\ge 1/n=\gamma.) By Gross’s equivalence, αLSI\alpha_{\mathrm{LSI}} governs the hypercontractive smoothing of etLe^{t\mathcal L} from L2(π)L^2(\pi) into Lq(t)(π)L^{q(t)}(\pi) (q(t)=1+e2αLSItq(t)=1+e^{2\alpha_{\mathrm{LSI}}t}) and replaces the initial-state factor log⁡(1/πmin⁡)\log(1/\pi_{\min}) by its double logarithm: for continuous-time mixing (and similarly for lazy discrete-time chains), τmix(ε)=O ⁣(1αLSIlog⁡log⁡1πmin⁡+1γlog⁡1ε). \tau_{\mathrm{mix}}(\varepsilon) =O\!\left(\frac{1}{\alpha_{\mathrm{LSI}}}\log\log\frac{1}{\pi_{\min}}+\frac{1}{\gamma}\log\frac{1}{\varepsilon}\right). For the random walk on {−1,1}n\{-1,1\}^n, where αLSI=γ=1/n\alpha_{\mathrm{LSI}}=\gamma=1/n and log⁡log⁡(1/πmin⁡)=log⁡(nlog⁡2)=Θ(log⁡n)\log\log(1/\pi_{\min})=\log(n\log 2)=\Theta(\log n), this bound recovers the sharp O(nlog⁡n)O(n\log n) mixing time directly from a functional inequality!

References. Persi Diaconis and Laurent Saloff-Coste, Logarithmic Sobolev Inequalities for Finite Markov Chains, Annals of Applied Probability 6 (1996), 695–750; Leonard Gross, Logarithmic Sobolev Inequalities, American Journal of Mathematics 97 (1975), 1061–1083; see also David A. Levin and Yuval Peres, Markov Chains and Mixing Times (2nd ed., AMS, 2017), Chapter 22.

References and source acknowledgments. The hypercube, variance tensorization, and Poincaré discussions draw on Chihao Zhang’s AA2025 Lecture 8, originally scribed by Xuanxuan Liu and Yuchen He. The analytic proof of Arrow’s theorem follows Kalai [6] and O’Donnell [3], §2.5; the collective coin-flipping, hypercontractivity, and KKL development follows Kahn, Kalai, and Linial [2] and O’Donnell [3,4]; the logarithmic Sobolev equivalence and entropy tensorization follow Gross [7] and O’Donnell [3], §10.1; and the semigroup clock continues Chapter 2.

  1. Manuel Blum, Michael Luby, and Ronitt Rubinfeld, Self-Testing/Correcting with Applications to Numerical Problems, Journal of Computer and System Sciences 47 (1993), 549–595; preliminary version in STOC 1990. Source for the linearity-testing and self-correction problem.
  2. Jeff Kahn, Gil Kalai, and Nathan Linial, The Influence of Variables on Boolean Functions, FOCS 1988, 68–80. Source for the influence bound and its collective coin-flipping consequences.
  3. Ryan O’Donnell, Analysis of Boolean Functions. Sections 1.2–1.4 and 2.4–2.5 supply the Fourier/noise/Arrow framework; §1.6 supplies BLR/local correction; §§9.2, 9.6, and 10.1 supply hypercontractivity, KKL, coalition applications, and logarithmic Sobolev inequalities.
  4. Ryan O’Donnell’s CMU course: linearity testing and local decoding, influences, coalitions, and tribes, and KKL and hypercontractivity. These are scribed, mostly unproofread notes. Here the local-decoding amplification requires error below 1/41/4, and the coalition proof uses direct maximization over controlled coordinates rather than a monotone-compression reduction.
  5. Yuval Filmus, Lianna Hambardzumyan, Hamed Hatami, Pooya Hatami, and David Zuckerman, Biasing Boolean Functions and Collective Coin-Flipping Protocols over Arbitrary Product Distributions, ICALP 2019, 58:1–58:13. Its introduction and definitions explain the single-round rushing model and distinguish the uniform setting proved here from more general product distributions.
  6. Gil Kalai, A Fourier-Theoretic Perspective on the Condorcet Paradox and Arrow’s Theorem, Advances in Applied Mathematics 29 (2002), 412–426. Source for the analytic correlation formula and the connection between consistency and spectral structure.
  7. Leonard Gross, Logarithmic Sobolev Inequalities, American Journal of Mathematics 97 (1975), 1061–1083. Source for the equivalence between hypercontractivity of Markov semigroups and the logarithmic Sobolev inequality.