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 n 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} to each pair of candidates—A versus B, B versus C, and C versus A—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>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)=xi), 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}, 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/logn) can still force the output with probability 0.99.
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 and tracking how its Fourier spectrum responds to coordinatewise noise Tρ. In the first meeting, negative correlation (ρ=−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) 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/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}n, Boolean outputs lie in {−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}n for linear maps over F2={0,1} with addition modulo two, related to signs by xi=(−1)ui.
I. Course Content
1.1 A mathematical model of voting and collective rankings
We want to combine the preferences of n 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≥1. The voters are labelled by [n]={1,…,n}, and the candidate set is 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 i’s preference ≻i is irreflexive (U≻iU for each candidate U), compares every two distinct candidates in exactly one direction, and is transitive: U≻iV and V≻iW⟹U≻iW. Write A>B>C for the ranking that places A first, B second, and C third. There are 3!=6 such rankings. Denote their set by L(C).
Definition 1 (Ballot and preference profile) Voter i submits a ballot πi∈L(C), consisting of a complete strict ranking of the candidates. A preference profile is the ordered tuple of all n ballots, π=(π1,…,πn)∈L(C)n. Every such profile is allowed, giving 6n 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). 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).
Turn each ballot into three binary comparisons
For a fixed profile, we record voter i’s pairwise preferences along the cyclic orientations A versus B, B versus C, and C versus A: xi={+1,−1,A≻iB,B≻iA,yi={+1,−1,B≻iC,C≻iB,zi={+1,−1,C≻iA,A≻iC. Orienting the three pairs cyclically (A→B→C→A) makes the forbidden cycles easy to spot: (xi,yi,zi)=(+1,+1,+1) means A≻iB≻iC≻iA, while (−1,−1,−1) gives the reverse cycle A≺iB≺iC≺iA. Neither can occur for a transitive voter. Checking all 3!=6 rankings shows that the remaining six sign triples all occur:
Orienting pairs along A→B→C→A turns the two directed cycles into the constant triples (+1,+1,+1) and (−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 turns Arrow’s theorem into a spectral identity.
Individual ranking
xi
yi
zi
A>B>C
+1
+1
−1
A>C>B
+1
−1
−1
B>A>C
−1
+1
−1
B>C>A
−1
+1
+1
C>A>B
+1
−1
+1
C>B>A
−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)}.
Collecting the comparisons across all n voters gives three column vectors in {−1,1}n: x=(x1,…,xn),y=(y1,…,yn),z=(z1,…,zn). Equivalently, a legal profile π∈L(C)n is an n×3 sign matrix with columns x,y,z whose ith row (xi,yi,zi) belongs to R for every voter i∈[n].
The pairwise procedure and its requirements
A Boolean comparison rule is a function f:{−1,1}n→{−1,1}. Its input is the column of n voters’ signs for one pair of candidates; output +1 means society prefers the first candidate in that pair’s stated orientation, and −1 means society prefers the second.
Applying the same rule f 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)). Its codomain {−1,1}3 has eight triples. Whenever Hf(π)∈R, the encoding table translates the triple back into a unique social ranking in L(C). When Hf(π) is (+1,+1,+1) or (−1,−1,−1), the pairwise decisions form a cycle and fail to define a ranking.
By construction, Hf 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 A and B is f(x), which depends only on column x). Using the same Boolean function f 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,fCA.
Definition 2 (The design requirements) The common-rule procedure Hf is:
unanimous if it respects unanimous agreement on any pair: f(1,…,1)=1,f(−1,…,−1)=−1;
consistent at every profile if it never outputs a cycle: Hf(π)∈Rfor every π∈L(C)n;
a dictatorship if there is a single fixed voter j∈[n] such that f(u)=ujfor every u∈{−1,1}n, in which case Hf(π)=(xj,yj,zj) always copies voter j’s ballot πj. A rule is non-dictatorial if no such j exists.
Our design question is now completely specified: does there exist a unanimous, non-dictatorial Boolean rule f that is consistent at every profile? Note that f 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=3, let f=Maj3, which outputs +1 when at least two of its three input signs are +1, and −1 otherwise. Consider the cyclic profile:
Voter
Ballot
xi (A vs. B)
yi (B vs. C)
zi (C vs. A)
1
A>B>C
+1
+1
−1
2
B>C>A
−1
+1
+1
3
C>A>B
+1
−1
+1
Reading down the columns gives x=(+1,−1,+1), y=(+1,+1,−1), and z=(−1,+1,+1). Since each column contains two +1 signs and one −1 sign, majority outputs +1 on all three: HMaj3(π)=(+1,+1,+1)∈/R. Society prefers A to B, B to C, and C to A. 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 22n possible rules without inspecting 6n 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 6n profiles one by one, suppose each voter i∈[n] independently draws a ballot Πi uniformly at random from the six rankings in L(C) (the impartial culture distribution). Because every profile π∈L(C)n has strictly positive probability 6−n, a rule f is consistent at every profile if and only if P(Hf(Π)∈/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), Y=y(Π), and Z=z(Π) be the three random comparison columns generated by Π. Look first at a single column, say X: for each voter i, exactly three of the six rankings place A above B (Xi=+1) and three place B above A (Xi=−1). Since voters are independent, X is distributed uniformly on the hypercube {−1,1}n, and the same holds marginally for Y and Z.
Across columns, however, a single voter’s signs are coupled by transitivity. When do Xi and Yi agree? Having Xi=Yi=+1 means A≻iB and B≻iC, which by transitivity forces the single ranking A>B>C; similarly, Xi=Yi=−1 forces C>B>A. In the other four rankings—whenever B is ranked first or last rather than in the middle—Xi and Yi must have opposite signs. Therefore, for each voter i, EXi=EYi=0,E[XiYi]=(+1)⋅62+(−1)⋅64=−31. By symmetry of the three cyclic orientations, (Yi,Zi) and (Zi,Xi) follow the exact same joint distribution. Because different voters choose their rankings independently, the pair of vectors (X,Y) consists of n independent coordinate pairs (Xi,Yi) with correlation −1/3.
Now we need an algebraic expression for the event that Hf(Π)=(f(X),f(Y),f(Z)) forms a cycle. For any triple of signs (a,b,c)∈{−1,1}3, consider the sum of the three pairwise products ab+bc+ca:
If a=b=c (one of the two cycles), all three pairs agree, so ab=bc=ca=+1 and ab+bc+ca=3.
If (a,b,c)∈R (a valid ranking), two signs agree and the third is opposite, so one product is +1 and two are −1, giving ab+bc+ca=−1.
Rescaling {−1,3} to {0,1} gives the exact polynomial indicator of a cycle: 1cycle(a,b,c)=41+ab+bc+ca. Notice what this quadratic formula buys us: even though the three vectors (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), (Y,Z), and (Z,X), we obtain P(cycle)=41+E[f(X)f(Y)]+E[f(Y)f(Z)]+E[f(Z)f(X)]=41+43E[f(X)f(Y)]. Our combinatorial design problem has been reduced to a single analytic question: when (X,Y) is a pair of uniform sign vectors whose coordinates are independent with correlation E[XiYi]=−1/3, how does E[f(X)f(Y)] depend on the rule f?
1.3 Product functions as a basis for collective rules
Why was E[f(X)f(Y)] hard to compute directly for a general rule f, yet effortless if f were a product of coordinate signs? Because the coordinate pairs (Xi,Yi) are independent across voters i∈[n], expectation factors over products of distinct coordinates! This observation suggests expanding any function f:{−1,1}n→R as a linear combination of product functions.
Throughout this section, single-cube expectations use a uniform random vector X∈{−1,1}n (independent fair signs X1,…,Xn). For each subset S⊆[n], define the parity function (or character) χS(x)=i∈S∏xi,χ∅(x)=1. Its value is +1 when an even number of coordinates in S are −1, and −1 when that number is odd. Because xi2=1 for every i, multiplying two characters cancels their shared coordinates and keeps those in the symmetric difference S△T=(S∖T)∪(T∖S): χS(x)χT(x)=χS△T(x).
Recall (Bits vs. signs). Why use {−1,1} rather than {0,1}? Under xi=(−1)ui=1−2ui, addition modulo 2 becomes multiplication: (−1)ui⊕vi=xiyi. The parity ⨁i∈Suimod2 thus becomes the monomial χS(x)=∏i∈Sxi—the group characters of F2n (or Walsh–Hadamard basis).
Under independent fair signs, the expectation of any nonempty product is ∏i∈SE[Xi]=0, whereas E[χ∅]=1. Hence the 2n characters are orthonormal with respect to the inner product ⟨f,g⟩=E[f(X)g(X)]: E[χS(X)χT(X)]={1,0,S=T,S=T. Since the vector space of real-valued functions on {−1,1}n has dimension 2n, these 2n 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:
Fix a single row a=(a1,…,an)∈{−1,1}n of the truth table. For each coordinate i, the linear factor 21+aixi equals 1 when xi=ai (since ai2=1) and 0 when xi=−ai. Multiplying these n factors gives the 0/1point indicator of a: 1{x=a}(x)=i=1∏n(21+aixi)=2−nS⊆[n]∑(i∈S∏ai)(i∈S∏xi)=2−nS⊆[n]∑χS(a)χS(x).
Now express any function f:{−1,1}n→R as the sum over all 2n rows of its truth table and exchange the sums over a and S: f(x)=a∈{−1,1}n∑f(a)1{x=a}(x)=a∈{−1,1}n∑f(a)i=1∏n(21+aixi)=S⊆[n]∑2−na∈{−1,1}n∑f(a)χS(a)χS(x)=S⊆[n]∑E[f(X)χS(X)]χS(x).
Combining this explicit truth-table interpolation with the orthonormality of {χS}S⊆[n] yields the fundamental theorem of Boolean Fourier analysis.
Theorem 1 (Fourier expansion and Parseval’s identity) Every function f:{−1,1}n→R has a unique multilinear expansion f(x)=S⊆[n]∑f(S)χS(x),f(S)=⟨f,χS⟩=E[f(X)χS(X)]=2−na∈{−1,1}n∑f(a)χS(a). For any f,g:{−1,1}n→R, orthogonality gives Plancherel’s identity E[f(X)g(X)]=S⊆[n]∑f(S)g(S). In particular, setting g=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.
Proof. Existence of the expansion f=∑Sf(S)χS was just shown by truth-table interpolation; for uniqueness, if f=∑ScSχS, taking the inner product with χT gives ⟨f,χT⟩=∑ScS⟨χS,χT⟩=cT. Expanding both f and g in E[fg] and applying E[χSχT]=1S=T eliminates all cross terms S=T. Finally, f(∅)=E[f⋅1]=Ef, so subtracting f(∅)2 from E[f2] leaves the sum over nonempty subsets.
When f:{−1,1}n→{−1,1} is Boolean-valued, f(x)2=1 at every input, so Parseval states that the squared Fourier coefficients form a probability distribution over the 2n subsets of [n]: S⊆[n]∑f(S)2=1. The cardinality ∣S∣ is called the degree or level of the coefficient f(S), and we group the total energy at degree k∈{0,1,…,n} into the level-k Fourier weightWk(f)=∣S∣=k∑f(S)2,k=0∑nWk(f)=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)), then f(X)χS(X) equals +1 on agreement and −1 on disagreement, so f(S)=E[f(X)χS(X)]=1−2dist(f,χS). Thus f(S) records how strongly f agrees with the parity χS (an interpretation we will use directly for linearity testing in Reading Contents), while the weights W0(f),W1(f),…,Wn(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 ∏i21+aixi and evaluating inner productsf(S)=E[f(X)χS(X)] using symmetry and conditioning—work hand in hand.
Example 2 (Dictator and Parity (single-character rules))
Dictatorf(x)=xi=χ{i}(x): since f is already a basis character, uniqueness of the Fourier expansion gives f({i})=1 and f(S)=0 for all S={i}. All Fourier weight sits on level 1: W1(f)=1 and Wk(f)=0 for k=1.
Parity on T⊆[n]f(x)=χT(x)=∏i∈Txi: similarly, f(T)=1 and f(S)=0 for S=T. If ∣T∣=k, all Fourier weight sits on level k: Wk(χT)=1. In particular, the full parity χ[n](x)=x1⋯xn concentrates all of its energy at the highest degree Wn=1.
Example 3 (Subcube indicator and ANDw (single-row interpolation)) Let A={x∈{−1,1}n:x1=⋯=xw=+1} be the subcube where the first w coordinates are +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 0/1-valued indicator 1A(x) and the corresponding {−1,1}-valued conjunction ANDw(x)=21A(x)−1?
Viewed on the first w coordinates, 1A has a single nonzero row a=(+1,…,+1) in its truth table! Applying our point-indicator formula across i=1,…,w immediately expands 1A(x): 1A(x)=i=1∏w(21+xi)=2−wS⊆[w]∑χS(x). By uniqueness of the Fourier expansion, 1A(S)=2−w for every S⊆[w] and 0 otherwise. (Equivalently, from the expectation definition 1A(S)=E[1A(X)χS(X)]: on the event A of probability 2−w, every xi with i∈[w] is +1, so χS(x)=+1 for all S⊆[w].) For w=2, this gives 1{x1=x2=+1}(x)=41+x1+x2+x1x2.
Converting from {0,1} to {−1,1} via ANDw(x)=21A(x)−1 yields the Fourier expansion of Boolean conjunction: ANDw(x)=(21−w−1)+∅=S⊆[w]∑21−wχS(x), so ANDw(∅)=21−w−1 and ANDw(S)=21−w for every nonempty S⊆[w]. Its level weights are W0(ANDw)=(1−21−w)2 and Wk(ANDw)=(kw)22−2w for 1≤k≤w, which sum to (1−21−w)2+(2w−1)22−2w=1.
Example 4 (Three-voter Majority Maj3 (truth-table interpolation vs. inner products & symmetry)) In §1.1 we used Maj3:{−1,1}3→{−1,1} to exhibit Condorcet’s paradox. Let us compute its Fourier expansion in three ways that highlight our different tools:
By expanding the truth table row by row: Maj3(a) equals +1 on four rows—(+1,+1,+1), (+1,+1,−1), (+1,−1,+1), (−1,+1,+1)—and −1 on their four negations −a. In the truth-table sum ∑a∈{−1,1}3Maj3(a)∏i=1321+aixi, pairing each positive row a with its negation −a cancels all even-degree terms: i=1∏3(21+aixi)−i=1∏3(21−aixi)=4a1x1+a2x2+a3x3+a1a2a3x1x2x3. Summing this expression over the four positive rows (+1,+1,+1), (+1,+1,−1), (+1,−1,+1), and (−1,+1,+1) gives Maj3(x)=41[(x1+x2+x3+x1x2x3)+(x1+x2−x3−x1x2x3)+(x1−x2+x3−x1x2x3)+(−x1+x2+x3−x1x2x3)]=21x1+21x2+21x3−21x1x2x3.
Directly from the inner product f(S)=E[Maj3(X)χS(X)]:
Degrees 0 and 2 (∣S∣ even): Because Maj3(−x)=−Maj3(x) is odd, replacing X by −X negates Maj3(X)χS(X) whenever ∣S∣ is even, so Maj3(∅)=Maj3({i,j})=0.
Degree 1 (S={i}): For S={1}, Maj3(X) agrees with X1 on 3 of the 4 equally likely pairs (X2,X3)∈{−1,1}2 (failing only when X2=X3=−X1), so Maj3({i})=43(+1)+41(−1)=21.
Degree 3 (S={1,2,3}): On the 2 unanimous inputs, Maj3(x) and x1x2x3 agree (product +1); on the 6 two-against-one inputs, Maj3(x) equals the majority sign while x1x2x3 equals the minority sign (product −1). Hence Maj3({1,2,3})=82(+1)+86(−1)=−21.
Shortcut via oddness and permutation symmetry: Since Maj3 is odd and symmetric under permuting x1,x2,x3, it must have the form Maj3(x)=α(x1+x2+x3)+βx1x2x3. Plugging in x=(+1,+1,+1) and x=(+1,+1,−1) gives 3α+β=1 and α−β=1, immediately recovering α=1/2 and β=−1/2.
In all three ways, Maj3 places three quarters of its spectral weight on level 1 and one quarter on level 3: W1(Maj3)=3(21)2=43,W3(Maj3)=(−21)2=41,W0=W2=0.
Example 5 (General odd-n Majority Majn: conditioning and level-1 weight) What does the Fourier spectrum of Majn(x)=sgn(x1+⋯+xn) look like for an arbitrary odd number of voters n=2m+1?
By oddness (Majn(−x)=−Majn(x)), all even levels vanish: W0(Majn)=W2(Majn)=⋯=0. To compute the degree-1 coefficient Majn({i})=E[Majn(X)Xi], condition on the other n−1=2m signs X−i and let S−i=∑j=iXj:
When S−i=0: since S−i is a nonzero even integer (∣S−i∣≥2), adding Xi=±1 cannot change its sign. Thus Majn(X)=sgn(S−i) is already fixed by X−i, and conditioning on X−i gives E[Majn(X)Xi∣X−i]=sgn(S−i)E[Xi]=0.
When S−i=0: the other 2m voters are tied (m pluses and m minuses), so voter i breaks the tie and Majn(X)=Xi. On this tie event, Majn(X)Xi=Xi2=1!
Taking expectations over X−i shows that Majn({i}) equals the probability that the other n−1 fair coins tie: Majn({i})=Pj=i∑Xj=0=2−(n−1)((n−1)/2n−1)∼πn2, where the asymptotic follows from Stirling’s formula (m2m)∼πm4m with m=(n−1)/2. Summing the squares of these n singleton coefficients yields the level-1 Fourier weight of Majn: W1(Majn)=i=1∑nMajn({i})2=n(2−(n−1)((n−1)/2n−1))2⟶π2≈0.6366(n→∞). Even as n→∞, nearly two-thirds (2/π≈63.7%) of Majority’s total Fourier weight remains on the linear level k=1! (We will see in §1.8 why this tie-breaking probability is also the coordinate influence Infi(Majn), and in Reading Contents we work out the higher odd levels W3,W5,… as well.)
Why 2/π? By the central limit theorem, for large n the normalized sum n1∑i=1nXi approaches a standard Gaussian. Sheppard’s arcsine formula (1899) states that for ρ-correlated standard Gaussians (G1,G2), E[sgn(G1)sgn(G2)]=π2arcsinρ=π2ρ+3π1ρ3+⋯. Matching the linear term in ρ with Stabρ(Majn)=∑kρkWk(Majn) explains why W1(Majn)→2/π.
Comparing these four examples already reveals how the Fourier spectrum reflects combinatorial structure: a dictator concentrates all energy on a single degree-1 coordinate (W1=1), full parity pushes all energy to the highest degree (Wn=1), ANDw spreads its variance across degrees 1 through w, and Majn places 2/π of its weight on level 1 while leaving the remaining 1−2/π on higher odd levels k≥3.
1.4 The output correlation we need: noise stability
In the ranking problem, the pair (X,Y) has coordinate correlation ρ=−1/3. It costs nothing—and pays off in Meeting 2—to analyze the same coordinatewise experiment for any correlation parameter ρ∈[−1,1].
Given a uniform sign vector X∈{−1,1}n, generate Y∈{−1,1}n by independently flipping each coordinate Xi with probability (1−ρ)/2 and keeping Yi=Xi with probability (1+ρ)/2. Then each coordinate satisfies E[Yi∣Xi]=(21+ρ−21−ρ)Xi=ρXi,E[XiYi]=ρ. When ρ≥0, flipping with probability (1−ρ)/2 is equivalent to retaining Xi with probability ρ and resampling a fresh fair sign with probability 1−ρ. The flip parameterization has the advantage of working for negative correlations as well: at ρ=−1/3, each sign is flipped with probability 2/3, which is precisely the marginal pair law of (X,Y) under uniform random rankings.
Definition 3 (Noise operator) For h:{−1,1}n→R and −1≤ρ≤1, the noise operatorTρ averages h(Y) conditioned on X=x: (Tρh)(x)=E[h(Y)∣X=x].
By the law of total expectation, the correlation between f(X) and f(Y) is the inner product of f with Tρf:
Definition 4 (Noise stability) For f:{−1,1}n→R and a ρ-correlated uniform pair (X,Y), the noise stability of f at correlation ρ is Stabρ(f)=E[f(X)f(Y)]=E[f(X)E[f(Y)∣X]]=⟨f,Tρf⟩.
Now we see why the character basis was tailor-made for this operator. Because the coordinate flips are independent across i∈[n], conditioning on X=x factors across any product of coordinates: (TρχS)(x)=E[i∈S∏YiX=x]=i∈S∏E[Yi∣Xi=xi]=ρ∣S∣χS(x).
Recall (Chapter 2 semigroup). When ρ=e−t∈(0,1], Te−t=etL is the continuous-time Markov semigroup on {−1,1}n that refreshes each coordinate independently at rate 1 (generator L=−∑i=1nDi). The identity TρχS=ρ∣S∣χS is the exact eigenbasis diagonalization of LχS=−∣S∣χS; see §2.6.
Every character χS is an eigenfunction of Tρ with eigenvalue ρ∣S∣! Expanding f=∑Sf(S)χS and applying Plancherel’s identity immediately diagonalizes the output correlation: Stabρ(f)=⟨f,Tρf⟩=S⊆[n]∑ρ∣S∣f(S)2=k=0∑nρkWk(f). For a Boolean rule f, where ∑k=0nWk(f)=1, Stabρ(f) is simply a weighted average of the powers 1,ρ,ρ2,…,ρn. When ρ<0, odd and even levels contribute with opposite signs—the exact leverage we need to analyze ρ=−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) from §1.4 expresses the probability of a Condorcet cycle directly in terms of the rule’s Fourier weights Wk(f)=∑∣S∣=kf(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} under independent uniform voter rankings, the probability of a cyclic outcome is P(cycle)=41+43Stab−1/3(f)=41+43k=0∑n(−31)kWk(f). Equivalently, using ∑k=0nWk(f)=1, P(cycle)=k=0∑n41+3(−1/3)kWk(f)=W0(f)+31W2(f)+92W3(f)+277W4(f)+⋯.
Proof. In §1.2 we showed that P(cycle)=41+43E[f(X)f(Y)], where (X,Y) is a uniform pair with coordinate correlation ρ=−1/3. Thus E[f(X)f(Y)]=Stab−1/3(f)=∑k=0n(−1/3)kWk(f), giving the first formula. Writing 41=41∑k=0nWk(f) combines the two terms into the level-by-level sum, where the k=1 coefficient is 41+3(−1/3)=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:
For a dictatorf(x)=xi (Example 2), W1(f)=1 and Wk(f)=0 for all k=1. Since the k=1 coefficient in the cycle formula is zero, P(cycle)=0.
For three-voter majorityMaj3 (Example 4), we found W1(Maj3)=3/4 and W3(Maj3)=1/4, so P(cycle)=92W3(Maj3)=181 (matching the 12 cyclic profiles—six with (+1,+1,+1) and six with (−1,−1,−1)—out of 63=216 total profiles). For general odd n, since W1(Majn)→2/π<1 (Example 5), the remaining weight 1−W1(Majn)→1−2/π sits on higher odd levels k≥3 and keeps P(cycle) bounded away from zero for every n≥3 (see Reading Contents).
Most importantly, for any Boolean rule f, the coefficient 41+3(−1/3)k is zero at k=1 and strictly positive (≥92) for every k=1. Because the weights Wk(f) are nonnegative, P(cycle)=0 forces Wk(f)=0 for every k=1, and therefore W1(f)=1! A universally consistent Boolean rule must have all of its Fourier mass supported on level one.
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} satisfies W1(f)=1, then f(x)=xi or f(x)=−xi for some coordinate i.
Proof. Because W1(f)=1 and ∑k=0nWk(f)=1, all Fourier coefficients outside level one vanish. Thus f(x)=∑i=1nbixi with ∑i=1nbi2=1. The Boolean range f(x)∈{−1,1} now imposes a much stronger pointwise constraint: f(x)2=1 at every input x. Expanding the square in the character basis gives 1=f(x)2=(i=1∑nbi2)χ∅(x)+21≤i<j≤n∑bibjχ{i,j}(x). Since the characters {χS} form a basis, matching coefficients of χ{i,j} on both sides forces bibj=0 for every distinct pair i<j. Hence at most one coefficient bi can be nonzero; since ∑i=1nbi2=1, that single coefficient must be ±1, so f(x)=±xi.
Theorem 3 (Arrow’s theorem: the three-candidate common-rule version) Suppose a unanimous Boolean rule f 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 for one fixed voter i.
Proof. The hypothesis is Hf(π)∈R for every profile π∈Rn. Because every profile has positive weight 6−n>0 under the uniform-ranking experiment, universal consistency implies P(cycle)=0. The spectral identity in §1.5 then forces W1(f)=1, and Lemma 1 gives f=±xi. Unanimity (f(1)=+1) rules out −xi. Conversely, the dictator rule f(x)=xi simply copies voter i’s strict ranking on every pair, which is always transitive.
General Arrow theorem. Classical social choice allows three distinct pairwise rules (fAB,fBC,fCA) 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 i is fixed by the rule f 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 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 n 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 n participants who want to generate a common random bit in a single broadcast round. Each participant i broadcasts a sign xi∈{−1,1}, and everyone applies a publicly known deterministic Boolean rule f:{−1,1}n→{−1,1} to the broadcast vector x=(x1,…,xn). When all participants are honest, their broadcasts are independent and uniform on {−1,1}. We call fbalanced when E[f(X)]=0, meaning that an all-honest execution produces a fair coin (P(f(X)=+1)=1/2).
Now suppose a coalition S⊆[n] of dishonest participants colludes to bias the outcome toward +1. We study a single-round rushing adversary: the coalition S is fixed before the protocol begins, waits to see the broadcasts X[n]∖S of the honest participants, and then coordinates its own signs yS∈{−1,1}S to force f(X[n]∖S,yS)=+1 whenever possible. Here xS=(xi)i∈S, and f(x[n]∖S,yS) denotes f evaluated on the combined vector taking xi outside S and yi inside S. Under uniform honest broadcasts X[n]∖S, the coalition’s maximum achievable probability of forcing +1 is p+(S)=PX[n]∖S(there exists yS∈{−1,1}S such that f(X[n]∖S,yS)=+1). Our design question is concrete: can we construct a balanced rule f for which every small coalition S keeps p+(S) close to 1/2? 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:
Parityf(x)=x1x2⋯xn: every variable looks completely symmetric, yet corrupting any single participant i (∣S∣=1) still gives p+({i})=1! Because the adversary rushes, participant i simply waits to see the product ∏j=iXj∈{−1,1} of the remaining n−1 honest signs and chooses yi to make the total product +1. (If participants had to commit before seeing others’ bits, parity would be perfectly resilient; the rushing order of information is decisive.)
MajorityMajn(x) (for odd n): outputs +1 when ∑i=1nxi>0. A coalition S of size k trying to force +1 should simply set yi=+1 for all i∈S. Then the output is +1 whenever the sum H=∑i∈/SXi of the n−k honest signs satisfies H>−k. Because H has mean 0 and standard deviation n−k≈n, a coalition of size k=Θ(n) shifts the threshold by several standard deviations and drives p+(S) close to 1 (for instance, Chebyshev’s inequality gives 1−p+(S)=P(H≤−k)≤(n−k)/k2).
Dictator and parity fall to a coalition of size 1; majority survives up to coalition size Θ(n). Can a cleverer balanced rule resist coalitions of size Θ(n)? We will prove that no balanced rule can: every balanced f admits a coalition of size O(n/logn)—a vanishing o(1) fraction of the participants—that forces +1 with probability at least 0.99. 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 n can we get? Ben-Or and Linial [5] showed that iterated majority on ternary trees resists rushing coalitions of size O(nlog32)≈O(n0.631). Ajtai and Linial (1993) later took an AND of O(logn) random Tribes functions on overlapping coordinate partitions to balance the +1 and −1 vulnerabilities, resisting coalitions of size Ω(n/log2n) in both directions—just one logn factor below the O(n/logn) upper bound we prove via KKL [2]. Closing that remaining logn gap between Ω(n/log2n) and O(n/logn) 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 i control the output of f? Exactly when the other n−1 honest broadcasts fall in a configuration where flipping coordinate i changes the value of f. Let x⊕i denote x with its i-th coordinate flipped (xi↦−xi).
Definition 5 (Influence and total influence) For a Boolean function f:{−1,1}n→{−1,1}, the influence of coordinate i is the probability that coordinate i is pivotal under a uniform random input: Infi(f)=PX∼Unif({−1,1}n)(f(X)=f(X⊕i)). The total influence of f is I(f)=∑i=1nInfi(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 f along coordinate i by Dif(x)=2f(x)−f(x⊕i). Because f takes values in {−1,1}, the difference f(x)−f(x⊕i) is 0 when f(x)=f(x⊕i) and ±2 when f(x)=f(x⊕i). Therefore hi(x)=Dif(x) takes values in {−1,0,1}, and its square is the exact 0/1 indicator of the pivotal event: (Dif(x))2=1{f(x)=f(x⊕i)},Infi(f)=E[(Dif(X))2]=∥Dif∥22. Now evaluate Di on the character basis. Flipping xi leaves χS(x) unchanged if i∈/S and negates χS(x) if i∈S. Consequently, 2χS(x)−χS(x⊕i) equals χS(x) when i∈S and 0 when i∈/S. By linearity, applying Di simply projects the Fourier expansion onto the characters that contain i: Dif=S∋i∑f(S)χS. Applying Parseval’s identity to Dif immediately yields the Fourier formulas for individual and total influence: Infi(f)=S∋i∑f(S)2,I(f)=i=1∑nInfi(f)=S⊆[n]∑∣S∣f(S)2.
Geometrically, Infi(f) is the fraction of direction-i edges whose endpoints get opposite signs (red edges above for Maj3, giving Inf1(Maj3)=1/2). Notation: our Dif keeps xi inside χS; O’Donnell [3] writes Dibookf=xiDif, which has the same ∣Dif∣ everywhere.
The first formula says that participant i’s influence is the total Fourier weight of all interactions involving i; the second says that total influence weights each Fourier level by its degree ∣S∣, because a coalition S of size k is counted once by each of its k members.
Example 6 (When is a voter pivotal in Majority?) For odd n, voter i is pivotal in Majn precisely when the other n−1 honest signs split into (n−1)/2 pluses and (n−1)/2 minuses. By Stirling’s approximation for the central binomial coefficient, Infi(Majn)=2−(n−1)((n−1)/2n−1)∼πn2,I(Majn)∼π2n. Compare this with Example 5 in §1.3: for Majn, the influence Infi(Majn) is equal to the degree-1 Fourier coefficient Majn({i})! (In Reading Contents we show that Infi(f)=f({i}) holds for every monotone Boolean function.) For n=3, Infi(Maj3)=2−2(12)=1/2; the Fourier formula Infi(f)=∑S∋if(S)2 gives the same answer from Example 4, since coordinate i participates in Maj3({i})=1/2 and Maj3({1,2,3})=−1/2, contributing (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] and wants to force +1. Define the residual success function on the remaining honest coordinates [n]∖S by fS(x[n]∖S)=yS∈{−1,1}Smaxf(x[n]∖S,yS), where −1<+1. By definition, fS(x[n]∖S)=+1 if and only if the coalition S can force +1 on the honest draw x[n]∖S, so P(fS=+1)=p+(S). Crucially, fS is again a Boolean function taking values in {−1,1} on the (n−∣S∣)-dimensional cube of untouched honest coordinates—even when the original rule f is neither monotone nor symmetric.
Now suppose the coalition recruits one additional participant i∈/S. Since maximizing over S∪{i} is the same as maximizing over yi∈{−1,1} after maximizing over yS∈{−1,1}S, we have fS∪{i}(x[n]∖(S∪{i}))=yi∈{−1,1}maxfS(x[n]∖(S∪{i}),yi). Fix the remaining honest signs x[n]∖(S∪{i}). There are three possibilities for the pair of values fS(⋅,+1) and fS(⋅,−1):
Both are +1: then fS=+1 and fS∪{i}=+1, so recruiting i changes nothing.
Both are −1: then fS=−1 and fS∪{i}=−1, so recruiting i changes nothing.
One is +1 and the other is −1 (that is, coordinate i is pivotal for fS): when i was honest, its random sign Xi∈{−1,1} gave +1 with probability 1/2; once i joins the coalition, it chooses the winning sign and gives +1 with probability 1.
Thus recruiting i increases the success probability from 1/2 to 1 on the pivotal event of probability Infi(fS), and changes nothing elsewhere! This gives the exact one-step gain identity: p+(S∪{i})−p+(S)=21Infi(fS).
How large an influence can our Fourier formulas guarantee on an arbitrary Boolean function h with r=n−∣S∣ remaining coordinates? Since ∣A∣≥1 for every nonconstant Fourier term A=∅, 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).
Recall (Poincaré from Chapter 2). For any reversible Markov chain with spectral gap γ, Chapter 2 showed Var(h)≤γ1E(h,h). On the hypercube with rate-1 coordinate refreshes, the Dirichlet form is E(h,h)=∑i=1r∥Dih∥22=I(h) and γ=1. On the Fourier side, this becomes the one-line inequality 1≤∣A∣ for A=∅.
By averaging over the r coordinates, some coordinate i must satisfy Infi(h)≥I(h)/r≥Var(h)/r. Because h∈{−1,1} with p=P(h=+1), its variance is Var(h)=4p(1−p), so greedy recruitment increases p+(S) by at least p+(S∪{i})−p+(S)≥2rVar(fS)=r2p(1−p).
Caution (Why Poincaré stalls at a trivial Θ(n) coalition).
Near p=1/2, Poincaré guarantees an increment of only Θ(1/n) per recruited coordinate. At rate 1/n, moving p+(S) from 1/2 to 0.99 requires Θ(n) coordinates—a constant fraction of all participants, which is trivial. To obtain a sublinear coalition of size o(n), we need an influence lower bound that beats Var(h)/r by an unbounded factor on every intermediate Boolean function h=fS.
Before trying to prove such an inequality, let us pause and ask a benchmark question: how small can the maximum individual influence maxiInfi(f) actually be for a balanced Boolean function? For Dictator it is 1; for Majority it is Θ(n−1/2). Can a balanced function push every coordinate’s influence all the way down to the Poincaré floor Θ(1/n), or is there a genuine barrier above 1/n?
1.10 Tribes: the Θ((logn)/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((logn)/n), far smaller than Majority’s Θ(n−1/2), but stop just Θ(logn) above the Poincaré floor 1/n.
Partition the n=mw coordinates into m disjoint blocks (“tribes”) T1,…,Tm, each of width w. Define f(x)=+1 if at least one tribe consists entirely of +1 signs, and f(x)=−1 otherwise (in 0/1 logic, an OR of m disjoint ANDs of width w).
Because the tribes are disjoint, each tribe is unanimously +1 with probability 2−w independently of the others. Hence the probability that no tribe succeeds is P(f(X)=−1)=(1−2−w)m.
Because (1−2−w)2w→e−1, choosing m≈(ln2)2w makes the all-fail probability (1−2−w)m≈e−ln2=1/2.
To make f approximately balanced (P(f(X)=−1)≈1/2), we choose m and w so that (1−2−w)m≈1/2. Taking logarithms gives mlog(1−2−w)≈−log2, so m∼(log2)2w and therefore n=mw∼(log2)w2w, which means w=log2n−log2log2n+O(1)=Θ(logn).
When is one coordinate pivotal in Tribes?
Participant i (belonging to tribe T1, say) can change the outcome of f if and only if two independent events hold simultaneously:
Inside i’s own tribe: all other w−1 members of T1 broadcast +1 (otherwise T1 fails regardless of xi), which happens with probability 2−(w−1);
Outside i’s tribe: none of the remaining m−1 tribes is unanimously +1 (otherwise f=+1 regardless of xi), which happens with probability (1−2−w)m−1≈1/2.
Multiplying these two independent probabilities gives the exact influence of every coordinate: Infi(f)=2−(w−1)(1−2−w)m−1. Since 2−w∼mlog2=nwlog2, the detailed asymptotic check in Reading Contents gives Infi(f)=Θ(nlogn),I(f)=nInfi(f)=Θ(logn),Var(f)→1.
This single example teaches us two essential lessons for the rest of the chapter:
Why Θ((logn)/n) is the exact target for KKL: Because Tribes has Var(f)≈1 and maxiInfi(f)=Θ((logn)/n), no universal theorem can improve on Poincaré’s Var(f)/n bound by more than a Θ(logn) factor. At the same time, a logn improvement is just enough to break the 1/n barrier in §1.9 and reduce the greedy coalition size from O(n) to O(n/logn)!
Small individual influence does not mean coalition resilience: Even though every single coordinate in Tribes has tiny influence Θ((logn)/n), a rushing coalition that corrupts just one full tribe of w=Θ(logn) participants can set all w signs to +1 and force f=+1 with probability 1. Thus Tribes shows that KKL’s individual-influence bound Ω((logn)/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 logn factor—or more precisely, a factor of log(1/maxiInfi(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/n bound in §1.9: 1≤i≤nmaxInfi(f)(1) averaging≥nI(f)(2) Poincareˊ≥nVar(f). For the combined bound maxiInfi(f)≥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:
Step (1) (Averaging) is tight when influence is spread evenly across all n coordinates, so every individual influence ai=Infi(f) is tiny (a=maxiai≈I(f)/n≪1).
Step (2) (Poincaré) is tight when replacing ∣S∣ by 1 in I(f)=∑S=∅∣S∣f(S)2=∑k=1nkWk(f) loses almost nothing—that is, when the variance Var(f)=∑k=1nWk(f) sits entirely on Fourier level 1 (W1(f)≈Var(f) and Wk(f)≈0 for all k≥2).
Now compare what happens for arbitrary real-valued functions versus Boolean-valued functions:
For real-valued functions g:{−1,1}n→R, (1) and (2) can be tight simultaneously. Take the normalized linear sum g(x)=n1i=1∑nxi,Dig(x)=nxi,∥Dig∥22=n1. Here g is purely linear (W1(g)=Var(g)=I(g)=1, so (2) holds with equality) and symmetric (maxi∥Dig∥22=1/n, so (1) holds with equality). This example delivers a crucial lesson: no purely L2 argument can improve on Var(f)/n. Any stronger inequality must exploit the Boolean range f(x)∈{−1,1}.
For Boolean functions f:{−1,1}n→{−1,1}, (1) and (2) are in head-on conflict. Suppose a balanced Boolean function makes Poincaré (step (2)) tight: I(f)=Var(f)=1, so W1(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)=1 are the dictatorsf(x)=±xj. And for a dictator, step (1) fails as badly as possible: one coordinate has influence 1 and the other n−1 have influence 0, giving maxiInfi(f)=1≫1/n! Conversely, if every coordinate has small influence (a=maxiInfi(f)≪1, so we are in the dispersed regime where (1) is tight), f is far from a dictator, so its variance cannot live on level 1—it is forced up to higher degrees k≥2, where I(f)=∑kkWk(f) pays a factor of k and makes Poincaré (step (2)) loose.
Real vs. Boolean derivatives. For the real-valued counterexample g(x)=n−1/2∑ixi, the derivative Dig(x)=n−1/2xi has tiny nonzero magnitude ∣Dig(x)∣=n−1/2 at every vertex of the cube. For a Boolean function f:{−1,1}n→{−1,1}, Dif(x)=2f(x)−f(x⊕i) can only take values in {−1,0,1}: small influence ai=∥Dif∥22≪1 therefore means Dif is sparse (zero almost everywhere, and ±1 on a rare pivotal set of measure ai).
Warm-up at degree 1: why small influences expel Fourier mass from W1(f)
Before tackling general degrees, let us see by a two-line calculation how the Boolean range f(x)∈{−1,1} prevents variance from sitting on level 1 whenever a=maxiInfi(f) is small. Write V=Var(f),ai=Infi(f),a=1≤i≤nmaxai. Because f is {−1,1}-valued, its discrete derivative hi=Dif∈{−1,0,1} is nonzero on a fraction ai=P(hi(X)=0) of the cube, so its L1 norm is E∣hi(X)∣=ai (unlike the real-valued derivative ∣Dig∣≡n−1/2, whose L1 norm is ai=n−1/2). Recall from §1.8 that Dif(x)=∑S∋if(S)χS(x). Multiplying by xi∈{−1,1} and taking expectations extracts the singleton coefficient S={i}, so ∣f({i})∣=∣E[XiDif(X)]∣≤E∣Dif(X)∣=ai. Squaring and summing over i=1,…,n immediately bounds the entire level-1 Fourier weight: W1(f)=i=1∑nf({i})2≤i=1∑nai2≤ai=1∑nai=aI(f). Look at what happened: because the {−1,0,1} sparsity of Dif gave f({i})2≤ai2 (an extra power of ai beyond the L2 bound ∥Dif∥22=ai), the level-1 weight W1(f) is at most aI(f). When every coordinate has tiny influence (a≪1), W1(f)≤aI(f)≪I(f) is only a tiny fraction of total influence I(f): thus if a noticeable fraction of the variance V sits on level 1 (as in Majority, where W1≈2/π and I(Majn)≥W1/a∼2n/π), I(f) is already huge of order V/a; and if I(f) is not that large (as in Tribes, where I(f)∼logn), then small individual influences force variance out of degree 1 and up to higher Fourier degrees.
From degree 1 to degree d: pushing variance above degree d
Being pushed from degree 1 to degree 2 would only double I(f)=∑kkWk(f). To prove I(f)≳log(1/a)V, we want to generalize W1(f)≤aI(f) to all degrees ≤d≍log(1/a), showing that when a=maxiai≪1, the low-degree weight up to degree d is a tiny fraction of I(f)!
Fix a degree cutoff d≥1 and split the variance V=∑S=∅f(S)2 into its low-degree part (≤d) and high-degree part (>d): V=Mlow+Mhigh,Mlow=1≤∣S∣≤d∑f(S)2,Mhigh=∣S∣>d∑f(S)2. Why do we win as soon as most of the variance sits in Mhigh? Because I(f)=∑S∣S∣f(S)2 multiplies every Fourier coefficient of degree ∣S∣>d by at least d+1: I(f)≥(d+1)∣S∣>d∑f(S)2=(d+1)Mhigh,or equivalently,Mhigh≤d+1I(f). Notice what this says: Mhigh/V is not small in the bottleneck regime—on the contrary, whenever Mhigh≥21V, we immediately get I(f)≥2d+1V! The only way I(f) could avoid paying this (d+1) factor is if the variance stayed at low degrees 1≤∣S∣≤d (inside Mlow). Therefore, our entire task is to generalize our degree-1 bound W1(f)≤aI(f) to all degrees ≤d, showing that when a≪1, Mlow≪I(f) up to d≍log(1/a)—so that combining Mlow≪I(f) with Mhigh≤I(f)/(d+1) forces V=Mlow+Mhigh≲I(f)/(d+1).
What does small Mlow mean combinatorially? For any subset S⊆[n] of ∣S∣≤d coordinates, revealing XS gives conditional mean E[f(X)∣XS]=∑T⊆Sf(T)χT(XS), whose variance is VarXS(E[f(X)∣XS])=∑∅=T⊆Sf(T)2≤Mlow. Thus Mlow≪V means that for any chosen set of at most d coordinates, the conditional mean stays close to the global mean E[f] for most revealed assignments XS—revealing ≤d bits gives almost no local hint about f(X), forcing its +1/−1 regions to interweave finely with boundary I(f)≥(d+1)Mhigh.
To bound Mlow in terms of the sparse derivatives hi=Dif∈{−1,0,1}, define the low-degree Fourier energy of hi by Bd(hi)=∣S∣≤d∑hi(S)2=S∋i,∣S∣≤d∑f(S)2. Since every nonempty set S has ∣S∣≥1 and is counted ∣S∣ times across ∑i=1nBd(hi), we have Mlow=1≤∣S∣≤d∑f(S)2≤1≤∣S∣≤d∑∣S∣f(S)2=i=1∑nBd(hi). Our entire problem now reduces to a single question about a sparse function hi∈{−1,0,1} supported on a fraction ai=P(hi=0) of the cube: how well can we bound its low-degree energy Bd(hi) in terms of ai?
Caution (Why naive L2 and L1 bounds on Bd(hi) fail for opposite reasons).
At degree 1, hi contributed only a single singleton coefficient f({i})2≤ai2. Up to degree d, however, Bd(hi) sums over all subsets ∣S∣≤d, and the two elementary bounds fail for opposite reasons:
Attempt 1 (Parseval / L2 orthogonality): Dropping the degree restriction ∣S∣≤d gives Bd(hi)≤S∑hi(S)2=∥hi∥22=ai. This bound is dimension-free because it uses the orthogonality of the characters χS, but it is completely blind to the Boolean sparsity hi∈{−1,0,1}: the real-valued counterexample ∣Dig∣≡n−1/2 has the exact same ∥Dig∥22=ai! Summing Bd(hi)≤ai over i just gives Mlow≤I(f), returning us to Poincaré.
Attempt 2 (Pointwise L1 bound on each coefficient, as in our degree-1 warm-up): Because ∣hi∣∈{0,1}, we have E∣hi∣=ai, so every individual Fourier coefficient satisfies ∣hi(S)∣=∣E[hiχS]∣≤E∣hi∣=ai and hence hi(S)2≤ai2. At d=1 there was only one term per i; up to degree d, summing hi(S)2≤ai2 over all sets of size at most d yields Bd(hi)≤(k=0∑d(kn))ai2. This bound uses the {−1,0,1} sparsity of hi to gain an extra power of ai, but bounding each coefficient separately destroys orthogonality and pays the combinatorial factor ∑k=0d(kn)∼nd/d!, which blows up with the ambient dimension n.
Look at the complementary virtues of the two attempts. Attempt 1 is dimension-free because it keeps all characters together in an L2 sum, while Attempt 2 extracts an extra power of ai because it measures the {−1,0,1}-valued derivative hi in a norm (L1) with exponent smaller than 2. 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 for universal constants C>1 and η>0independent of n (paying an exponential factor Cd in the degree d rather than nd in the dimension n). Indeed, if such an estimate holds, summing over i=1,…,n gives Mlow≤i=1∑nCdai1+η≤Cdaηi=1∑nai=CdaηI(f), and combining with Mhigh≤I(f)/(d+1) immediately yields V=Mlow+Mhigh≤I(f)(Cdaη+d+11). Choosing d≍log(1/a) so that Cdaη≪1 will then force V≲log(1/a)I(f), giving the exact logarithmic gain over Poincaré!
Our analytic goal is now crystal clear: for a sparse function h∈{−1,0,1} with P(h=0)=α, prove a dimension-free low-degree bound Bd(h)≤Cdα1+η. The next two sections show how the noise operator Tρ 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} be supported on a set A={x:h(x)=0} of probability α=P(A). Why was it hard to bound the low-degree sum Bd(h)=∑∣S∣≤dh(S)2 directly? Because the sharp degree cutoff 1{∣S∣≤d} has no clean representation in the physical domain {−1,1}n where the sparsity ∣h∣=1A lives.
Step 1: Smooth degree damping replaces the sharp cutoff
In §1.4 we saw that the noise operator Tρ multiplies each character χS by ρ∣S∣. In Meeting 1 we took ρ=−1/3 to encode ranking correlations; now take a fixed positive parameter ρ∈(0,1). By Parseval, the L2 energy of the smoothed function Tρh is ∥Tρh∥22=S⊆[n]∑ρ2∣S∣h(S)2. Instead of cutting off abruptly at degree d, Tρ damps every Fourier level smoothly by ρ2∣S∣. Because ρ2∣S∣≥ρ2d whenever ∣S∣≤d, every low-degree term retains at least a ρ2d fraction of its squared mass in ∥Tρh∥22: Bd(h)=∣S∣≤d∑h(S)2≤ρ−2d∣S∣≤d∑ρ2∣S∣h(S)2≤ρ−2d∥Tρh∥22. Look at what this simple inequality achieves: it produces the exact exponential degree factor Cd=ρ−2d we anticipated in §1.11, with no dependence on the dimension n! Moreover, unlike the sharp cutoff Bd(h), the operator Tρ is a genuine averaging operator on the cube: for 0≤ρ≤1, generating Y from x by keeping each coordinate xi with probability ρ and resampling it uniformly with probability 1−ρ gives E[Yi∣x]=ρxi, so Tρh(x)=E[h(Y)∣X=x].
Step 2: What ∥Tρh∥22 means probabilistically
Before bounding ∥Tρh∥22 analytically, let us see what it measures in the physical domain. Because Tρ is diagonal in the orthonormal basis {χS}, it is self-adjoint and satisfies the semigroup law: ⟨g,Tρh⟩=S∑ρ∣S∣g(S)h(S)=⟨Tρg,h⟩,TρTσ=Tρσ. Therefore, if (X,Z) is a pair of uniform inputs with coordinate correlation ρ2, ∥Tρh∥22=⟨Tρh,Tρh⟩=⟨h,Tρ2h⟩=E[h(X)h(Z)]. When h=1A is the indicator of a rare event A of measure α, ∥Tρ1A∥22=P(X∈A,Z∈A) is simply the probability that A persists after random noise! If we applied no noise (ρ=1, so Z=X), this probability would be α; if we resampled completely (ρ=0, so Z⊥X), it would be α2. For a fixed noise rate 0<ρ<1, a rare event should still struggle to survive resampling, so we expect P(X∈A,Z∈A)≤α1+η for some η>0.
For example, take the two-coordinate subcube A={x1=x2=+1} of probability α=1/4, and set ρ2=1/3. For each coordinate i∈{1,2}, keeping the sign with probability 1/3 and resampling otherwise gives P(Xi=Zi=+1)=21(31+32⋅21)=31. By independence across coordinates, P(X∈A,Z∈A)=(1/3)2=1/9, which is already strictly smaller than α=1/4 (in fact, 1/9<1/8=α3/2). We can verify the exact same number from the Fourier expansion 1A(x)=41+x1+x2+x1x2 in Example 3: ∥Tρ1A∥22=S⊆{1,2}∑ρ2∣S∣(41)2=161+2ρ2+ρ4=(41+ρ2)2=91. Our goal is to prove this α3/2 persistence bound for every set A and every {−1,0,1}-valued function h of support α, regardless of the dimension n or the geometry of A.
Step 3: Recording rarity in L4/3 and dualizing to a 2→4 norm bound
How do we express the rarity of h∈{−1,0,1} analytically without losing orthogonality? For any p≥1, define the Lp norm on the cube by ∥h∥p=(E∣h(X)∣p)1/p. Because ∣h(x)∣∈{0,1} at every point, ∣h(x)∣p=1A(x) for every exponent p, so E∣h∣p=α and hence ∥h∥p=α1/p. When p=2, ∥h∥2=α1/2, so ∥h∥22=α gives no extra power of α. But whenever we choose an exponent p<2, the 1/p power is strictly larger than 1/2! In particular, for p=4/3, ∥h∥4/3=α3/4,∥h∥4/32=α3/2. Thus, if we can prove the norm inequality ∥Tρh∥2≤∥h∥4/3 for some fixed ρ>0, squaring both sides will immediately give ∥Tρh∥22≤∥h∥4/32=α3/2 (so η=1/2).
Why is p=4/3 so convenient? Because its Hölder conjugate exponent q (satisfying p1+q1=1, i.e., 43+41=1) is q=4, an even integer! Using the L2 duality identity ∥u∥2=sup∥g∥2=1∣⟨g,u⟩∣, the self-adjointness of Tρ, and Hölder’s inequality ∣E[hk]∣≤∥h∥4/3∥k∥4 (proved in Reading Contents), we can pass the operator Tρ from the sparse function h to a test function g∈L2: ∥Tρh∥2=∥g∥2=1sup∣⟨g,Tρh⟩∣=∥g∥2=1sup∣⟨Tρg,h⟩∣≤∥h∥4/3∥g∥2=1sup∥Tρg∥4.
Recall (Hölder via double Cauchy–Schwarz). For conjugate exponents p1+q1=1, ∣E[hk]∣≤∥h∥p∥k∥q (proved for p=4/3,q=4 in §2.11). When ∣h∣=1A with P(A)=α, you can also see (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.
Every step of our reduction is now complete. To establish ∥Tρh∥2≤∥h∥4/3—and with it the rare-event bound Bd(hi)≤ρ−2dai3/2 and KKL—we only need to find a fixed noise rate ρ>0 such that ∥Tρg∥4≤∥g∥2 for every function g on {−1,1}n. And because q=4 is an even integer, ∥Tρg∥44=E[(Tρg)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 is asking for. On any probability space, Cauchy–Schwarz implies ∥g∥2=(E[g2])1/2≤(E[g4])1/4=∥g∥4: the L4 norm is always larger than the L2 norm because fourth powers heavily penalize sharp spikes. For instance, if g=α−1/21A on a small set A of measure α, then ∥g∥2=1 whereas ∥g∥4=α−1/4→∞ as α→0. For ∥Tρg∥4≤∥g∥2 to hold, the noise averaging Tρ must smooth out those spikes enough to reverse the natural ordering between L4 and L2. An operator that maps Lp into a higher norm Lq (q>p) is called hypercontractive.
A sparse spike g=α−1/21A has ∥g∥2=1 but huge fourth moment ∥g∥4=α−1/4. Noise averaging T1/3 spreads the spike across neighbors so that its L4 norm drops below ∥g∥2.
How small must ρ be for Tρ to reverse the L4-versus-L2 comparison? We can discover the exact threshold ρ on a single bit (n=1), and then verify that the same threshold works in every dimension n by induction.
On a single uniform sign x∈{−1,1}, every real-valued function has the linear Fourier expansion g(x)=u+vx, where u=g(∅) and v=g({1}). Its second moment is ∥g∥22=u2+v2, so ∥g∥24=(u2+v2)2=u4+2u2v2+v4. Now apply Tρ: we get Tρg(x)=u+ρvx. Averaging its fourth power over x=±1 kills the odd powers of x and gives ∥Tρg∥44=2(u+ρv)4+(u−ρv)4=u4+6ρ2u2v2+ρ4v4. Compare the two polynomials term by term: ∥g∥24 has middle coefficient 2u2v2, whereas ∥Tρg∥44 has middle coefficient 6ρ2u2v2! Without noise (ρ=1), the coefficient 6 exceeds 2. To make u4+6ρ2u2v2+ρ4v4≤u4+2u2v2+v4 hold for all u,v∈R, we need 6ρ2≤2 (and ρ4≤1), which forces ρ≤31. At the exact threshold ρ=1/3, we have 6ρ2=2 and ρ4=1/9≤1, so ∥T1/3g∥44=u4+2u2v2+91v4≤u4+2u2v2+v4=(u2+v2)2=∥g∥24. Taking fourth roots proves ∥T1/3g∥4≤∥g∥2 on a single bit! Remarkably, this one-bit calculation already gives the exact constant for the n-dimensional hypercube.
Theorem 4 (The dimension-free 2→4 hypercontractive inequality (Bonami)) For every n≥1 and every real-valued function g on {−1,1}n, ∥T1/3g∥4≤∥g∥2.
Historical note. The general hypercontractive inequality ∥Tρg∥q≤∥g∥p for ρ2≤q−1p−1 (1<p≤q<∞) was proved by Aline Bonami (1970) in harmonic analysis on Z2n, 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 n. The base case n=1 was just proved above.
For the induction step, split the input into the first n−1 coordinates x′∈{−1,1}n−1 and the last coordinate xn∈{−1,1}, and decompose g accordingly: g(x′,xn)=g0(x′)+xng1(x′),x′∈{−1,1}n−1. Set ρ=1/3. Because Tρ acts independently on each coordinate, applying Tρ on {−1,1}n applies the (n−1)-coordinate operator Tρ to g0 and g1 and multiplies xn by ρ: Tρg(x′,xn)=A(x′)+ρxnB(x′),where A=Tρg0,B=Tρg1. Notice that A(x′) and B(x′) depend only on x′ and are independent of the uniform sign xn. Taking the expectation of (A(x′)+ρxnB(x′))4 first over xn∈{−1,1} (using E[xn]=E[xn3]=0 and xn2=xn4=1) and then over x′ gives ∥Tρg∥44=E[A4]+6ρ2E[A2B2]+ρ4E[B4]. Write u=∥g0∥2 and v=∥g1∥2. By the induction hypothesis on {−1,1}n−1, we have ∥A∥4≤∥g0∥2=u and ∥B∥4≤∥g1∥2=v, so E[A4]≤u4,E[B4]≤v4. To bound the cross term E[A2B2], apply the Cauchy–Schwarz inequality to the random variables A2 and B2: E[A2B2]≤(E[A4])1/2(E[B4])1/2≤u2v2. Substituting ρ=1/3 (6ρ2=2 and ρ4=1/9) therefore yields ∥Tρg∥44≤u4+2u2v2+91v4≤(u2+v2)2. Finally, averaging g(x′,xn)2=g0(x′)2+2xng0(x′)g1(x′)+g1(x′)2 over xn and x′ gives ∥g∥22=∥g0∥22+∥g1∥22=u2+v2. Thus ∥Tρg∥44≤∥g∥24, and taking fourth roots completes the induction.
Look at why the induction succeeded without losing any constant factor across n steps: Cauchy–Schwarz on E[A2B2] reduced the (n−1)-dimensional cross term to the exact same algebraic inequality u4+2u2v2+91v4≤(u2+v2)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=T1/3. 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 h on {−1,1}n, ∥Th∥2≤∥h∥4/3. In particular, if h∈{−1,0,1} is supported on a set A={x:h(x)=0} of probability α=P(A), then ∥Th∥22≤α3/2.
Proof. In the duality inequality ∥Th∥2≤∥h∥4/3sup∥g∥2=1∥Tg∥4 from §1.12, Theorem 4 shows that sup∥g∥2=1∥Tg∥4≤1, proving ∥Th∥2≤∥h∥4/3. When ∣h∣=1A with P(A)=α, we have ∥h∥4/3=α3/4, so squaring gives ∥Th∥22≤α3/2.
Before returning to derivatives, notice what Corollary 1 says about any event A⊆{−1,1}n of probability α>0: for a pair of inputs (X,Z) with coordinate correlation ρ2=1/3, P(X∈A,Z∈A)=∥T1A∥22≤α3/2⟹P(Z∈A∣X∈A)≤α. No matter how A is embedded in the n-dimensional cube, resampling two-thirds of the coordinates forces the conditional survival probability of A below α.
Now apply Corollary 1 to the pivotal derivative hi=Dif∈{−1,0,1}, which is nonzero on a set of probability ai=Infi(f). Combining the degree-damping bound Bd(Dif)≤ρ−2d∥TDif∥22 (where ρ−2=(1/3)−2=3) with ∥TDif∥22≤ai3/2 gives Bd(Dif)≤3d∥TDif∥22≤3dai3/2. This is precisely the missing low-degree rarity estimate Bd(hi)≤Cdai1+η from §1.11, with concrete constants C=3 and η=1/2! Whenever 3dai<1, it beats the naive Parseval bound Bd(Dif)≤ai by a factor of 3dai, paying 3d for the degree d rather than nd for the dimension n.
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)≤3dai3/2 over all i=1,…,n and using ai≤a=maxiai bounds the low-degree variance of f: Mlow≤i=1∑nBd(Dif)≤3di=1∑nai3/2≤3dai=1∑nai=3daI(f). Adding the high-degree bound Mhigh≤I(f)/(d+1) gives, for every integer d≥0, V=Mlow+Mhigh≤I(f)(3da+d+11). Look at the trade-off inside the parentheses as we vary the cutoff degree d: raising d pushes down the high-degree factor d+11 at the cost of growing 3da. When a is small, we can choose d on the scale of log(1/a) so that 3d absorbs only part of a (say 3d≤a−1/4, making 3da≤a1/4), while d+11 shrinks to O(1/log(1/a)).
Lemma 2 (Maximal-influence trade-off) For every nonconstant Boolean function f:{−1,1}n→{−1,1} with 0<a=maxiInfi(f)<1, I(f)≥6Var(f)loga1.
Specifically, choosing d=⌊4log3log(1/a)⌋ ensures 3d≤a−1/4, so 3da+d+11≤a1/4+d+11≤log(1/a)6 for all a∈(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} on n≥2 uniform bits, some coordinate i∈[n] satisfies Infi(f)≥121Var(f)nlogn.(The constant 1/12 is chosen for clean arithmetic rather than optimality.)
Proof. If f is constant, Var(f)=0 and the inequality is trivial. Otherwise let a=maxiInfi(f)>0. We distinguish two cases:
Case 1 (a≥n−1/2): Since Var(f)≤1 and nlogn≤n−1/2 for all n≥2, we already have a≥n−1/2≥Var(f)nlogn>121Var(f)nlogn.
Case 2 (0<a<n−1/2): Here loga1>21logn. Using na≥∑i=1nai=I(f) together with Lemma 2 gives na≥I(f)≥6Var(f)loga1>12Var(f)logn, and dividing by n yields a>121Var(f)nlogn.
For any balanced Boolean function (Var(f)=1), Theorem 5 guarantees a coordinate of influence at least 12nlogn—matching the Θ((logn)/n) benchmark of Tribes from §1.10! Even better, Theorem 5 requires no symmetry and no monotonicity, and its dependence on the bias of f enters solely through the variance Var(f)=4p(1−p). That universality is exactly what allows us to apply Theorem 5 repeatedly to every residual function fS 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>0 such that, for every balanced function f:{−1,1}n→{−1,1} with n≥2 and every 0<ε<1/2, there exists a coalition S⊆[n] of size ∣S∣≤Clognnlogε1 with a rushing strategy that forces +1 with probability p+(S)≥1−ε. In particular, for ε=0.01, a coalition of O(n/logn)=o(n) participants can force the output +1 with probability at least 0.99.
Proof. We build the coalition greedily starting from S0=∅. At step t≥0, let ht=fSt be the residual Boolean function on the rt=n−t untouched coordinates, and write pt=P(ht=+1)=p+(St),et=1−pt=P(ht=−1). Since f is balanced, p0=1/2 (e0=1/2), and adding coordinates to the coalition can only increase pt (and decrease the failure probability et). As long as et>ε and t<n/2, apply Theorem 5 to ht on its rt-dimensional cube to choose a remaining coordinate it∈/St of maximum influence, and set St+1=St∪{it}.
By the one-step gain identity (§1.9), Theorem 5, and the variance formula Var(ht)=4pt(1−pt)=4ptet, et−et+1=pt+1−pt=21Infit(ht)≥241Var(ht)rtlogrt=61ptetrtlogrt. Suppose n≥4 and t≤n/2. Then n/2≤rt≤n and pt≥p0=1/2, so rtlogrt≥nlog(n/2)≥2nlogn. Substituting pt≥1/2 and rtlogrt≥2nlogn into the display above gives et−et+1≥61⋅21et⋅2nlogn=24nlognet, and rearranging for et+1 yields the geometric decay et+1≤(1−24nlogn)et⟹et≤21exp(−24ntlogn). Now set K=⌈logn24nlog2ε1⌉. If K≤n/2, then after at most K greedy steps the right-hand side is at most ε, so et≤ε and p+(St)≥1−ε. If instead K>n/2 (or if n∈{2,3}), the coalition can simply corrupt all n coordinates and achieve p+([n])=1. In both cases, ∣S∣≤Clognnlogε1 for a universal constant C.
Finally, notice that the sequence of residual functions ht and chosen coordinates it depends only on the truth table of f, so the coalition S is completely determined before any honest bits are broadcast. Once the honest participants broadcast x[n]∖S with fS(x[n]∖S)=+1, the coalition simply outputs a witnessing assignment yS∈{−1,1}S with f(x[n]∖S,yS)=+1.
Applying Theorem 6 to −f produces a coalition of the same size that forces −1 with probability at least 1−ε (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 maxiInfi(f)=Ω((logn)/n) tight, and forcing Tribesm,w=−1 indeed requires corrupting one coordinate in almost every tribe (Ω(n/logn) coordinates). Yet to force Tribesm,w=+1, a coalition only needs a single tribe of size w=Θ(logn)! Why does greedy recruitment finish in w steps instead of Θ(n/logn)? Because as soon as the coalition sets one coordinate i1∈T1 to +1, the remaining w−1 members of T1 see their influence on the residual function f{i1}double from 2−w to 2−(w−1), then 2−(w−2), and so on—snowballing inside T1 rather than staying at Θ((logn)/n).
Let us step back and take stock of both classroom stories and the scope of what we proved:
What the logarithm did: In §1.9, Poincaré gave a per-step gain of only Θ(Var(ht)/n), which stalled at linear coalition size Θ(n). Using the noise operator T1/3 and Bonami’s 2→4 hypercontractive inequality to extract an extra power ai3/2 from rare pivotal events boosted the per-step gain by Θ(logn), accelerating the geometric decay of et by a logarithmic factor so that O(n/logn) steps suffice.
What the protocol theorem assumes: Theorem 6 is an information-theoretic existence theorem for single-round, full-information, rushing protocols. It does not claim that a small coalition can find the winning response yS in polynomial time if f is only given as a cryptographic circuit, nor does it rule out multi-round coin-flipping protocols or protocols with cryptographic commitments. What it proves is that in a single broadcast round without cryptography, no Boolean aggregation rule can prevent a sublinear O(n/logn) coalition from seizing control of the bit.
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}. A linear function has the form ℓa(u)=a1u1+⋯+anun(mod2),a∈{0,1}n. The vector a is not supplied to us. We want to test whether the program follows some linear rule, without reading all 2n 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)). Our task is a promise problem: accept every linear function, and reject with high probability whenever g has distance at least ε from every linear function. Functions between these two cases need not receive a prescribed answer.
The desired query bound should depend on ε and the failure probability, but not on 2n. Query complexity counts calls to the program; generating an n-bit query still takes time proportional to n.
A local identity and its three-query test
Every linear function obeys g(u⊕v)=g(u)+g(v)(mod2), where u⊕v is coordinatewise addition modulo two. Conversely, if the identity holds for every pair, then g(0)=0, and repeatedly applying it to the standard basis vectors gives g(u)=∑iuig(ei)(mod2). The identity characterizes linearity exactly.
This suggests a randomized test: choose independent uniform U,V, query g(U), g(V), and g(U⊕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)ui and f(x)=(−1)g(u). Addition of bits becomes multiplication of signs. Write xy=(x1y1,…,xnyn). A linear rule becomes a parity χS(x)=i∈S∏xi,S={i:ai=1}, and its local identity becomes χS(xy)=χS(x)χS(y).
Definition 6 (The BLR test) Choose independent uniform X,Y∈{−1,1}n. Query f(X), f(Y), and f(XY), and accept if f(X)f(Y)=f(XY). The test makes at most three queries. The empty parity χ∅=1 is included among the linear rules.
Recall (Pairwise vs. mutual independence). The three query points (X,Y,XY) are pairwise independent—any two are uniform and independent—but completely dependent as a triple since X⋅Y⋅(XY)=1. This is the n-dimensional cousin of the classic three-coin-event counterexample from introductory probability.
The local test statistic involves three values of f. 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: XY is determined by X,Y. Each pair among X,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=21+f(X)f(Y)f(XY). Expand the three factors in the character basis. A term with indices A,B,C has expectation E[χA(X)χB(Y)χC(XY)]=E[χA(X)χC(X)]E[χB(Y)χC(Y)]=1A=C1B=C. The first step uses the character identity and independence of X,Y; the second uses orthogonality. All terms disappear except those with A=B=C. Therefore E[f(X)f(Y)f(XY)]=S∑f(S)3, and the rejection probability δ satisfies 1−2δ=S∑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 f with probability δ, then some parity χS satisfies dist(f,χS)≤δ. Equivalently, if f is at distance at least ε from every parity, then a single test rejects with probability at least ε.
Proof. Let M=maxSf(S). Squared coefficients are nonnegative, so Parseval gives 1−2δ=S∑f(S)3≤MS∑f(S)2=M. Choose S attaining M. The correlation–distance identity yields dist(f,χS)=21−M≤δ.
The maximum here is of the coefficients themselves, not their absolute values. A negative parity −χS 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 r times and accept only if every trial accepts. A linear function is always accepted. If f is ε-far, then P(all trials accept)≤(1−ε)r≤e−εr. For failure probability 0<η<1, take r=⌈εlog(1/η)⌉. The algorithm uses at most 3r queries, independent of n. 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 f is promised to be within distance α<1/4 of an unknown parity χS. We are given a particular input x, and want the correct value χS(x) even if the table entry f(x) is corrupted.
Querying f(x) may repeatedly read the same corrupted entry. The parity identity gives another route: for any y, the correct value at x equals χS(y)χS(xy). We can choose y uniformly so that both queried locations are uniform, spreading the risk over the whole table. This suggests returning f(Y)f(xY). For the uncorrupted parity, χS(Y)χS(xY)=χS(x). Both Y and xY are uniform, so each queried entry is corrupted with probability at most α. They need not be independent; a union bound is sufficient.
Proposition 1 (Two-query local correction) If dist(f,χS)≤α, then for every fixed x, PY(f(Y)f(xY)=χS(x))≥1−2α.
The guarantee holds at every target input, with probability over the decoder’s randomness. If α<1/4, independent repetitions and a majority vote amplify it. Writing γ=1/2−2α>0, r repetitions have failure probability at most e−2rγ2 by Hoeffding. For a fixed error bound below 1/4, the number of queries needed for failure probability η is O(log(1/η)).
The table (χS(x))x∈{−1,1}n is a Hadamard codeword encoding S. 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 S. The BLR test is therefore a concrete example of local information controlling global algebraic structure.
2.3 Further Fourier expansions: monotone rules, Majn, and Tribes
In §1.3 (Example 4 and Example 5) we computed the full expansion of Maj3 and the degree-1 coefficients of Majn. Here we develop three extensions that connect those calculations to the rest of the chapter: why degree-1 Fourier coefficients equal influences for all monotone functions, how the higher odd levels of Majn 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} is monotone (or increasing) if flipping any coordinate from −1 to +1 can never decrease the output: f(x−i,+1)≥f(x−i,−1) for every i∈[n] and x−i∈{−1,1}[n]∖{i}. Majority, ANDw, ORw, and Tribes are all monotone, whereas Parity (for n≥2) is not.
Proposition 2 (Linear Fourier coefficients of a monotone Boolean function) If f:{−1,1}n→{−1,1} is monotone, then for every coordinate i∈[n], f({i})=Infi(f),i=1∑nf({i})=I(f).
Proof. Recall from §1.8 that Dif(x)=2f(x)−f(x⊕i)=∑S∋if(S)χS(x). Multiplying by xi∈{−1,1} gives xiDif(x)=2f(x−i,+1)−f(x−i,−1)=S∋i∑f(S)χS∖{i}(x). Taking expectations over X extracts the S={i} term on the right: E[XiDif(X)]=f({i}). On the other hand, because f is monotone and {−1,1}-valued, the difference 2f(x−i,+1)−f(x−i,−1) is +1 whenever coordinate i is pivotal and 0 otherwise—it is never −1! Hence xiDif(x)=1{f(x)=f(x⊕i)}=(Dif(x))2 at every input x. Taking expectations gives f({i})=E[(Dif(X))2]=Infi(f).
Notice the sharp contrast between the two formulas for Infi(f) when f is monotone: Infi(f)=f({i})=S∋i∑f(S)2. For Majn, where Infi(Majn)∼πn2, summing the unsquared degree-1 coefficients gives I(Majn)=∑i=1nMajn({i})∼π2n, whereas summing their squares gives the level-1 Fourier weight W1(Majn)=∑i=1nMajn({i})2∼π2.
Higher Fourier levels of Majn and five-voter Majority
How do we compute Majn(S) for larger odd subsets ∣S∣=k? Let us first work out n=5 explicitly, and then record the general formula.
Example 7 (Five-voter Majority Maj5 and higher odd degrees) By oddness (Maj5(−x)=−Maj5(x)) and permutation symmetry, only odd levels k∈{1,3,5} appear, and all coefficients within a level are equal: Maj5(x)=a∣S∣=1∑χS(x)+b∣S∣=3∑χS(x)+cχ[5](x). From Example 5, we already know a=Maj5({1})=2−4(24)=166=83. To find b and c, evaluate Maj5(x)=+1 on two inputs with different numbers of +1’s:
At x=(+1,+1,+1,+1,+1), every character χS(x) equals +1, so 5a+10b+c=1.
At x=(+1,+1,+1,+1,−1) (four +1’s and one −1), the singleton sum is 4−1=3, the (35)=10 triples split into (34)=4 with product +1 and (24)=6 with product −1 (sum 4−6=−2), and χ[5](x)=−1, giving 3a−2b−c=1.
Adding the two equations gives 8a+8b=2, so b=41−a=−81, and then c=1−5a−10b=83: Maj5(x)=83∣S∣=1∑χS(x)−81∣S∣=3∑χS(x)+83χ[5](x). Its spectral weights across levels k=1,3,5 are W1(Maj5)=5(83)2=6445,W3(Maj5)=10(−81)2=6410,W5(Maj5)=(83)2=649, which sum to 6445+10+9=1.
More generally, for any odd n=2m+1 and any subset S⊆[n] of odd size k=2r+1, picking an element i∈S and using xiDiMajn(x)=1{∑j=ixj=0} expresses Majn(S) as the expectation of ∏j∈S∖{i}Xj on the tie event {∑j=iXj=0}, yielding the closed form (see O’Donnell [3], Theorem 5.19): Majn(S)=(−1)r(2r2m)(rm)2−2m(m2m)(∣S∣=2r+1,n=2m+1). As n→∞, the level-k weight Wk(Majn) converges for each odd k≥1 to πk2((k−1)/2k−1)2−(k−1) (W1→π2≈0.6366, W3→3π1≈0.1061, W5→20π3≈0.0477, …). Under Kalai’s cycle formula (§1.5), the central limit theorem implies that as n→∞, the correlated pair (n1∑iXi,n1∑iYi) converges to a bivariate normal pair with correlation ρ=−1/3, so Stab−1/3(Majn)→−π2arcsin(1/3) and n→∞limP(cycle under Majn)=41−2π3arcsin(31)≈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 Tribesm,w) Recall that Tribesm,w partitions [n] into m disjoint tribes T1,…,Tm of width w, outputting −1 if and only if every tribe Tj fails to be unanimously +1. Let 1Aj(x)=∏i∈Tj21+xi=2−w∑S⊆TjχS(x) be the 0/1 indicator that tribe Tj is unanimously +1. Then the 0/1 indicator of {Tribesm,w(x)=−1} is the product ∏j=1m(1−1Aj(x)), so in {−1,1} sign notation: Tribesm,w(x)=1−2j=1∏m(1−1Aj(x)). Now substitute 1−1Aj(x)=(1−2−w)−2−w∑∅=Sj⊆TjχSj(x) into the product. Because the tribes T1,…,Tm are pairwise disjoint, any subset S⊆[n] decomposes uniquely into its intersections Sj=S∩Tj with the m tribes! Expanding the product across j=1,…,m immediately gives every Fourier coefficient:
For S=∅, picking the constant term 1−2−w from all m factors gives Tribesm,w(∅)=1−2(1−2−w)m.
For any nonempty S⊆[n], let r(S)={j∈[m]:S∩Tj=∅}≥1 be the number of tribes that S intersects. Picking −2−wχS∩Tj(x) from those r(S) tribes and the constant 1−2−w from the remaining m−r(S) tribes gives Tribesm,w(S)=−2(−2−w)r(S)(1−2−w)m−r(S).
Notice how this formula checks against our earlier calculations: when S={i} is a singleton, it intersects r({i})=1 tribe, so Tribesm,w({i})=2−(w−1)(1−2−w)m−1—which matches the exact influence Infi(Tribesm,w) from §1.10, just as Proposition 2 requires! Moreover, unlike Majority (where 2/π of the Fourier weight sits on level 1), a singleton in Tribes has squared coefficient f({i})2=Θ(n2log2n), so the total level-1 weight of Tribes is W1(Tribesm,w)=nf({1})2=Θ(nlog2n)→0. Almost all of Tribes’ variance lives around degree k=Θ(logn), which is why its total influence I(f)=∑kkWk(f)=Θ(logn) grows logarithmically with n.
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 f satisfies Var(f)≤i=1∑n∥Dif∥22. For Boolean-valued f, this says Var(f)≤I(f). In particular, a balanced Boolean function has I(f)≥1.
Proof. Every nonempty set has size at least one. Therefore Var(f)=S=∅∑f(S)2≤S∑∣S∣f(S)2=i∑∥Dif∥22. If f is balanced and Boolean, Var(f)=Ef2=1.
For the rate-one-per-coordinate generator, −⟨f,Lf⟩=∑i∥Dif∥22 and the spectral gap is one. For the discrete random-coordinate kernel, the energy is smaller by n and the gap is 1/n. Both versions give the same displayed inequality.
Theorem 9 (Equality forces a dictator) If f:{−1,1}n→{−1,1} is balanced, then I(f)=1 if and only if f=±xi for some i.
Proof. Equality in Poincaré removes every Fourier level of degree at least two. Balance removes level zero, so W1(f)=1. The level-one rigidity proved in Lemma 1 gives f=±xi. Conversely, a signed dictator has exactly one coordinate of influence one.
In geometric terms, let A={x:f(x)=1}. The number of undirected cube edges crossing from A to its complement is ∣∂A∣=2n−1I(f). Each crossing edge is counted twice by the directed pivotal events. Among sets of size 2n−1, the smallest possible edge boundary is 2n−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 i, write u=f(x−i,1),v=f(x−i,−1). The conditional mean is (u+v)/2, and the conditional variance is (u−v)2/4. Both values of Dif have squared magnitude (u−v)2/4. Thus ∥Dif∥22=E[Var(f(X)∣X−i)]. Poincaré becomes Var(f(X))≤i=1∑nE[Var(f(X)∣X−i)].
There is an equivalent resampling form. Let X(i) be obtained by replacing Xi by an independent uniform sign, with all other coordinates unchanged. Conditional on X−i, the two function values are independent and identically distributed. Therefore E[(f(X)−f(X(i)))2]=2E[Var(f(X)∣X−i)]. We obtain the Efron–Stein form on the cube: Var(f(X))≤21i∑E[(f(X)−f(X(i)))2]. For Boolean-valued f, a resampling changes the output with probability Infi(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 X be uniform and Y its ρ-correlated noisy version. Both marginals are uniform. We use 0≤ρ≤1 for the random-perturbation comparisons below.
Recall the noise stability Stabρ(f)=E[f(X)f(Y)] defined in the Arrow argument.
By orthogonality, Stabρ(f)=S∑ρ∣S∣f(S)2. For a Boolean-valued function, f(X)f(Y) is 1 when the two outputs agree and −1 otherwise, so P(f(X)=f(Y))=21−Stabρ(f). 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 Stabρ(x1)Stabρ(χ[n])Stabρ(Maj3)=ρ,=ρn,=43ρ+41ρ3. For fixed ρ<1, parity loses its correlation as n 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: dρdStabρ(f)ρ=1=I(f). Influence controls the initial response to small noise; stability at a fixed ρ depends on the whole distribution of Fourier degrees.
Connecting the clocks
Let P be the discrete chain that refreshes one uniformly chosen coordinate. A character χS survives an update if the selected coordinate is outside S, and otherwise its conditional mean becomes zero. Hence PχS=(1−n∣S∣)χS. For the continuous process that refreshes each coordinate at rate one, use L=n(P−I)=−i=1∑nDi. Then LχS=−∣S∣χS, and etL=Te−t. This clock is n times faster than the total-rate-one Poissonization et(P−I) in Chapter 2. The latter is Te−t/n. Stating the clock avoids losing a factor of n in the spectral gap or Dirichlet form.
2.7 Spectral tails and small-noise bounds
For a Boolean function, define its degree-d truncation by f≤d=∣S∣≤d∑f(S)χS. Since every discarded level has degree at least d+1, ∥f−f≤d∥22=∣S∣>d∑f(S)2≤d+1I(f). 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. 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))=21S∑(1−ρ∣S∣)f(S)2≤21−ρI(f), where 1−ρk≤k(1−ρ). If δ=(1−ρ)/2 is the bit-flip probability, the bound is δI(f).
The first derivative at zero noise is determined by influence: δ↓0limδP(f(X)=f(Y))=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 n.
2.8 The degree cutoff and its constants
Proof: choosing the degree cutoff
Put L=log(1/a)>0 and take d=⌊L/(4log3)⌋. Then 3da≤e−L/4,d+11≤L4log3. Since supL>0Le−L/4=4/e, e−L/4+L4log3≤L4/e+4log3<L6. Substitute into the low/high-degree estimate and rearrange. The argument includes d=0; there is no need to assume a is very small.
2.9 Tribes parameters and exact finite examples
In §1.10 we defined the Tribes function on n=mw coordinates (m disjoint tribes of width w) and found the exact formulas P(f(X)=−1)=(1−2−w)m,Infi(f)=2−(w−1)(1−2−w)m−1. Here we record the exact asymptotic calculation showing that Infi(f)=Θ((logn)/n) along a sequence of nearly balanced Tribes functions, together with a concrete finite example.
For each integer width w≥1, choose the number of tribes to be mw=⌊(log2)2w⌋, and set nw=mww. Because log(1−2−w)=−2−w+O(4−w) and mw=(log2)2w+O(1), we have mwlog(1−2−w)=−log2+O(2−w),(1−2−w)mw=21+O(2−w). Consequently E[f(X)]=O(2−w)→0 and Var(f)=1−(Ef)2=1−O(4−w)→1 as w→∞. (Because mw and w must be integers, (1−2−w)mw cannot equal 1/2 exactly, which is why Theorem 5’s dependence on Var(f) rather than exact balance Ef=0 is so convenient.)
Now express the individual influence in terms of n=nw. Since (1−2−w)mw−1→1/2, we have Infi(f)=2−(w−1)(1−2−w)mw−1∼2−w. Meanwhile, n=mww∼(log2)w2w. Taking base-2 logarithms gives log2n=w+log2w+log2log2+o(1)⟹w=log2n−log2log2n−log2log2+o(1), and substituting 2−w∼n(log2)w∼nlogn yields the sharp asymptotics Infi(f)∼nlogn,I(f)=nInfi(f)∼logn.
Example 10 (A concrete 8-coordinate example) Take m=4 tribes of width w=2, so n=8. Each tribe of size 2 is unanimously +1 with probability 1/4, so P(f(X)=+1)=1−(43)4=256175≈0.684, and every coordinate has influence Infi(f)=2−1(43)3=12827≈0.211. This small example illustrates that the exact probability and influence formulas hold for every (m,w), while approximate balance only emerges once m is chosen near (log2)2w.
Finally, notice the striking asymmetry between forcing +1 and forcing −1 on Tribes:
To force +1, a rushing coalition only needs to corrupt one tribe of size w=Θ(logn) and set all its bits to +1.
To force −1, a coalition must prevent every tribe from becoming unanimously +1. A deterministic coalition that guarantees −1 on every input must place at least one member in each of the m tribes, requiring m=Θ(n/logn) participants! (Even a rushing coalition that only aims for success probability 1−ε=0.99 on random honest inputs must still touch a (1−log2log(1/(1−ε)))≈98.55% fraction of the m tribes, requiring Θ(n/logn) participants.)
2.10 One coalition that can choose either outcome
For a coalition S, define its probability of full control by q(S)=PX[n]∖S(f(X[n]∖S,⋅) takes both signs). 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+ and S− be the coalitions obtained by applying the classroom theorem to f and −f, with error ε/2. Put S=S+∪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+) and p−(S)≥p−(S−), with each probability evaluated on its corresponding remaining uniform cube.
On the remaining cube for S, let E+ and E− be the events that the two target signs are achievable. Their probabilities are each at least 1−ε/2. A union bound gives q(S)=P(E+∩E−)≥1−ε,∣S∣=O(lognnlogε1). 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 for the positive target and all −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 g has degree at most d, define h by h(S)=3∣S∣/2g(S). Then T1/3h=g, so ∥g∥4≤∥h∥2≤3d/2∥g∥2. Low-degree polynomials cannot concentrate their entire L2 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.
Proof of the finite-space Hölder inequality
If either norm is zero, the inequality is immediate. Otherwise set a=∣h∣/∥h∥4/3 and b=∣k∣/∥k∥4, so Ea4/3=Eb4=1. For nonnegative a,b, Young’s inequality in the required exponents is ab≤43a4/3+41b4. To verify it directly, put t=a1/3. The difference between right and left sides factors as 4b4+3t4−4t3b=4(b−t)2(b2+2bt+3t2)≥0. Averaging gives E[ab]≤1, 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<∞ and 0≤ρ≤q−1p−1, every real-valued function satisfies ∥Tρg∥q≤∥g∥p.
We quote this full statement; the special p=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 g has degree at most d and q≥2, rescale its Fourier coefficients as in the degree-form Bonami argument to get ∥g∥q≤(q−1)d/2∥g∥2. This converts a degree bound into moment bounds. For d≥1, normalize ∥g∥2=1, take t≥2d/2, and choose the real exponent q=1+t2/d/2≥2. Markov’s inequality gives P(∣g∣≥t)≤(t(q−1)d/2)q=2−dq/2≤exp(−4dlog2t2/d). For degree one this is Gaussian-type decay; higher-degree polynomials have a correspondingly weaker tail scale.
Second, consider a small set A of uniform mass μ. Let X,Y be ρ-correlated. Self-adjointness and the semigroup identity give P(X∈A,Y∈A)=⟨1A,Tρ1A⟩=∥Tρ1A∥22. For 0<ρ<1, use hypercontractivity with input exponent 1+ρ and output exponent 2: P(X∈A,Y∈A)≤μ2/(1+ρ). Thus P(Y∈A∣X∈A)≤μ(1−ρ)/(1+ρ) when μ>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−t use the rate-one-per-coordinate clock. From Chapter 2, dtdVar(Ptf)=−2i∑∥DiPtf∥22. Because Ptf→Ef, integration gives the exact identity Var(f)=2∫0∞i∑∥DiPtf∥22dt. To expose the decay associated with coordinate i, write hi(x−i)=xiDif(x)=S∋i∑f(S)χS∖{i}(x). This function does not depend on xi. We obtain DiTρf=ρxiTρhi, where the operator on the right acts on the other n−1 coordinates. Changing variables ρ=e−t yields Var(f)=i∑∫012ρ∥Tρhi∥22dρ. The factor ρ must not be lost when switching derivative conventions.
The general hypercontractive theorem gives ∥Tρhi∥2≤∥hi∥1+ρ2. Since ∣hi∣ is an indicator of mass ai, Var(f)≤i∑∫01ai2/(1+u)du,u=ρ2. If 0<ai<1, let Li=log(1/ai). Because 2/(1+u)≥1+(1−u)/2 on [0,1], ∫01ai2/(1+u)du≤ai∫01e−Li(1−u)/2du≤aimin{1,2/Li}. The elementary comparison min{1,2/L}≤3/(1+L) now gives Var(f)≤3i∑1+log(1/ai)ai. The summand is defined as zero at ai=0; at ai=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 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−t is equivalent to a single differential inequality at t=0—the logarithmic Sobolev inequality—which replaces the variance Var(f) in Poincaré’s inequality by an entropy functionalEnt(f2) 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,∞) (with the convention 0log0=0 and 0log(0/0)=0), the entropy of g with respect to the uniform measure on {−1,1}n is Ent(g):=E[g(X)logE[g(X)]g(X)]=E[g(X)logg(X)]−E[g(X)]logE[g(X)].
By Jensen’s inequality applied to the strictly convex function ϕ(u)=ulogu on [0,∞), we always have Ent(g)≥0, with equality if and only if g is constant. Moreover, when E[g(X)]=1, g is the likelihood ratio (density) of a probability distribution ν(x)=2−ng(x) relative to the uniform distribution μ(x)=2−n, and Ent(g)=DKL(ν∥μ) is simply the Kullback–Leibler divergence from μ to ν (a central object in Chapters 4 and 5). For general g≥0 with E[g]>0, homogeneity Ent(cg)=cEnt(g) means that Ent(g)=E[g]DKL(νg∥μ) measures how nonuniformly the mass of g is concentrated across the cube.
Gross’s differentiation bridge: hypercontractivity at t=0
Recall from §2.6 and §2.12 that Pt=Te−t=etL is the continuous-time Markov semigroup with generator L=−∑i=1nDi. In the general hypercontractive theorem (Theorem 10), fix the input exponent p=2 and set ρ=e−t=q(t)−11, so that the output exponent grows with time t≥0 according to q(t)=1+e2t(q(0)=2,q′(0)=2). Then hypercontractivity asserts that for every real-valued function f, the log-norm curve Λ(t):=log∥Ptf∥q(t)=q(t)1logE[∣Ptf(X)∣q(t)] satisfies Λ(t)≤Λ(0)=log∥f∥2 for all t≥0. Because Λ(t) attains its maximum at t=0, its right derivative at t=0 must satisfy Λ′(0)≤0!
Let us compute Λ′(0) for a strictly positive function f>0 (the general case f:{−1,1}n→R follows by applying the result to f2+ε and letting ε↓0). Write ut=Ptf>0 and M(t)=E[utq(t)]. Differentiating Λ(t)=q(t)1logM(t) by the product and chain rules gives Λ′(t)=−q(t)2q′(t)logM(t)+q(t)M(t)1E[utq(t)(q′(t)logut+q(t)ut∂tut)]. At t=0, we have q(0)=2, q′(0)=2, u0=f, M(0)=E[f2], and ∂tut∣t=0=Lf=−∑i=1nDif. Using 2logf=log(f2) and the Dirichlet-form identity −⟨f,Lf⟩=∑i=1n⟨f,Dif⟩=∑i=1n∥Dif∥22=E(f,f), we obtain Λ′(0)=2E[f2]1(E[f2log(f2)]−E[f2]logE[f2]+2⟨f,Lf⟩)=2E[f2]1(Ent(f2)−2i=1∑n∥Dif∥22). Therefore, the infinitesimal condition Λ′(0)≤0 is identical to the inequality Ent(f2)≤2∑i=1n∥Dif∥22! Conversely, Gross (1975) showed that applying this single inequality at every time t≥0 to the rescaled function ft=∣Ptf∣q(t)/2 implies Λ′(t)≤0 for all t≥0, integrating back to the full (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 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], let Ei[g](x)=2g(x−i,+1)+g(x−i,−1) denote the expectation over Xi with the remaining n−1 coordinates X−i held fixed, and define the conditional one-coordinate entropy Enti(g):=Ei[glogEi[g]g].
Lemma 3 (Entropy tensorization (subadditivity) on the cube) For every nonnegative function g:{−1,1}n→[0,∞), Ent(g)≤i=1∑nE[Enti(g)].
Proof: telescoping via the chain rule and joint convexity
By approximation (g+ε with ε↓0), we may assume g>0. For 0≤k≤n, let gk=E[g∣X1,…,Xk] be the conditional expectation given the first k coordinates, so gn=g and g0=E[g]. Writing logE[g]g=∑i=1nloggi−1gi and using the tower property E[gloggi−1gi]=E[giloggi−1gi] decomposes the total entropy into a sum of one-step increments: Ent(g)=i=1∑nE[giloggi−1gi]. Now fix i∈[n] and let E>i denote averaging over (Xi+1,…,Xn) conditioned on (X1,…,Xi). By independence of the coordinates, gi=E>i[g] and gi−1=Ei[gi]=E>i[Eig]. Because the two-variable relative-entropy function Ψ(u,v)=ulog(u/v) is jointly convex on (0,∞)2 (its Hessian (1/u−1/v−1/vu/v2) is positive semidefinite), Jensen’s inequality applied to the average E>i gives giloggi−1gi=Ψ(E>i[g],E>i[Eig])≤E>i[Ψ(g,Eig)]. Taking full expectations and using the tower property E[Ψ(g,Eig)]=E[Ei[Ψ(g,Eig)]]=E[Enti(g)] across i=1,…,n yields the claim.
Theorem 11 (Logarithmic Sobolev inequality on the uniform cube (Gross)) For every real-valued function f:{−1,1}n→R, Ent(f2)≤2i=1∑n∥Dif∥22=2E(f,f).
Proof. By Lemma 3 applied to g=f2, it suffices to prove the one-bit inequality Enti(f2)≤2Ei[(Dif)2] for each fixed assignment of x−i. Write a=f(x−i,+1) and b=f(x−i,−1); then Ei[(Dif)2]=(2a−b)2, so we must show that for all a,b∈R, 2a2loga2+b22a2+2b2loga2+b22b2≤2(a−b)2. Since ∣∣a∣−∣b∣∣≤∣a−b∣, replacing (a,b) with (∣a∣,∣b∣) leaves the left-hand side unchanged and can only decrease the right-hand side, so we may assume a,b≥0. If a=b=0 both sides vanish; otherwise, by symmetry and homogeneity of degree 2, we may assume a≥b≥0 and normalize a+b=2, writing a=1+t and b=1−t with t∈[0,1]. Then 2(a−b)2=2t2 and 2a2+b2=1+t2, so the left-hand side becomes φ(t):=(1+t)2log(1+t)+(1−t)2log(1−t)−(1+t2)log(1+t2). For t∈(0,1), differentiating twice with respect to t gives φ′(t)=2(1+t)log(1+t)−2(1−t)log(1−t)−2tlog(1+t2),φ′′(t)=2log1+t21−t2+1+t24≤4, since log1+t21−t2≤0 and 1+t24≤4. Because φ(0)=φ′(0)=0, integrating φ′′(t)≤4 twice on [0,t] (and extending to t=1 by continuity) yields φ(t)≤2t2 for all t∈[0,1]. Averaging over x−i and summing over i=1,…,n completes the proof.
Notice that in the one-bit proof above, φ′′(0)=4=(2t2)′′∣t=0: for small perturbations f=1+εh around a constant function (with E[h]=0), expanding Ent((1+εh)2)=2ε2Var(h)+O(ε3) and 2E(1+εh,1+εh)=2ε2E(h,h) shows that the Poincaré inequality Var(h)≤E(h,h) is simply the second-order linearization of the logarithmic Sobolev inequality around constant functions!
Small subcubes of codimension k (α=2−k, out-degree k vs. bound (ln2)k)
Geometric intuition: small-set expansion and the 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 0/1 indicator f=1A of a small set A⊆{−1,1}n of volume α=P(X∈A)=2−k:
Boundary out-degree of small sets: Along coordinate i, the derivative Di1A(x)=21A(x)−1A(x⊕i) equals ±21 on boundary edges crossing between A and Ac, and 0 elsewhere. Hence ∥Di1A∥22=41P(1A(X)=1A(X⊕i)), and summing over i=1,…,n relates the Dirichlet energy directly to the size ∣∂A∣ of the undirected edge boundary between A and Ac: E(1A,1A)=i=1∑n∥Di1A∥22=21⋅2n∣∂A∣=2α⋅∣A∣∣∂A∣, where ∣A∣∣∂A∣ is the average boundary out-degree of a vertex in A (the average number of coordinates i∈[n] whose flip exits A).
What Poincaré sees: Because Var(1A)=α(1−α), Poincaré gives α(1−α)≤2α∣A∣∣∂A∣, or ∣A∣∣∂A∣≥2(1−α). Even when α=2−k is exponentially small, Poincaré only guarantees a constant O(1) average out-degree, because it is calibrated to half-cubes (α=1/2).
What Log-Sobolev sees: Evaluating the entropy of 1A2=1A gives Ent(1A2)=E[1Alogα1A]=αlogα1. Plugging this into Theorem 11 yields the small-set isoperimetric inequality on the cube: αlogα1≤2E(1A,1A)=α⋅∣A∣∣∂A∣⟹∣A∣∣∂A∣≥logα1=(ln2)k. Every subset A⊆{−1,1}n of volume α=2−k has average boundary out-degree at least (ln2)k=Ω(log(1/α))! Geometrically, the most compact set of volume 2−k in the hypercube is a codimension-k subcube (fixing k coordinates), where every vertex in A has exactly k=log2(1/α) outgoing edges to Ac.
Where the log(1/a) factor in KKL comes from: Return to a Boolean function f:{−1,1}n→{−1,1} whose discrete derivatives hi=Dif∈{−1,0,1} have small influences ai=Infi(f)=P(hi=0)≤a≪1. Because hi2=1Ai is the 0/1 indicator of the rare pivotal set Ai of volume ai, its entropy is Ent((Dif)2)=ailogai1. Applying Theorem 11 to Dif (where DjDif=∑S⊇{i,j}f(S)χS for all i,j∈[n], including j=i) and summing over i=1,…,n gives loga1⋅I(f)≤i=1∑nailogai1=i=1∑nEnt((Dif)2)≤2i=1∑nj=1∑n∥DjDif∥22=2S⊆[n]∑∣S∣2f(S)2 (and applying Theorem 11 to the xi-free derivative xiDif=∑S∋if(S)χS∖{i} drops the j=i term, tightening ∣S∣2 to ∣S∣(∣S∣−1)). This immediately shows that small individual influences force the quadratic degree moment ∑S∣S∣2f(S)2 to exceed 21log(1/a)I(f)! In §2.12, integrating hypercontractivity along the semigroup Pt=Te−t converted this exact log(1/ai) small-set entropy gain on hi into the variance bound Var(f)≤3∑i1+log(1/ai)ai.
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} with independent Ber(p) laws. Let g:{0,1}n→{0,1} be increasing in every coordinate, and put θ(p)=Epg. This is the probability of the increasing event {g=1} as the input density varies.
Theorem 12 (Russo–Margulis formula) For 0<p<1, θ′(p)=i=1∑nInfi,p(g), where Infi,p(g) is the probability, under the other n−1 biased bits, that g(X−i,1)=g(X−i,0).
Proof. Give the coordinates separate biases p1,…,pn and let F(p1,…,pn) be the resulting expectation of g. Conditional on all bits except i, E[g∣X−i]=pig(X−i,1)+(1−pi)g(X−i,0). Differentiating with respect to pi gives ∂pi∂F=E[g(X−i,1)−g(X−i,0)]. For increasing Boolean g, the difference is either zero or one, and equals the pivotal indicator. Apply the chain rule to θ(p)=F(p,…,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=1/2; it does not by itself prove a general sharp-threshold theorem.
For example, if g is the AND of n bits, then θ(p)=pn and every biased influence is pn−1. The identity gives θ′(p)=npn−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 C such that if a Boolean function satisfies ∣S∣≥2∑f(S)2≤ε, then it differs from a constant function or a signed dictator on at most Cε 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, the stability formula implies W1(f)≥1−9ε/2. The degree-one form of FKN then gives O(ε) 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] lets us estimate a specified coefficient by sampling. There are exponentially many possible subsets S. 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, the Goldreich–Levin algorithm can, with high probability, produce a polynomial-size list containing every S with ∣f(S)∣≥τ, in time polynomial in n, 1/τ, and the logarithm of the inverse failure probability. Parseval already bounds the number of such coefficients by 1/τ2; 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 f is Boolean and 0<ε<1, there is a Boolean function g depending on at most exp(O(I(f)/ε)) coordinates such that P(f=g)≤ε. Such a function g 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 p varies.
Main result. Fix 0<δ<1/2 and 0<ε<1/2. For nonconstant increasing Boolean functions invariant under a transitive permutation group on n coordinates, any transition from probability ε to 1−ε lying in p∈[δ,1−δ] has width Oδ(lognlog(1/ε)). This is an interior-parameter consequence of the Friedgut–Kalai sharp-threshold framework. Thresholds approaching p=0 or 1 require the corresponding biased estimates with their explicit dependence on p.
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 n equal influences have order (logn)/n.
Main result. Let fn be a sequence of Boolean functions on uniform cubes. If i∑Infi(fn)2⟶0, then for every fixed 0≤ρ<1, with X,Y a ρ-correlated pair, E[fn(X)fn(Y)]−(Efn)2⟶0. 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 1/2 under any fixed positive noise level.
3.6 Logarithmic Sobolev inequalities and Markov chain mixing
Background. In Chapter 2 (§1.13–1.14), the spectral gap γ (or Poincaré inequality Varπ(f)≤γ1E(f,f)) of a lazy, irreducible reversible Markov chain yielded the general L2(π) mixing-time upper bound τmix(ε)≤⌈γ1(21logπmin1+log2ε1)⌉,πmin=x∈Ωminπ(x). On the n-dimensional hypercube under the single-coordinate-per-step walk (γ=1/n and πmin=2−n), this generic spectral-gap bound gave O(n2), missing the true Θ(nlogn) mixing time by a factor of Θ(n/logn). As we saw in §2.13, the reason for that loss is geometric: at t=0, a walk started at a deterministic vertex x0 has relative density h0(x)=1{x=x0}/π(x0) supported on a tiny set of stationary measure π(x0)=2−n, and Poincaré fails to capture the rapid small-set expansion of rare spikes.
Main result. Let P be an irreducible reversible transition matrix on a finite state space Ω with stationary distribution π, continuous-time generator L=P−I, and Dirichlet form E(f,f)=⟨f,(I−P)f⟩π. Define the logarithmic Sobolev constantαLSI as the largest constant α>0 such that every f:Ω→R satisfies Entπ(f2):=Eπ[f2logEπ[f2]f2]≤α2E(f,f). (Linearizing around f=1+δg gives αLSI≤γ; for the single-coordinate-per-step walk on {−1,1}n, rescaling Theorem 11 by 1/n gives the matching lower bound αLSI≥1/n=γ.) By Gross’s equivalence, αLSI governs the hypercontractive smoothing of etL from L2(π) into Lq(t)(π) (q(t)=1+e2αLSIt) and replaces the initial-state factor log(1/πmin) by its double logarithm: for continuous-time mixing (and similarly for lazy discrete-time chains), τmix(ε)=O(αLSI1loglogπmin1+γ1logε1). For the random walk on {−1,1}n, where αLSI=γ=1/n and loglog(1/πmin)=log(nlog2)=Θ(logn), this bound recovers the sharp O(nlogn) 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.
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.
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.