Real Analysis in Computer Science:

A collection of Open Problems

Yuval Filmus, Hamed Hatami, Steven Heilman, Elchanan Mossel,
Ryan O’Donnell, Sushant Sachdeva, Andrew Wan, and Karl Wimmer∗

December 8, 2014

Abstract
We list a collection of open problems in real analysis in computer science, which complements,
updates and extends a previous list curated by Ryan O’Donnell (2012). The object of study in
these problems are boolean functions f : {0, 1}n → {0, 1}, and their analytic and combinatorial
properties. Many of the questions originate from in theoretical computer science or the theory
of voting. The formulation of many of the questions has a strong combinatorial and analytical
flavor including the use of the discrete Fourier expansion.

Contents
1 Introduction 2

2 Notations and Definitions 2

3 Boolean functions 6

4 Fourier analysis on the symmetric group 15

5 Real-Valued Functions 16

6 Extremal and additive combinatorics 16

7 Fourier analysis and noise sensitivity 19

8 Computational hardness 27
∗
The authors attended the Real Analysis in Computer Science program at the Simons Institute for the Theory of
Computing, and they thank the Institute for its hospitality.

1
1 Introduction
Fourier analysis of boolean functions is an active area of research which brings together mathe-
matical analysis, theoretical computer science and the theory of voting. During the 2012 Simons
symposium on the area, O’Donnell [O’D12] curated a list of open problems in the area. The current
collection curated by the authors while attending the special semester on real analysis in computer
science at the Simons Institute during the fall of 2013, includes additional open problems as well
as update on the status of some of the open problems presented in [O’D12].
Common notations and definitions are listed in Section 2, which is followed by several sections
listing open problems organized thematically. For general background we refer to reader to the
recent book [O’D14].

2 Notations and Definitions

We use the notation [n] = {1, . . . , n}. We denote the cardinality of a set S by |S|. The finite field
with p points is denoted Fp . The support of a function f , denoted Supp(f ), is the set of points in
the domain on which f is non-zero. The notation N (µ, σ 2 ) denotes the normal distribution with
mean µ and variance σ 2 .

2.1 Analysis on the boolean cube

Boolean cube. The boolean cube {0, 1}n is endowed with the uniform measure. The inner
product between two functions f, g : {0, 1}n → R is given by hf, gi = Ex∈{0,1}n f (x)g(x), and the L2
p
norm is given by kf k = kf k2 = hf, f i. Other norms are given by kf kp = (Ex∈{0,1}n |f (x)|p )1/p .

Fourier expansion. A boolean function on n inputs is a function f : {0, 1}n → {0, 1}. It is
often prudent to consider more generally real-valued Q functions f : {0, 1}n → R. The Fourier basis
x
consists of the functions χS (x1 , . . . , xn ) = (−1) i∈S . (Sometimes we use the notation WS instead
i

of χS .) The Fourier basis is an orthonormal basis of the space of all functions from {0, 1}n to R,
and so every function f : {0, 1}n → R has a unique Fourier expansion
X
f= fˆ(S)χS , where fˆ(S) = hf, χS i = E f (x)χS (x).
x∈{0,1}n
S⊆[n]

Parseval’s identity states that X

E [f (x)2 ] = fˆ(S)2 .
x∈{0,1}n
S⊆[n]

In particular, X
V [f ] = fˆ(S)2 .
x∈{0,1}n
S6=∅

Functions given by polynomials. Every boolean function can be expressed uniquely as a

multilinear polynomial. A polynomial of degree d is a boolean function given by a multilinear
polynomial of degree d.

2
Influence. Let f : {0, 1}n → {0, 1} be a boolean function. The influence of the ith variable is

Inf i (f ) := Pr [f (x) 6= f (∆i x)],

x∈{0,1}n

where ∆i x is the vector differing from x in the ith coordinate only. In the literature sometimes the
alternative notation Ii (f ) is used. Direct calculation gives an equivalent spectral definition:
X
Inf i (f ) = 4 fˆ(S)2 ,
i∈S

and this definition makes sense for arbitrary real-valued functions f : {0, 1}n → R. Going back full
circle, we discover the general spatial definition

Inf i (f ) := E [(f (x) − f (∆i x))2 ].

x∈{0,1}n

The total influence (or average sensitivity) of a function f : {0, 1}n → {0, 1} is the sum of the
individual influences:
n
X
Inf(f ) := Inf i (f ).
i=1

The matching spectral expression is

X
Inf(f ) = |S|fˆ(S)2 .
S⊆[n]

Poincaré’s inequality Inf(f ) ≥ 4 V(f ) easily follows.

Some authors prefer defining the influence without the constant factor 4. The spatial definition
then has to be adjusted accordingly.

Smoothing. For ρ ∈ [−1, 1], the smoothing operator (also known as the noise operator or the
Ornstein–Uhlenbeck operator) Tρ , operating on functions f : {0, 1}n → R, is given spectrally by
X
(Tρ f )(x) := ρ|S| fˆ(S)χS (x).
S⊆[n]

The spatial definition is slightly more complicated. For x ∈ {0, 1}n , let xρ ∈ {0, 1}n be obtained
by independently perturbing each coordinate so that the correlation coefficient between xi and xρ,i
is ρ. Then
(Tρ f )(x) = E f (xρ ).
xρ

The noise operators form a semigroup: Tα Tβ = Tαβ . For this reason, sometimes the parametriza-
tion Tρ0 = Teρ is used: it makes Tα0 Tβ0 = Tα+β
0 .

3
Replacing {0, 1} with {−1, 1}. So far we have considered functions with domain {0, 1}n . Some-
times it is nicer to consider the domain {−1, 1}n ; here +1
Q corresponds to 0 and −1 corresponds to 1.
The Fourier characters then become χS (x1 , . . . , xn ) = i∈S xi . All other definitions are the same.
The spatial definition of the smoothing operator becomes simpler: xρ is obtained by independently
perturbing each coordinate of x so that E[xi xρ,i ] = ρ.
Similarly, sometimes we want to consider boolean functions whose range is {−1, 1}, using the
same correspondence. The spatial definition of influence has to be changed slightly:

Inf i (f ) := 4 Pr [f (x) 6= f (∆i x)].

x∈{0,1}n

Some authors prefer keeping the original spatial definition (without the factor 4), and in that case
the spectral definition becomes simply
X
Inf 0i (f ) = fˆ(S)2 .
i∈S

Biased measures. Our definitions used the uniform inner product hf, gi = Ex∈{0,1}n f (x)g(x).
It is sometimes advantageous to consider a skewed distribution on {0, 1}n instead. For 0 ≤ p ≤ 1,
the p-biased measure, denoted µp or µnp , is given by µp (x) = p|x| (1 − p)n−|x| . All notions considered
so far can be extended to the biased case; see O’Donnell’s book [O’D14].

2.2 Analysis on [m]n

The Fourier basis is just the basis of characters of the abelian group Zn2 . We can similarly define
the Fourier expansionPof a function f : [m]n → R by taking the characters of Znm . The Fourier basis
n
is χS (x1 , . . . , xn ) = ω i=1 Si xi , where S ∈ [m]n and ω is a primitive mth root of unity.
There are several different extensions of the notion of influence; since the notion doesn’t come
up in the problems, we do not define any of them here.
The noise operator is given by
X
Tρ f := ρ|S| fˆ(S)χS .
S∈[m]n

The corresponding spatial definition is similar to the previous one: given ρ ∈ [−1, 1] and a vector
x ∈ [m]n , define a vector xρ ∈ [m]n , define xρ,i = xi with probability (1 + ρ)/2, and choose xρ,i
randomly from the values different from xi with probability (1 − ρ)/2. Then

Tρ f = E [f (xρ )].
xρ

2.3 Analysis on Rn
The invariance principle allows us to reason about functions on the boolean cube using function
on Rn . All the notions we have considered can be extended to this case as well, though we will
only describe how to extend the Ornstein–Uhlenbeck operator. For a function f : Rn → R and
ρ ∈ [−1, 1], define p
(Tρ f )(x) := E [f (ρx + 1 − ρ2 y)].
y∼N (0n ,In )

4
Here N (0n , In ) is the multivariate normal distribution of dimension n. In other words, each co-
ordinate of y is an independent normally distributed random variable with zero mean and unit
variance.

2.4 Computational complexity

Circuits. Boolean circuits on n bits are boolean functions {0, 1}n → {0, 1} specified by labelled
DAGs (directed acyclic graphs) with a unique sink (node with out-degree 0) of the following form:

• Each leaf is labelled with one of 0, 1, x1 , . . . , xn .

• Each internal node is labelled by one of NOT, AND, OR; the label NOT can only be used if
the in-degree is 1.

We define inductively the value of a node in the circuit in the obvious way:

• A leaf has the value of the constant or variable which it is labelled by.

• A node labelled NOT has the value 1 if its child has the value 0, and it has the value 0 if its
child has the value 1.

• A node labelled AND has the 1 if all its children have value 1, and the value 0 otherwise.

• A node labelled OR has the 0 if all its children have value 0, and the value 1 otherwise.

The function computed by the circuit is the value of the unique sink. Nodes are usually known
as gates, and the sink is known as the root. If there are no NOT gates, the circuit is monotone,
and one can check that it computes a monotone function (a monotone function is one satisfying
f (x) ≤ f (y) for x ≤ y). If the out-degree of every node is at most 1, the circuit is a formula.
The size of the circuit is the number of nodes. The depth of the circuit is the maximal length of a
root-to-leaf path (here length is the number of edges; some authors prefer to count vertices). What
we have described is circuits with unbounded fan-in (the fan-in of a node is its in-degree). Some
circuit classes are defined with a restriction on the fan-in, but we will not need them here.

Circuit classes. The class AC0 consists of sequences of boolean functions fn : {0, 1}n → {0, 1}
such that fn can be computed by a circuit of size nO(1) and depth O(1). In the computation
complexity literature, this class is known more accurately as non-uniform AC0 or AC0 /poly.
The class TC0 is defined similarly, in terms of threshold circuits, which are circuits with the
addition
P of a new type of gate, a threshold gate: this gate computes the indicator function of
i αi xi ≥ T , where xi are the inputs and αi , T are parameters. If all αi are non-negative then
the threshold gate is monotone. A threshold circuit is monotone if there are no NOT gates and all
threshold gates are monotone.

Decision trees. Decision trees are a different model for defining boolean functions. A decision
tree on n bits is a tree in which each internal node has a exactly two children, a 0-child and a 1-
child. Each node is labelled by a variable, and each leaf is labelled by a binary value. The decision
tree gives instructions for computing a function: Starting with the root, for each internal node,
computation proceeds to either the 0-child or the 1-child according to the value of the variable that

5
labels the node. The computation eventually reaches a leaf, which is the value computed by the
decision tree. The depth of the decision tree is the length of the maximum root-to-leaf path.
A randomized decision tree is a probability distribution over decision trees. The worst-case
depth of a randomized decision tree is the maximum over all inputs of the expected length of the
root-to-leaf computed by a random decision tree chosen according to the given distribution.

3 Boolean functions
KKL inequality under K-wise independence
Statement: Suppose f : {0, 1}n → {0, 1} satisfies fb(S) = 0 for every S ⊆ [n] with |S| ≤ Kn ,
where limn→∞ Kn = ∞. Is it true that maxi∈[n] Inf i (f ) = ω( logn n )? In other words, is it true that
maxi∈[n] Inf i (f ) ≥ ωn log
n
n
for some ωn with limn→∞ ωn = ∞?
Source: Hamed Hatami.
Remarks: A counter example based on the tribes function and good codes is provided in [HMO14].

Harper’s inequality under K-wise independence

Statement: Suppose f : {0, 1}n → {0, 1} satisfies fb(S) = 0 for every S ⊆ {1, . . . , n} with 1 ≤ |S| ≤
Kn where limn→∞ Kn = ∞. PnLet E be expectation with1 respect to the uniform probability measure
n
on {0, 1} . Is it true that i=1 Inf i (f ) ≥ ωn E[f ] log E[f ] for some ωn with limn→∞ ωn = ∞?
Source: Gil Kalai

A counter example based on the tribes function and good codes is provided in [HMO14].
Largest influence and the degree
Statement: What is the smallest possible value for maxi∈[n] Inf i (f ) for a balanced (i.e. E[f ] = 1/2)
boolean function f : {0, 1}n → {0, 1} with fb(S) = 0 for all |S| > d, S ⊂ {1, . . . , n}?
Source: Hamed Hatami
Remarks:
• By a result of Nisan and Smolensky (See [BdW02]) f has a decision tree of depth O(d4 ), and
then a result of O’Donnell, Saks, Schramm, and Servedio [OSSS05] implies maxi∈[n] Inf i (f ) ≥
Ω(d−4 ).

Small influences and influential tribes

Statement: For every c > 0 and  > 0 there exists b > 0 such that, if an increasing f : {0, 1}n →
{0, 1} and for all i ∈ {1, . . . , n}, if Inf i (f ) ≤ c log(n)/n, then there exists a small “tribe”, i.e., a set J
of variables with |J| ≤ b log(n), such that either E[f (x)|xJ = ~1] ≥ 1−o(1) or E[f (x)|xJ = ~0] ≤ o(1)?
Source: Friedgut [Fri04].
Remarks:
• One can also formulate a general form of the conjecture without the increasing assumption
(either maxy∈{0,1}J E[f (x)|xJ = y] ≥ 1 − o(1) or miny∈{0,1}J E[f |xJ = y] ≤ o(1).)

Influential tribes for continuous cube

Statement: Let f : [0, 1]n → {0, 1} be an increasing measurable function. There exists a set
J ⊆ {1, . . . , n} with |J| = o(n) such that either E[f (x)|xJ = ~1] ≥ 1 − o(1) or E[f (x)|xJ = ~0] ≤ o(1).
Source: Friedgut [Fri04]
Remarks:

6
 • It was wrongfully claimed in [BKK+ 92] that it is always possible to achieve E[f (x)|xJ = ~1] ≥
1 − o(1) provided E[f ] = Ω(1). However defining f (x1 , . . . , xn ) = 1 if and only if xi > n1 for
all 1 ≤ i ≤ n refutes this.

Stability of the edge-isoperimetric inequality

Statement: For every K > 0, there are positive real numbers K 0 , δ > 0 depending on K such that
Pn assertion holds. If a monotone boolean function f : {0, 1} → {0,
the following n 1} with E[f ] ≤ 12
satisfies i=1 Inf i (f ) ≤ KE[f ] log(1/E[f ]), then there is a set J of at most −K 0 log E[f ] coordinates
such that E[f (x)|xJ = ~1] ≥ (1 + δ)E[f ].
Source: [KK07].
Remarks:

• The case of E[f ] = Ω(1) follows from Friedgut’s junta theorem [Fri98]. See also [FS07] for
some relevant results and discussions.

Expected vs. critical threshold

Statement: For a graph H, let pc (H) denote the critical probability defined as Pr[G(n, pc ) ⊇ H] =
1 0
2 , and let PE (H) be the expected threshold defined as the least p such that for all H ⊆ H, we
have
1
E[number of copies H 0 in G(n, p)] ≥ .
2
Given  > 0 is there a K such that any H with  < pc (H) < 1 −  satisfies pc (H) < KpE (H)?
Source: [KK07].
Remarks:

• For arbitrary H (e.g. a hamiltonian cycle) it is possible that pc (H) & (log n)pE (H). However
it is also conjectured by Kahn and Kalai [KK07] that pc (H) . (log n)pE (H) holds for every
graph H.

Monotone DNFs and pseudo-juntas

Statement: Recall that a DNF is an OR of ANDs, and a DNF is called monotone if it only contains
un-negated boolean variables. Let  > 0 and let f : {0, 1}n → {0, 1} be defined by a monotone
DNF. Let p = p(n) and denote by µnp the p-biased product distribution on {0, 1}n . If for a random
assignment x ∼ µnp , the expected number of variables that belong to at least one satisfied clause is
K. Is there a DNF g with clauses of size at most O,K (1) such that Prx∼µnp [f (x) 6= g(x)] ≤ ?
Source: [Hat12].
Remarks:

• The main result of [Hat12] implies that this is equivalent to [Fri99, Conjecture 1.5].

The daisies problem

Statement: For x ∈ {0, 1}n and S ⊆ {1, . . . , n}, let ∆S (x) ∈ {0, 1}n be the vector that is equal to
x in the coordinates in {1, . . . , n} \ S and different from x in the coordinates in S. For every  > 0,

7
and integers 0 ≤ s ≤ t, if n is sufficiently large, and A ⊆ {0, 1}n satisfies E[A] > , then there exists
T ⊆ {1, . . . , n} of size t, and x ∈ {0, 1}n such that

∆S (x) ∈ A for all S ⊆ T with |S| = s.

Source: [BLM11].
Remarks:

• This is open even in the case of t = 4 and s = 2.

Extremal sets in the hypercube

Statement: Consider integers 1 ≤ d ≤ n and a family F of subsets F ⊆ {0, 1}d . We say that
S ⊆ {0, 1}n is F-free if every embedding ι : {0, 1}d → {0, 1}n satisfies ι(F ) 6⊆ S for all F ∈ F. Is it
true that among all F-free sets S ⊆ {0, 1}n , there is a symmetric largest set?
Source: [JT10].
Remarks:

• A positive answer would imply a positive answer to the daisies problem [BLM11] mentioned
above.

Testing correlation with cubic polynomials

Statement: Does there exist a test which queries a given f : Fn2 → F2 on a constant number of
points, and distinguishes whether f has noticeable or negligible correlation with cubic polynomials
p : Fn2 → F2 ?
More precisely, is there a universal constant k, and a protocol (an adaptive procedure) such
that given a function f : Fn2 → F2 , makes k queries to f , and for every  > 0,

• Accepts f with probability α() > 1

2 if maxp:deg(p)≤3 E(−1)f (x)+p(x) ≥ ;

• Rejects f with probability α() > 1

2 if maxp:deg(p)≤3 E(−1)f (x)+p(x) < δ, where δ > 0 can
depend on .

Source: [HL11].
Remarks:

• The Gowers U 2 and U 3 norms provide such tests for correlation with linear and quadratic
polynomials, respectively. However the example of the symmetric polynomial of degree 4
in [LMS11, GT09a] shows that the U 4 norm does not provide such a test for correlation with
cubic polynomials.

Small support implies large biased subcubes

Statement: Suppose f : {0, 1}n → R satisfy |Supp(f )| = m. Show that there is a S ⊂ {1, . . . , n}
of size O(log m) such that |E[f (x)|xS = y]| > kf k1 /poly(m) for some fixed y ∈ {0, 1}S .
Source: Shachar Lovett.
Remarks:

8
 • The case of boolean function f : {0, 1}n → {−1, 0, 1} follows from the fact that one can find
a set S and an y such that there exists only one x ∈ Supp(f ) with xS = y. The statement is
also true for symmetric functions [Shachar Lovett, personal communication].

Corners in quasi-random groups

Statement: Let δ > 0. Is it true that for sufficiently large q and G := PSL2 (q), the following
|A| 2−δ , there exists x, y, z ∈ G with z 6= 1
holds? For every subset A ⊆ G × G with density |G| 2 ≥ |G|

such that (x, y), (xz, y), (x, zy) ∈ A.

Source: Hamed Hatami.
Remarks:

• Bounds for similar problems are proven in [Tao13, Gow08, BT14, LM07, Shk06]. Proving
the conjecture for any group G would be as interesting. The motivation for this problem
comes from communication complexity. Consider the function f : G × G × G → {0, 1} defined
as f (x, y, z) = 1 if and only if xyz = 1. Using a simple hashing argument, one can see
that f has a O(1) randomized communication protocol in the NOF model. If the above
conjecture is true, then one needs Ω(log |G|) bits of communication in the deterministic NOF
model. To see this let A0 be a cylinder-intersection (i.e. A0 = (B1 × C1 × G) ∩ (B2 ×
G × D2 ) ∩ (G × C3 × D3 )) containing only elements (x, y, z) with f (x, y, z) = 1. Let A =
{(x, y) : (x, y, y −1 x−1 ) ∈ A0 }. If |A0 | = |A| ≥ |G|2−δ , then the conjecture would imply the
existence of (x, y, y −1 x−1 ), (xz, y, y −1 z −1 x−1 ), (x, zy, y −1 z −1 x−1 ) ∈ A0 . Since A0 is a cylinder-
intersection, this leads to (x, y, y −1 z −1 x−1 ) ∈ A0 contradicting z 6= 1. Hence one needs at
least |G|δ monochromatic cylinder-intersections to cover all of G × G × G, and this provides
the desired communication complexity lower-bound.

AC0 and Fourier spectrum

Statement: Let  > 0 be a fixed real number and c ≥ 1 be a fixed integer. Let f : {0, 1}n → {0, 1}
be a monotone property expressed by a depth-c circuit of size M . There is a set S of polynomial
size in M (where the polynomial depends on c and ) so that
X
fb(S)2 ≤ .
S6∈S

Source: [BKS99, Conjecture 7.2].

TC0 and Fourier spectrum

Statement: In a threshold circuit each gate is a weighted majority function with an arbitrary
number of inputs. In a monotone threshold circuit all the weights in every gate are non-negative.
Let f be a boolean function given by a monotone threshold circuit of depth c and size M . Then

X fb(S)2
≤ O (log M )c−1 .


|S|
S6=∅

Source: [BKS99, Conjecture 7.3].

9
 Graph properties and AC0
Statement: There are no balanced graph properties in AC0 . Here a graph property is a subset of
graphs which is invariant under permutation of the vertices, and balanced means that a G(n, 1/2)
random graph satisfies the property with probability tending to 1/2.
Source: Ben Rossman.
Remarks:
• It follows from Rossman’s work that containing a clique of size ∼ 2 log n is such a property
[Hamed Hatami, Personal communication].

Almost-resiliency of monotone functions

Statement: Let Cn be the class of boolean functions f : {0, 1}n → {0, 1} that are monotone and
depend on all n variables. For each n ≥ 3, let d(n) be the largest integer d such that for some
f ∈ Cn , fb(S) = 0 for all S where 2 ≤ |S| ≤ d. What is the behavior of d(n) as n → ∞?
Source: Karl Wimmer.
Remarks:
√
• The best bounds known are 2 ≤ d(n) ≤ Ω( n). Because f is monotone and depends on all
n variables, fb({i}) 6= 0 for 1 ≤ i ≤ n. If the monotonicity condition is dropped and replaced
with fb({i}) 6= 0 for 1 ≤ i ≤ n, the best bounds known are 5 ≤ d(n) ≤ n − Ω(1).

L0 sparsity of spectrum
Statement: Let f : → {0, 1}. Suppose that fˆ({i}) 6= 0 for all 1 ≤ i ≤ n. Show that the
{0, 1}n
number of non-zero coefficients of f is at least Ω(n2 ).
Source: John Wright (1st workshop at the Real Analysis special semester)
Remarks:
r
• The address function f : {0, 1}r+2 → {0, 1} defined by
f (x0 , . . . , x2r −1 , y1 , . . . , yr ) = xy1 ,...,rr .

is a function that satisfies fˆ(j) 6= 0 for all 0 ≤ j < 2r but has O(22r ) non-zero coefficients.

Size of the junta in Bourgain’s theorem

Statement: Let f : {0, 1}n → {0, 1} be a boolean function and  = k |S|>k fˆ(S)χS k2 . Show that
P
√ O(k)
f is O( k)-close to some 2  -junta.
Source: [KKO].
Remarks:
2O(k)
• [KKO] get a junta of size 4
.

Decision-tree complexity of recursive 3-majority

Statement: Let Maj : {0, 1}3 → {0, 1} denote the majority function on 3 input bits. Let x =
h
(x1 , . . . , x3h ) ∈ {0, 1}3 . Let Maj0 : {0, 1} → {0, 1} be the identity function. For every h ≥ 0,
h ∈ Z, define the recursive 3-Majority function so that
Majh (x)


:= Maj Majh−1 (x1 , . . . , x3h−1 ), Majh−1 (x1+3h−1 , . . . , x2·3h−1 ), Majh−1 (x1+2·3h−1 , . . . , x3h ) .

10
 We’re looking for a randomized decision tree which minimizes the worst-case (over inputs) cost
of computing recursive 3-majority. The best bounds we currently know are 2.5n (lower bound) and
≈ 2.65n (upper bound, 2011).
Source: [MNS+ 13]
Remarks:
• The following algorithm isn’t optimal: query two children (recusively), and query the third
one only if the first two disagree. This gives (8/3)n since worst-case inputs have 2-vs.-1 on
every internal node. Can be improved by looking at two levels at once. What’s the best that
can be achieved?

Low-degree polynomials and distance from indicators

Statement: Let p : Rn → R be a polynomial of degree d such that Var[p] ≥ 1. It is known (see
[KO12]) that for some constant c:
c
E[(|p| − 1)2 ] ≥ √ ,
d
where the expectation is with respect to the standard Gaussian distribution. What is the correct
conjectured that c is of the form cm (1 − od (1)), where cm is the constant obtained
value of c? It isP
by setting p = |S|≤d fˆ(S) · HS , where the HS are the n-variate Hermite polynomials. The same
question may be asked for polynomials over {−1, 1}n , with the requirement that all influences are
at most some τ ; then the lower bound is allowed an additive factor depending on τ .
Source: Guy Kindler, personal communication.

Correlation Bounds for Polynomials

Statement: Find an explicit (i.e., in NP) function f : Fn2 → F2 such that we have the correlation
bound |E[(−1)hf (x),p(x)i ]| ≤ 1/n for every F2 -polynomial p : Fn2 → F2 of degree at most log2 n.
Source: Folklore dating back to [Raz87, Smo87].
Remarks:
√
• The problem appears to be open even with correlation bound 1/ n replacing 1/n.

• Define the mod3 function to be 1 if and only if the number of 1’s in its input is congruent
to 1 modulo 3. Smolensky [Smo87] showed that mod3 has correlation at most 2/3 with every
√
F2 -polynomial of degree at most c n (where c > 0 is an absolute constant). For related
√
bounds using his techniques, there seems to be a barrier to obtaining correlation o(1/ n).

• Babai, Nisan, and Szegedy [BNS92] implicitly showed a function in P which has correlation
at most exp(−nΘ(1) ) with any F2 -polynomial of degree at most .99 log2 n; see also [VW08].
Bourgain [Bou05] (see also [GRS05]) showed a similar (slightly worse) result for the mod3
function.

Tomaszewski’s Conjecture
Statement: Let a ∈ Rnhave kak2 = 1. Then Prx∼{−1,1}n [|ha, xi| ≤ 1] ≥ 1/2.
Source: Question attributed to Tomaszewski in [Guy86].
Remarks:
√ √
• The bound of 1/2 would be sharp in light of a = (1/ 2, 1/ 2).

11
 • Holman and Kleitman [HK92] proved the lower bound 3/8. In fact they proved the inequality
Prx∼{−1,1}n [|ha, xi| < 1] ≥ 3/8 (assuming ai 6= ±1 for all i), which is sharp in light of
a = (1/2, 1/2, 1/2, 1/2).

Bernoulli Conjecture
Statement: Let T be a finite collection of vectors in Rn . Define b(T ) = Ex∼{−1,1}n [maxt∈T ht, xi],
and define g(T ) to be the same quantity except with x ∼ Rn Gaussian. Then there exists a finite
collection of vectors T 0 such that g(T 0 ) ≤ O(b(T )) and ∀t ∈ T ∃t0 ∈ T 0 kt − t0 k1 ≤ O(b(T )).
Source: [Tal94].
Remarks:

• The quantity g(T ) is well-understood in terms of the geometry of T , thanks to Talagrand’s

majorizing measures theorem.

• Talagrand offers a $5000 prize for proving this, and a $1000 prize for disproving it.

• The conjecture was prove by Bednorz and Latala [BL13].

Aaronson–Ambainis Conjecture
Statement: Let f : {−1, 1}n → [−1, 1] have degree at most k. Then there exists i ∈ [n] with
Inf i (f ) ≥ (V[f ]/k)O(1) .
Source: [Aar08, AA11a].
Remarks:

• True for f : {−1, 1}n → {−1, 1}; this follows from a result of O’Donnell, Schramm, Saks, and
Servedio [OSSS05].

• The weaker lower bound (V[f ]/2k )O(1) follows from a result of Dinur, Kindler, Friedgut, and
O’Donnell [DFKO07].

Bhattacharyya–Grigorescu–Shapira Conjecture
Statement: Let M ∈ Fm×k 2 and σ ∈ {0, 1}k . Say that f : Fn2 → {0, 1} is (M, σ)-free if there
does not exist X = (x , . . . , x(k) ) (where each x(j) ∈ Fn2 is a row vector) such that M X = 0 and
(1)

f (x(j) ) = σj for all j ∈ [k]. Now fix a (possibly infinite) collection {(M 1 , σ 1 ), (M 2 , σ 2 ), · · · } and
consider the property Pn of functions f : Fn2 → {0, 1} that f is (M i , σ i )-free for all i. Then there is
a one-sided error, constant-query property-testing algorithm for Pn .
Source: [BGS10].
Remarks:

• The conjecture is motivated by a work of Kaufman and Sudan [KS08] which proposes as
an open research problem the characterization of testability for linear-invariant properties of
functions f : Fn2 → {0, 1}. The properties defined in the conjecture are linear-invariant.

• Every property family (Pn ) defined by {(M 1 , σ 1 ), (M 2 , σ 2 ), · · · }-freeness is subspace-hereditary;

i.e., closed under restriction to subspaces. The converse also “essentially” holds. [BGS10].

12
 • For M of rank one, Green [Gre05b] showed that (M, 1k )-freeness is testable. He conjectured
this result extends to arbitrary M ; this was confirmed by Král’, Serra, and Vena [KSV12]
and also Shapira [Sha09]. Austin [Sha09] subsequently conjectured that (M, σ)-freeness is
testable for arbitrary σ; even this subcase is still open.

• The conjecture is known to hold when all M i have rank one [BGS10]. Also, Bhattacharyya,
Fischer, and Lovett [BFL12] have proved the conjecture in the setting of Fp for affine con-
straints {(M 1 , σ 1 ), (M 2 , σ 2 ), . . . } of “Cauchy–Schwarz complexity” less than p.

Linear CoefficientsPversus Total p Degree

Statement: Let f : {−1, 1}n → {−1, 1}. Then ni=1 fb(i) ≤ deg(f ).
Source: Parikshit Gopalan and Rocco Servedio, ca. 2009.
Remarks:

• More ambitiously, one could propose the upper bound k · k−1

1−k
k−1 2 , where k = deg(f ). This
2
is achieved by the Majority function on k bits.

• Apparently, no bound better than the trivial ni=1 fb(i) ≤ Inf (f ) ≤ deg(f ) is known.
P

k-wise Independence for PTFs

Statement: Fix d ∈ N and  ∈ (0, 1). Determine the least k = k(d, ) such that the following
holds: If p : Rn → R is any degree-d multivariate polynomial, and X is any Rn -valued random
variable with the property that each Xi has the standard Gaussian distribution and each collection
Xi1 , . . . , Xik is independent, then |Pr[p(X) ≥ 0] − Pr[p(Z) ≥ 0]| ≤ , where Z has the standard
n-dimensional Gaussian distribution.
Source: [DGJ+ 10].
Remarks:

• For d = 1, Diakonikolas, Gopalan, Jaiswal, Servedio, and Viola [DGJ+ 10] showed that k =
O(1/2 ) suffices. For d = 2, Diakonikolas, Kane, and Nelson [DKN10] showed that k =
O(d)
O(1/8 ) suffices. For general d, Kane [Kan11b] showed that Od (1) · −2 suffices and that
Ω(d2 /2 ) is necessary.

-biased Sets for DNFs

Statement: Is it true for each constant δ > 0 that s−O(1) -biased densities δ-fool size-s DNFs? I.e.,
that if f : {0, 1}n → {−1, 1} is computable by a size-s DNF and ϕ is an s−O(1) -biased density on
{0, 1}, then |Ex∼{0,1}n [f (x)] − Ey∼ϕ [f (y)]| ≤ δ.
Source: [DETT10], though the problem of pseudorandom generators for bounded-depth circuits
dates back to [AW85].
Remarks:

• De, Etesami, Trevisan, and Tulsiani [DETT10] show the result for exp(−O(log2 (s) log log s))-
biased densities. Assuming Mansour’s Conjecture, their result improves to exp(−O(log2 s)).
More precisely, they show that exp(−O(log2 (s/δ) log log(s/δ)))-biased densities δ-fool size-s
DNF. They also give an example showing that s−O(log(1/δ)) -biased densities are necessary.
Finally, they show that s−O(log(1/δ)) -biased densities suffice for read-once DNFs.

13
 PTF Sparsity for Inner Product Mod 2
Statement: Is it true that any PTF representation of the inner product mod 2 function on 2n bits,
IP2n : F2n n
2 → {−1, 1}, requires at least 3 monomials?
Source: Srikanth Srinivasan, 2010.
Remarks:

• Rocco Servedio independently asked if the following much stronger statement is true: Sup-
pose f, g : {−1, 1}n → {−1, 1} require PTFs of sparsity at least s, t, respectively; then
f ⊕ g : {−1, 1}2n → {−1, 1} (the function (x, y) 7→ f (x)g(y)) requires PTFs of sparsity at
least st.

Servedio–Tan–Verbin Conjecture
Statement: Fix any  > 0. Then every monotone f : {−1, 1}n → {−1, 1} is -close to a
poly(deg(f ))-junta.
Source: Elad Verbin (2010) and independently Rocco Servedio and Li-Yang Tan (2010)
Remarks:

• One can equivalently replace degree by decision-tree depth or maximum sensitivity.

• Disproved by Daniel Kane, 2012.

Approximate Degree for Approximate Majority

Statement: What is the leastP possible degree of a function f : {−1, 1}n → [−1, −2/3]∪[2/3, 1] which
has f (x) ∈ [2/3, 1] whenever ni=1 xi ≥ n/2 and has f (x) ∈ [−1, −2/3] whenever ni=1 xi ≤ −n/2?
P
Source: Srikanth Srinivasan, 2010.
Remarks:

• Note that f (x) is still required to be in [−1, −2/3] ∪ [2/3, 1] when −n/2 < ni=1 xi < n/2.
P

Uncertainty Principle for Quadratic Fourier Analysis

Statement: Suppose q1 , . . . , qm : Fn2 → F2 are polynomials of degree at most 2 and suppose the
indicator function of (1, . . . , 1) ∈ Fn2 , namely AND : Fn2 → {−1, 1}, is expressible as AND(x) =
P m qi (x) for some real numbers c . What is a lower bound for m?
i=1 ci (−1) i
Source: Hamed Hatami, 2011.
Remarks:

• Hatami can show that m ≥ n is necessary but conjectures m ≥ 2Ω(n) is necessary. Note that
if the qi ’s are of degree at most 1 then m = 2n is necessary and sufficient.

• The Constant-Degree Hypothesis is a similar conjecture made by Barrington, Straubing, and

Thérien [BST90] in 1990 in the context of finite fields.

14
4 Fourier analysis on the symmetric group
Friedgut–Kalai–Naor for functions on the symmetric group
Statement: A t-coset is a subset of the symmetric group on n elements Sn of the form

{π ∈ Sn : π(a1 ) = b1 , . . . , π(at ) = bt }, where ai 6= aj , bi 6= bj .

The subspace of R[Sn ] spanned by the characteristic functions of the t-cosets is denoted Ut . Let
F ⊆ Sn be a set of permutations, and let t ∈ N. Denote the characteristic function of F by f , and
let ft be the orthogonal projection of f onto Ut . If

E[(f − ft )2 ] ≤ ε E[f ].

then there exists a family G ⊆ Sn which is a union of t-cosets such that

|F4G| = O()|F|,

where the implied constant does not depend on n, t.

Source: [EFF13a, EFF13c, EFF13b].
Remarks:

• Some special cases are proved in [EFF13a, EFF13c, EFF13b], with O(O(1) ) instead of : the
case t = 1 is treated in the regime E[f ] = c/n for small c [EFF13a] and in the regime E[f ] ∈
(δ, 1−δ) [EFF13c], and the general case is treated in the regime E[f ] = c/n(n−1) · · · (n−t+1)
for small c [EFF13b].

• The question can be generalized to the subspaces Uλ spanned by the characteristic functions
of the λ-cosets. Here λ = λ1 , . . . , λ` is a partition of n and a λ-coset is

{π ∈ Sn : π(A1 ) = B1 , . . . , π(A` ) = B` }, where |Ai | = |Bi | = λi , Ai ∩ Aj = ∅, Bi ∩ Bj = ∅.

Edge-isoperimetry sets in the transposition graph of Sn

Notation: Consider the Cayley graph on Sn generated by all transpositions. The boundary ∂S of
a set A ⊆ Sn is the set of edges connecting A and Sn \ A. The initial segment of the lexicographic
order on Sn of size k consists of the first k permutations in lexicographic order (when written out
in full, i.e. 123, 132, 213, 231, 312, 321).
Statement: For any 0 ≤ k ≤ n!, the minimum of |∂A| on sets A ⊆ Sn of size k is attained by the
initial segment of the lexicographic order on Sn of size k.
Source: Conjectured by Limor Ben Efraim [Efr]; mentioned in [EFF13a, EFF13c, EFF13b].
Remarks:

• Proved for k = c(n − 1)! (for 0 ≤ c ≤ n) by Diaconis and Shahshahani [DS81]. Related
stability results in [EFF13a, EFF13c, EFF13b].

15
5 Real-Valued Functions
L1 total influence as a function of the degree
Statement: Let f : {0, 1}n → [−1, 1] satisfy fb(S) = 0 for all |S| > d, S ⊂ {1, . . . , n}. Show that
n
X f (x1 , . . . , xn ) − f (x1 , . . . , xi−1 , 1 − xi , xi+1 , . . . , xn )
E ≤ d.
x∼{0,1}n 2
i=1

Source: [AA11b].
Remarks:

• Bačkurs and Bavarian [BB13] prove a bound O(d3 log d), which can be improved to O(d3 ). A
different argument (unpublished) gives a bound d2 .

Effect of noise on L1 norm of bounded-degree functions

Statement: Let f : {0, 1}n → R satisfy fb(S) = 0 for all |S| > d, S ⊂ {1, . . . , n}. Show that

kT1−O(1/d) f k1 = Ω(kf k1 ).

(That is, there exist constants c0 , c1 such that kT1−c0 /d f k1 ≥ c1 kf k1 .)

Source:
Remarks:
2 ,n)
• Oleszkiewicz showed that kT1− f kp ≥ (1 − )min(d kf kp for 1 ≤ p ≤ ∞.

6 Extremal and additive combinatorics

Extensions of the Ahlswede–Khachatrian theorem
Statement: Let F be a collection of subsets of {1, . . . , n}. Let t ∈ Z, t > 0. We say that F is
t-intersecting, if, for every A1 , A2 ∈ F, |A1 ∩ A2 | ≥ t. We say that F is r-wise t-intersecting if,
for every A1 , . . . , Ar ∈ F, |A1 ∩ · · · ∩ Ar | ≥ t. For everyPp ∈ [0, 1], define the product probability
n Pn
n
measure µp on {0, 1} by the formula µp (x1 , . . . , xn ) := p i=1 x i (1 − p) i=1 (1−x i ) . Let m(p, t) be the

supremum of µp (F) over all t-intersecting families F. The Ahlswede–Khachatrian theorem states
that for t ≥ 2 and p < 1/2,

m(p, t) = max µp ({A ⊆ [t + 2s] : |A| ≥ t + s}),

s≥0

and the maximum is attained only at F isomorphic to one of these families.

1. Let m(p, t, r) be the supremum of µp (F) over all r-wise t-intersecting families F (every r sets
in F have at least t points in common). Show that for t ≥ 2 and p < 1 − 1/r,

m(p, t, r) = max µp ({A ⊆ [t + rs] : |A| ≥ t + s}),

s≥0

and the maximum is attained only at F isomorphic to one of these families.

16
 2. Let m0 (p, t) be the supremum of µp (F)µp (G) over all cross-t-intersecting families F, G on
p

the same point set (every set in F intersects every set in G in at least t p
points). Determine
whether m0 (p, t) = m(p, t) for t ≥ 2 and p < 1/2. (Variant: replace µp (F)µp (F) with
(µp (F) + µp (G))/2.)

3. Let t ≥ 2 and p < 1/2, and suppose that F is a t-intersecting family satisfying µp (F) ≥
(1 − )m(p, t). Show that there exists a t-intersecting family G which is Op,t ()-close to F and
has measure µp (G) = m(p, t).

Source: Item (1) appears in Tokushige [Tok05].

Remarks:

• When p < 1/(t+1), the optimal family is a t-junta, and so item (1) follows from the Ahlswede–
Khachatrian theorem, and item (3) was proved by Friedgut [Fri08] using Fourier analysis.
Analogs of items (1–3) are known in the case t = 1. When p > 1 − 1/r, m(p, t, r) = 1 by
taking all subsets of [n] of size at least d[(r − 1)n + t]/re.

Intersecting families of graphs

Statement: Let F be a family of graphs (a family of subsets of Kn , the complete graph on n vertices)
endowed with some µp measure. For a collection C of graphs, we say that F is C-intersecting if
the intersection of every two graphs in F contains a subgraph isomorphic to C.

1. If F is odd-cycle-intersecting then µp (F) ≤ p3 for all p < 3/4.

2. If F is cycle-intersecting then µp (F) ≤ p3 for all p < 1/2.

Source: [EFF12]
Remarks:

• Every two graphs with at least ≈ 34 n2 edges intersect in at least ≈ 14 n2 edges, so their

intersection contains an odd cycle by Turán’s

1theorem, so we need p < 3/4 in item (1).
1 n
Similary, every two graphs with at least 2 2 + 2 n edges intersect in at least n edges, so their
intersection contains a cycle, so we need p < 1/2 in item (2).

• Ellis, Filmus and Friedgut [EFF12] proved item (1) for p ≤ 1/2. They also obtained unique-
ness and stability. The Ahlswede–Khachatrian theorem implies item (2) for p ≤ 1/4. Unique-
ness and stability follow from Friedgut [Fri08] when p < 1/4.

Triangle removal in Fn2

Statement: Let A ⊆ Fn2 . Suppose that at least 2n elements must be removed from A in order
that A be “triangle-free”, that is, there are no pairs x, y ∈ A such that x + y ∈ A. Is it true that
Pr[x, y, x + y ∈ A] ≥ poly()?
Source: [Gre05b].
Remarks:

• Bhattacharyya and Xie [BX10] constructed an A for which the probability is at most roughly
3.409 .

17
 • The result of Fox [Fox11], along with a reduction from directed cycle removal to trian-
gle removal in all groups by Král, Serra, and Vena [KSV09], implies a lower bound of
1/Tower(log(1/)). Here Tower(0) := 1, and Tower(i) := 2Tower(i−1) .

• A Fourier-analytic proof of the above result for Fn2 was given by Hatami, Sachdeva, and
Tulsiani [HST13].

Subspaces in Sumsets
Statement: Fix a constant α > 0. Let A ⊆ Fn2 have density at least α. Is it true that A + A
√
contains a subspace of codimension O( n)?
Source: [Gre05b]
Remarks:

• The analogous problem for the group ZN dates back to Bourgain [Bou90].
√
• By considering the Hamming ball A = {x : |x| ≤ n/2 − Θ( n)}, it is easy to show that
√
codimension O( n) cannot be improved. This example is essentially due to Ruzsa [Ruz99],
see [Gre05b].

• The best bounds are due to Sanders [San11], who shows that A+A must contain a subspace of
1−α
codimension dn/(1+log2 ( 1−2α ))e. Thinking of α as small, this means a subspace of dimension
α
roughly ln 2 · n. Thinking of α = 1/2 −  for  small, this is codimension roughly n/ log2 (1/).
√
In the same work Sanders also shows that if α ≥ 1/2 − .001/ n then A + A contains a
subspace of codimension 1.

• As noted in the remarks on the Polynomial Freiman–Ruzsa/Bogolyubov Conjectures, it is

also interesting to consider the relaxed problem where we only require that A + A contains
99% of the points in a large subspace. Here it might be conjectured that the subspace can
have codimension O(log(1/α)).

Polynomial Freiman–Ruzsa Conjecture (in the Fn2 setting)

Statement: Suppose ∅ = 6 A ⊆ Fn2 satisfies |A + A| ≤ C|A|. Then A can be covered by the union of
poly(C) affine subspaces, each of cardinality at most |A|.
Source: Attributed to Marton in [Ruz99]; for the Fn2 version, see e.g. [Gre05a]
Remarks:

• The following conjecture is known to be equivalent: Suppose f : Fn2 → Fn2 satisfies Prx,y [f (x)+
f (y) = f (x + y)] ≥ , where x and y are independent and uniform on Fn2 . Then there exists
a linear function f : Fn2 → Fn2 such that Pr[f (x) = `(x)] ≥ poly().

• The PFR Conjecture is known to follow from the Polynomial Bogolyubov Conjec-
ture [GT09b]: Let A ⊆ Fn2 have density at least α. Then A + A + A contains an affine
subspace of codimension O(log(1/α)). One can slightly weaken the Polynomial Bogolyubov
Conjecture by replacing A + A + A with kA for an integer k > 3. It is known that any such
weakening (for fixed finite k) is enough to imply the PFR Conjecture.

18
 • Sanders [San12] has the best result in the direction of these conjectures, showing that if
A ⊆ Fn2 has density at least α then A + A contains 99% of the points in a subspace of
codimension O(log4 (1/α)), and hence 4A contains all of this subspace. This suffices to give
4
the Freiman–Ruzsa Conjecture with 2O(log C) in place of poly(C).

• Green and Tao [GT09b] have proved the Polynomial Freiman–Ruzsa Conjecture in the case
that A is monotone.

Routing on the hypercube

Statement: Let S, T ⊂ {0, 1}n
with |S| = |T | and S ∩ T = ∅. Let x = (x1 , . . . , xn ) ∈ {0, 1}n . Let
f : S → T be a bijection such that f (x)i ≥ xi for all i = 1, . . . , n. Does there exist an edge-disjoint
routing on the hypercube from S to T (not necessarily respecting f )? Can you do it with monotone
paths?
Source: Sushant Sachdeva

7 Fourier analysis and noise sensitivity

The propeller problem
Statement: Let n ≥ 2. Let P1 , P2 , P3 ⊆ R2 be three disjoint 120 degree sectors with cusps at the
origin. Let A1 , . . . , An+1 ⊆ Rn be a partition of Rn . That is, Ai ∩ Aj = ∅ for all i, j ∈ {1, . . . , n + 1},
i 6= j, and ∪n+1 n
i=1 Ai = R . Then

n+1
X Z 2 3
X Z 2
xdγn (x) ≤ xdγ2 (x) .
i=1 Ai `n
2 i=1 Pi `22

Source: [KN09, KN13].

Remarks:

• Allowing more than n + 1 partition elements in Rn reduces to the above inequality.

Standard Simplex Conjecture, analytic form

Statement: Let n ≥ 2, ρ ∈ (−1, 1), 3 ≤ k ≤ n + 1. Let {A1 , . . . , Ak } be a partition of Rn with
γn (Ai ) = 1/k for all i ∈ {1, . . . , k}. Let z1 , . . . , zk ∈ Rn be the vertices of a regular simplex centered
at the origin. Define a partition of Rn so that, for all i ∈ {1, . . . , k}, Bi := {x ∈ Rn : hx, zi i =
maxj=1,...,k hx, zi i}.

(a) If ρ ∈ (0, 1), then

k Z
X k Z
X
1Ai Tρ 1Ai dγn ≤ 1Bi Tρ 1Bi dγn .
i=1 Rn i=1 Rn

(b) If ρ ∈ (−1, 0), then

k Z
X k Z
X
1Ai Tρ 1Ai dγn ≥ 1Bi Tρ 1Bi dγn .
i=1 Rn i=1 Rn

19
Source: [IM12].
Remarks:
• For ρ ∈ (−1/(k − 1), 1), this conjecture is equivalent to the Plurality is Stablest conjecture
(described below).
• The case ρ < 0 implies optimal UGC hardness of approximation for the MAX−k-CUT prob-
lem.
• For k = 3, and 0 < ρ < ρ(n), the conjecture holds [Hei12].
• If k ≥ 3 and the set Ai all have positive measure but the measures are not the same, then it
is known that there isn’t an optimal partition that is given by a simplex [HMN14]

Standard Simplex Conjecture, probabilistic form

Statement: Fix 0 ≤ ρ ≤ 1. Then among all partitions of Rn into 3 ≤ q ≤ n + 1 parts of equal
Gaussian measure, the maximal noise stability at ρ occurs for a “standard simplex partition”. By
this it is meant a partition A1 , . . . , Aq satisfying Ai ⊇ {x ∈ Rn : hai , xi > haj , xi ∀j 6= i}, where
1
a1 , . . . , aq ∈ Rn are unit vectors satisfying hai , aj i = − q−1 for all i 6= j. Further, for −1 ≤ ρ ≤ 0
the standard simplex partition minimizes noise stability at ρ.
Source: [IM12]
Remarks:
• Implies the Plurality Is Stablest Conjecture from [KKMO07]; in turn, the Plurality Is Stablest
1
Conjecture implies it for ρ ≥ − q−1 .

Symmetric Gaussian Problem, analytic form

Statement: Let ρ ∈ (−1, 1). Let A ⊆ Rn satisfy −A = A. Let B, C ⊆ Rn and b, c > 0 so that
B = {x ∈ Rn : ||x||22 < b}, C = {x ∈ Rn : ||x||22 > c}, and so that γn (A) = γn (B) = γn (C). Then
Z Z Z 
1A Tρ 1A dγn ≤ max 1B Tρ 1B dγn , 1C Tρ 1C dγn .
Rn Rn Rn

Source: [CR11, O’D12].

Remarks:
• The case where the restriction A = −A is removed is well understood [MN12, Eld13].
• Even the endpoint case ρ → 1 is not completely resolved [CIMW13]. Here, the statement
then becomes one of minimizing Gaussian surface area among symmetric sets.
• (Krzysztof Oleszkiewicz) In the case of minimizing the Gaussian surface area of symmetric
sets, it seems impossible to approximate the Gaussian measure by spheres, since already for
the 3-dimensional sphere, the set of two caps is not optimal [Bar01].

Symmetric Gaussian Problem, probabilistic form

Statement: Fix 0 ≤ ρ, µ, v ≤ 1. Suppose A, B ⊆ Rn have Gaussian measure µ, v respectively.
Further, suppose A is centrally symmetric: A = −A. What is the minimal possible value of
Pr[x ∈ A, y ∈ B], when (x, y) are ρ-correlated n-dimensional Gaussians?
Source: [CR11]
Remarks:

20
 • It is equivalent to require both A = −A and B = −B.
• Without the symmetry requirement, the minimum occurs when A and B are opposing halfs-
paces; this follows from the work of Borell [Bor85].
• A reasonable conjecture is that the minimum occurs when A is a centered ball and B is the
complement of a centered ball.

Grothendieck’s inequality and the tiger partition

Statement: Let n ≥ 2. Compute the following quantity, which involves the complex-time Ornstein–
Uhlenbeck operator.
Z Z
2 2
max f (x)g(y) sin(hx, yi)e−(||x||2 +||y||2 )/2 dxdy.
f,g : Rn →{−1,1} Rn Rn

Source: [BMMN11, Question 3.1, Question 3.2, Conjecture 5.5].

Remarks:
• For n = 1, the maximum is attained by f (x) = g(x) = sign(x) [BMMN11, Theorem 1.3].
• In the case n = 2, there is a conjecture describing f, g attaining the maximum value. It is an
open problem to rigorously describe such f, g.
• Finding the best f, g would lead to improved bounds on the real Grothendieck constant.

The real Grothendieck constant

Statement: Find the minimum constant KG over all K ∈ (0, ∞) such that, for every m, n ∈ N and
for every real matrix (aij ), we have
m X
X n m X
X n
max aij hxi , yj i ≤ K max aij εi δj .
{xi }m n
i=1 ,{yj }j=1 ⊆S
n+m−1 {εi }m n
i=1 ,{δj }j=1 ⊆{−1,1}
i=1 j=1 i=1 j=1

Source: [BMMN11].
Remarks:
• It is known that KG ∈ (1.676, 2 log(1+
π √
2)
).

• There is a corresponding conjecture for complex scalars.

The Plurality is Stablest Conjecture

Notation: Let m ≥ 2, k ≥ 3. For each j ∈ [k], let ej = (0, . . . , 0, 1, 0, . . . , 0) ∈ Rk be the jth unit
coordinate vector. Let σ ∈ [k]n . Define PLURm,k : [k]m → ∆k such that

ej
 , if |{i ∈ [m] : σi = j}| > |{i ∈ [m] : σi = r}| ,
PLURm,k (σ) := ∀ r ∈ [k] \ {j}
 1 Pk

k i=1 ei , otherwise
1
Statement: Let n ≥ 2, k ≥ 3, ρ ∈ [− k−1 , 1], ε > 0. Let h·, ·i denote the standard inner product on
R . Then there exists τ > 0 such that, if f : {1, . . . , k}n → ∆k satisfies σ∈{1,...,k}n : σj 6=1 (fbi (σ))2 ≤
n
P

τ for all i ∈ {1, . . . , k}, j ∈ {1, . . . , n}, then

21
 1 P 1 Pk
(a) If ρ ∈ (0, 1], and if kn σ∈{1,...,k}n f (σ) = k i=1 ei , then

1 X 1 X
hf (σ), Tρ f (σ)i ≤ lim hPLURm,k (σ), Tρ (PLURm,k )(σ)i + ε.
kn m→∞ k m
σ∈{1,...,k}n σ∈{1,...,k}m

(b) If ρ ∈ [−1/(k − 1), 0), then

1 X 1 X
hf (σ), Tρ f (σ)i ≥ lim hPLURm,k (σ), Tρ (PLURm,k )(σ)i − ε.
kn m→∞ k m
σ∈{1,...,k}n σ∈{1,...,k}m

Source: [IM12, Hei12]

Remarks:

• The case k = 2 was proved in [MOO10], and it is known as the Majority is Stablest Theorem.
The proof used Borell’s Theorem [BS01, MN12, Eld13] (a Gaussian isoperimetric result for
two Euclidean sets) and an invariance principle (which relates moments of functions on the
discrete hypercube {−1, 1}n to moments of functions on Euclidean space).

Majority is Stablest under coefficient bounds

Statement: Let f : {−1, 1}n → {−1, 1} and let x, y ∈ {−1, 1}n be 1 > ρ > 0 correlated, so that
different coordinates are independent, x and y are uniformly distributed and E[xi yi ] = ρ for all
i = 1, . . . , n Suppose that E[f ] = 0. Is it true that
2
E[f (x)f (y)] ≤ arcsin ρ + ρ (max |fˆ(S)|),
π
where
ρ (δ) → 0 as δ → 0?
Source: This question was asked in the context of [MOO10] where a weaker result was proven
assuming low influence instead of low coefficients.
Remarks:

• The examples of the function n−1/2 (x1 − x2 )(x3 + . . . xn ) shows that invariance does not hold
assuming low coefficients [MOO10].

Majority Is Least Stable Conjecture

Statement: Let f : {−1, 1}n → {−1, 1} be a linear threshold function, n odd. Then for all ρ ∈ [0, 1],
Stabρ [f ] ≥ Stabρ [Majn ].
Source: [BKS99]
Remarks:
√ √
• Slightly weaker version: If f is a linear threshold function then NSδ [f ] ≤ π2 δ + o( δ).

• The best result towards the weaker version is Peres’s q Theorem [Per04], which shows that
2
√ 3/2
every linear threshold function f satisfies NSδ [f ] ≤ π δ + O(δ ).

22
 • By taking ρ → 0, the conjecture has the following consequence, which isP also open: Let
f : {−1, 1} → {−1, 1} be a linear threshold function with E[f ] = 0. Then ni=1 fb(i)2 ≥ π2 .
n

The best known lower bound here is 21 , which follows from the Khinchine–Kahane inequality;
see [GL94].

Talagrand’s “Convolution with a Biased Coin” Conjecture

Statement: Let f : {−1, 1}n → R≥0 have E[f ] = 1. Fix any 0 < ρ < 1. Then Pr[Tρ f ≥ t] < o(1/t).
Source: [Tal89]
Remarks:

• Talagrand in fact suggests the bound O( t√log

1
t
).

• Talagrand offers a $1000 prize for proving this.

• Even the “special case” when f ’s domain is Rn with Gaussian measure is open. In this
Gaussian setting, Ball, Barthe, Bednorz, Oleszkiewicz, and Wolff [BBB+ 13] have shown the
1
upper bound O( t√log t
) for n = 1 and the bound O( log
√ log t ) for any fixed constant dimension.
t log t

Sensitivity versus Block Sensitivity

Statement: For any f : {−1, 1}n → {−1, 1} it holds that deg(f ) ≤ poly(sens[f ]), where sens[f ] is
the (maximum) sensitivity, maxx |{i ∈ [n] : f (x) 6= f (x1 , . . . , xi−1 , −xi , xi+1 , . . . , xn )}|.
Source: [CFGS88, Sze89, GL92, NS94]
Remarks:

• As the title suggests, it is more usual to state this as bs[f ] ≤ poly(sens[f ]), where bs[f ] is
the “block sensitivity”. However the version with degree is equally old, and in any case the
problems are equivalent since it is known that bs[f ] and deg(f ) are polynomially related.

• The best known gap is quadratic ([CFGS88, GL92]) and it is suggested ([GL92]) that this
may be the worst possible.

Gotsman–Linial Conjecture
Statement: Among degree-k polynomial threshold functions f : {−1, 1}n → {−1, 1}, the one with
maximal total influence is the symmetric one f (x) = sgn(p(x1 + · · · + xn )), where p is a degree-k
univariate polynomial which alternates sign on the k + 1 values of x1 + · · · + xn closest to 0.
Source: [GL94]
Remarks:

• The case k = 1 is easy.

√
• Slightly weaker version: degree-k PTFs have total influence O(k) · n.
√
• Even weaker version: degree-k PTFs have total influence Ok (1) · n.

• The weaker versions are open even in the case k = 2. The k = 2 case may be related to the
following old conjecture of Holzman: If g : {−1, 1}n → R has degree 2 (for n even), then g
n


has at most n/2 local strict minima.

23
 √
• It is known that bounding
√ total influence by c(k) · n is equivalent to a bounding δ-noise
sensitivity by O(c(k)) · δ.

• The “Gaussian special case” was solved by Kane [Kan11a].

k
• The best upper bounds known are 2n1−1/2 and 2O(k) · n1−1/O(k) [DHK+ 10].

• Recently, Daniel Kane [Kan13] proved that for any degree-k PTF, the maximum total influ-
ence is at most ck n1/2+o(1) . See [Kan13, O’D12] for a detailed history.

Mansour’s Conjecture
Statement: Let f : {−1, 1}n
→ {−1, 1} be computable by a DNF of size s > 1 and let  ∈ (0, 1/2].
Then f ’s Fourier spectrum is -concentrated on a collection F with |F| ≤ sO(log(1/)) .
Source: [Man94]
Remarks:

• Weaker version: replacing sO(log(1/)) by sO (1) .

• The weak version with bound sO(1/) is known to follow from the Fourier Entropy–Influence
Conjecture.

• Proved for “almost all” polynomial-size DNF formulas (appropriately defined) by Klivans,
Lee, and Wan [KLW10].

• Mansour [Man95] obtained the upper-bound (s/)O(log log(s/) log(1/)) .

Fourier Entropy–Influence Conjecture

Statement: There is a universal constant
P b C2 such that for any f : {−1, 1}n → {−1, 1} it holds that
2 2 1
H[f ] ≤ C · I[f ], where H[f ] = S f (S) log2 b 2 is the spectral entropy and I[f ] is the total
b b
f (S)
influence.
Source: [FK96]
Remarks:

• Proved for “almost all” polynomial-size DNF formulas (appropriately defined) by Klivans,
Lee, and Wan [KLW10].

• Proved for symmetric functions and functions computable by read-once decision trees by
O’Donnell, Wright, and Zhou [OWZ11].

• An explicit example showing that C ≥ 60/13 is necessary is known. (O’Donnell, unpublished.)

• Weaker version: the “Min-Entropy–Influence Conjecture”, which states that there exists S
such that fb(S)2 ≥ 2−C·I[f ] . This conjecture is strictly stronger than the KKL Theorem, and
is implied by the KKL Theorem in the case of monotone functions.

Optimality of Majorities for Non-Interactive Correlation Distillation

Statement: Fix r ∈ N, n odd, and 0 <  < 1/2. For f : {−1, 1}n → {−1, 1}, define P (f ) =
Pr[f (y (1) ) = f (y (2) ) = · · · f (y (r) )], where x ∼ {−1, 1}n is chosen uniformly and then each y (i) is

24
(independently) an -noisy copy of x. Is it true that P (f ) is maximized among odd functions f by
the Majority function Majk on some odd number of inputs k?
Source: [MO05] (originally from 2002)
Remarks:

• It is possible (e.g., for r = 10, n = 5,  = .26) for neither the Dictator (Maj1 ) nor full Majority
(Majn ) to be maximizing.

Noise Sensitivity of Intersections of Halfspaces

Statement: Let f :√{−1, 1}n√→ {−1, 1} be the intersection (AND) of k linear threshold functions.
Then NSδ [f ] ≤ O( log k) · δ.
Source: [KOS04]
Remarks:
√
• The bound O(k) · δ follows easily from Peres’s Theorem and is the best known.

• The “Gaussian special case” follows easily from the work of Nazarov [Naz03].

• An upper bound of the form polylog(k) · δ Ω(1) holds if the halfspaces are sufficiently “regu-
lar” [HKM12].

Average sensitivity of intersection of half-spaces

Statement: Let f : {0, 1}n → {0, 1} be the indicator function of an intersection of k half-spaces.
Then there exists a universal constant C > 0 such that
X
2−n
p
|{i ∈ {1, . . . , n} : f (x) 6= f (x1 , . . . , xi−1 , 1 − xi , xi+1 , . . . , xn )}| ≤ C n log k.
x=(x1 ,...,xn )∈{0,1}n
p
Inf (f ) ≤ C n log k.
Source: [Kan]
Remarks:

• This conjecture was settled by Daniel Kane [Kan].

Non-Interactive Correlation Distillation with Erasures

Statement: Let f : {−1, 1}n → {−1, 1} be an unbiased function. Let z ∼ {−1, 0, 1}n be a “random
restriction” in which each coordinate zi is (independently) ±1 with probability p/2 each, and 0
with probability 1 − p. Assuming p < 1/2 and n odd, is it true that Ez [|f (z)|] is maximized when
f is the majority function? (Here we identify f with its multilinear expansion.)
Source: [Yan07]
Remarks:

• For p ≥ 1/2, Yang conjectured that Ez [|f (z)|] is maximized when f is a dictator function;
this was proved by O’Donnell and Wright [OW12].

• Mossel [Mos10] shows that if f ’s influences are assumed at most τ then the following inequality
holds: Ez [|f (z)|] ≤ Ez [|Majn (z)|] + oτ (1).

25
 Average versus Max Sensitivity for Monotone Functions
Statement: Let f : {−1, 1}n → {−1, 1} be monotone. Then I[f ] < o(sens[f ]).
Source: Rocco Servedio, Li-Yang Tan, 2010
Remarks:

• The tightest example known has I[f ] ≈ sens[f ].61 ; this appears in a work of O’Donnell and
Servedio [OS07].

Noise correlation bounds

Statement: Let (Ωi , µi ) be pairwise independent distributions and let Ω = K Ωi , µ = K
Q Q
i=1 i=1 µi ,
n
P
Ωi = {1, . . . , ki }. Let fi : Ωi → [−1, 1], fi = σ∈{1,...,ki }n f (σ)Wσ satisfy
b
X
k fb(σ)Wσ k22 ≤ (1 − γ)d .
σ∈{1,...,ki }n : |σ|≥d

Suppose further that f1 satisfies |fˆ1 (σ)| ≤ δ for all σ ∈ {1, . . . , k1 }n then

E [f1 (x1 ) . . . fK (xK )] ≤ (δ), (δ) → 0 as δ → 0.

µn

Source: Austrin and Mossel [AM13].

Remarks:

• In the special case where fi are linear forms over the same additive group, this was shown to
hold by Hamed Hatami using the Gowers–Cauchy–Schwartz inequality (see [AM13]).

KKL for continuous domains

Statement: We have f : [0, 1]R → {1, . . . , q}, where [0, 1] is endowed with its Haar measure. We
want to prove versions of [KKL88] and majority is stablest in this setting. The influence can be
defined as

Inf i (f ) = Pra,b∼[0,1] f (x1 , . . . , xi−1 , a, xi+1 , . . . , xR ) = f (x1 , . . . , xi−1 , b, xi+1 , . . . , xR ).

(This is different from the [BKK+ 92] definition.) We want f to be balanced : no value is obtained
with probability more than . We want to prove that the sum of influences is at least 2−φ()−oR (1)
(rather than 1); the Poincaré inequality gives only 1.
Now suppose that furthermore all influences are at most τ . Is it true that that the η-noise
stability (the probability that the function changes when we flip each coordinate with probability
η) is at most 1 − 2? This is tight for a discretized version of max.
Source: Prasad Raghavendra

26
8 Computational hardness
MAX-2-LIN(2) on the Hypercube
Statement: Suppose you have variables on the vertices of a k-hypercube. For each of the edges, we
have a constraint xi + xj = c. Is it Unique Games hard to approximately satisfy 1 −  fraction of
√
the constraints vs. 1 −  fraction of the constraints? Is the corresponding question for 1 −  vs.
1 − M  easy (perhaps even using Goemans–Williamson)?
Source: [KKMO07]

Approximation Resistance of the republic function

Statement: Define the republic predicate as follows:
k
X
sign(ckX0 + Xi ).
i=1

Is this predicate approximation-resistant? In other words, given a set of m such constraints which
is 1 −  satisfiable, can we find an assignment that satisfies 1/2 +  fraction of the constraints?
Source: Johan Håstad

27
References
[AA11a] Scott Aaronson and Andris Ambainis, The need for structure in quantum speedups,
Proceedings of the 2nd Annual Symposium on Innovations in Computer Science, 2011.
12

[AA11b] Scott Aaronson and Andris Ambainis, The need for structure in quantum speedups,
ITCS, 2011. 16

[Aar08] Scott Aaronson, How to solve longstanding open problems in quantum computing us-
ing only Fourier Analysis, Invited lecture at Banff International Research Station,
http://www.scottaaronson.com/talks/openqc.ppt, 2008. 12

[AM13] Per Austrin and Elchanan Mossel, Noise correlation bounds for uniform low degree
functions, Ark. Mat. 51 (2013), no. 1, 29–52. MR 3029335 26

[AW85] Miklós Ajtai and Avi Wigderson, Deterministic simulation of probabilistic constant
depth circuits, Proceedings of the 26th Annual IEEE Symposium on Foundations of
Computer Science, 1985, pp. 11–19. 13

[Bar01] F. Barthe, An isoperimetric result for the Gaussian measure and unconditional sets,
Bull. London Math. Soc. 33 (2001), no. 4, 408–416. MR 1832552 (2002c:60027) 20

[BB13] Artūrs Bačkurs and Mohamma Bavarian, On the sum of l1 influences, ITCS, 2013. 16

[BBB+ 13] K. Ball, F. Barthe, W. Bednorz, K. Oleszkiewicz, and P. Wolff, L1 -smoothing for the
Ornstein-Uhlenbeck semigroup, Mathematika 59 (2013), no. 1, 160–168. MR 3028177
23

[BdW02] Harry Buhrman and Ronald de Wolf, Complexity measures and decision tree complex-
ity: a survey, Theoret. Comput. Sci. 288 (2002), no. 1, 21–43, Complexity and logic
(Vienna, 1998). MR 1934888 (2003i:68035) 6

[BFL12] Arnab Bhattacharyya, Eldar Fischer, and Shachar Lovett, Testing low complexity
affine-invariant properties, preprint, arXiv:1201.0330, 2012. 13

[BGS10] Arnab Bhattacharyya, Elena Grigorescu, and Asaf Shapira, Unified framework for
testing linear-invariant properties, 2010 IEEE 51st Annual Symposium on Foundations
of Computer Science FOCS 2010, IEEE Computer Soc., Los Alamitos, CA, 2010,
pp. 478–487. MR 3025221 12, 13

[BKK+ 92] Jean Bourgain, Jeff Kahn, Gil Kalai, Yitzhak Katznelson, and Nathan Linial, The
influence of variables in product spaces, Israel J. Math. 77 (1992), no. 1-2, 55–64. MR
1194785 (94g:05091) 7, 26

[BKS99] Itai Benjamini, Gil Kalai, and Oded Schramm, Noise sensitivity of Boolean functions
and applications to percolation, Inst. Hautes Études Sci. Publ. Math. (1999), no. 90,
5–43 (2001). MR 1813223 (2001m:60016) 9, 22

[BL13] W. Bednorz and R. Latala, On the boundedness of Bernoulli processes, unpublished,

arXiv:1305.4292. 12

28
[BLM11] Béla Bollobás, Imre Leader, and Claudia Malvenuto, Daisies and other Turán problems,
Combin. Probab. Comput. 20 (2011), no. 5, 743–747. MR 2825587 (2012i:05334) 8

[BMMN11] Mark Braverman, Konstantin Makarychev, Yury Makarychev, and Assaf Naor, The
Grothendieck constant is strictly smaller than Krivine’s bound, 2011 IEEE 52nd Annual
Symposium on Foundations of Computer Science—FOCS 2011, IEEE Computer Soc.,
Los Alamitos, CA, 2011, pp. 453–462. MR 2932721 21

[BNS92] László Babai, Noam Nisan, and Márió Szegedy, Multiparty protocols, pseudorandom
generators for logspace, and time-space trade-offs, J. Comput. System Sci. 45 (1992),
no. 2, 204–232, Twenty-first Symposium on the Theory of Computing (Seattle, WA,
1989). MR 1186884 (93m:68048) 11

[Bor85] Christer Borell, Geometric bounds on the Ornstein-Uhlenbeck velocity process, Z.

Wahrsch. Verw. Gebiete 70 (1985), no. 1, 1–13. MR 795785 (87k:60103) 21

[Bou90] J. Bourgain, On arithmetic progressions in sums of sets of integers, A tribute to Paul

Erdős, Cambridge Univ. Press, Cambridge, 1990, pp. 105–109. MR 1117007 (92e:11011)
18

[Bou05] Jean Bourgain, Estimation of certain exponential sums arising in complexity theory,
C. R. Math. Acad. Sci. Paris 340 (2005), no. 9, 627–631. MR 2139268 (2006a:11104)
11

[BS01] A. Burchard and M. Schmuckenschlger, Comparison theorems for exit times, Geometric
& Functional Analysis GAFA 11 (2001), no. 4, 651–692 (English). 22

[BST90] David A. Mix Barrington, Howard Straubing, and Denis Thérien, Nonuniform au-
tomata over groups, Inform. and Comput. 89 (1990), no. 2, 109–132. MR 1080845
(92e:68054) 14

[BT14] Vitaly Bergelson and Terence Tao, Multiple recurrence in quasirandom groups, Geo-
metric and Functional Analysis (2014), 1–48 (English). 9

[BX10] Arnab Bhattacharyya and Ning Xie, Lower bounds for testing triangle-freeness in
Boolean functions, Proceedings of the Twenty-First Annual ACM-SIAM Symposium
on Discrete Algorithms (Philadelphia, PA), SIAM, 2010, pp. 87–98. MR 2809663
(2012i:68273) 17

[CFGS88] F. R. K. Chung, Zoltán Füredi, R. L. Graham, and P. Seymour, On induced subgraphs

of the cube, J. Combin. Theory Ser. A 49 (1988), no. 1, 180–187. MR 957216 (89i:05101)
23

[CIMW13] Tobias Holck Colding, Tom Ilmanen, William P. Minicozzi, II, and Brian White, The
round sphere minimizes entropy among closed self-shrinkers, J. Differential Geom. 95
(2013), no. 1, 53–69. MR 3128979 20

[CR11] Amit Chakrabarti and Oded Regev, An optimal lower bound on the communication
complexity of gap Hamming distance, Proc. 43rd Annual ACM Symposium on the
Theory of Computing, 2011, pp. 51–60. 20

29
[DETT10] Anindya De, Omid Etesami, Luca Trevisan, and Madhur Tulsiani, Improved pseudo-
random generators for depth 2 circuits, Approximation, randomization, and combina-
torial optimization, Lecture Notes in Comput. Sci., vol. 6302, Springer, Berlin, 2010,
pp. 504–517. MR 2755859 (2012g:68169) 13

[DFKO07] Irit Dinur, Ehud Friedgut, Guy Kindler, and Ryan O’Donnell, On the Fourier tails
of bounded functions over the discrete cube, Israel J. Math. 160 (2007), 389–412. MR
2342503 (2008m:43012) 12

[DGJ+ 10] Ilias Diakonikolas, Parikshit Gopalan, Ragesh Jaiswal, Rocco A. Servedio, and
Emanuele Viola, Bounded independence fools halfspaces, SIAM J. Comput. 39 (2010),
no. 8, 3441–3462. MR 2721612 (2011j:60044) 13

[DHK+ 10] Ilias Diakonikolas, Prahladh Harsha, Adam Klivans, Raghu Meka, Prasad Raghaven-
dra, Rocco A. Servedio, and Li-Yang Tan, Bounding the average sensitivity and noise
sensitivity of polynomial threshold functions, STOC’10—Proceedings of the 2010 ACM
International Symposium on Theory of Computing, ACM, New York, 2010, pp. 533–
542. MR 2743302 (2011i:68064) 24

[DKN10] Ilias Diakonikolas, Daniel M. Kane, and Jelani Nelson, Bounded independence fools
degree-2 threshold functions, 2010 IEEE 51st Annual Symposium on Foundations of
Computer Science FOCS 2010, IEEE Computer Soc., Los Alamitos, CA, 2010, pp. 11–
20. MR 3024771 13

[DS81] Persi Diaconis and Mehrdad Shashahani, Generating a random permutation with ran-
dom transpositions, Z. Wahrsch. Verw. Gebeite 57 (1981), 159–179. 15

[EFF12] David Ellis, Yuval Filmus, and Ehud Friedgut, Triangle-intersecting families of graphs,
J. Eur. Math. Soc. 14 (2012), no. 3, 841–885. 17

[EFF13a] , A quasi-stability result for dictatorships in Sn , Submitted (2013). 15

[EFF13b] , A quasi-stability result for low-degree boolean functions on Sn , Manuscript

(2013). 15

[EFF13c] , A stability result for balanced dictatorships in Sn , Rand. Struct. Alg. (2013).
15

[Efr] Limor Ben Efraim, Personal communication. 15

[Eld13] Ronen Eldan, A two-sided estimate for the gaussian noise stability deficit, Preprint,
arXiv:1307.2781, 2013. 20, 22

[FK96] Ehud Friedgut and Gil Kalai, Every monotone graph property has a sharp threshold,
Proc. Amer. Math. Soc. 124 (1996), no. 10, 2993–3002. MR 1371123 (97e:05172) 24

[Fox11] Jacob Fox, A new proof of the graph removal lemma., Annals of Mathematics 174
(2011), no. 1, 561–579. 18

[Fri98] Ehud Friedgut, Boolean functions with low average sensitivity depend on few coordi-
nates, Combinatorica 18 (1998), no. 1, 27–35. MR 1645642 (99g:28021) 7

30
[Fri99] , Sharp thresholds of graph properties, and the k-sat problem, J. Amer. Math.
Soc. 12 (1999), no. 4, 1017–1054, With an appendix by Jean Bourgain. MR 1678031
(2000a:05183) 7

[Fri04] , Influences in product spaces: KKL and BKKKL revisited, Combin. Probab.
Comput. 13 (2004), no. 1, 17–29. MR 2034300 (2004m:06021) 6

[Fri08] Ehud Friedgut, On the measure of intersecting families, uniqueness and stability, Com-
binatorica 28 (2008), no. 5, 503–528. 17

[FS07] Dvir Falik and Alex Samorodnitsky, Edge-isoperimetric inequalities and influences,
Combin. Probab. Comput. 16 (2007), no. 5, 693–712. MR 2346808 (2008g:05061) 7

[GL92] C. Gotsman and N. Linial, The equivalence of two problems on the cube, J. Combin.
Theory Ser. A 61 (1992), no. 1, 142–146. MR 1178392 (93d:05074) 23

[GL94] Craig Gotsman and Nathan Linial, Spectral properties of threshold functions, Combi-
natorica 14 (1994), no. 1, 35–50 (English). 23

[Gow08] W. T. Gowers, Quasirandom groups, Combin. Probab. Comput. 17 (2008), no. 3, 363–
387. MR 2410393 (2009f:20105) 9

[Gre05a] Ben Green, Finite field models in additive combinatorics, Surveys in combinatorics
2005, London Math. Soc. Lecture Note Ser., vol. 327, Cambridge Univ. Press, Cam-
bridge, 2005, pp. 1–27. MR 2187732 (2006j:11030) 18

[Gre05b] , A Szemerédi-type regularity lemma in abelian groups, with applications, Geo-

metric And Functional Analysis 15 (2005), 340–376, 10.1007/s00039-005-0509-8. 13,
17, 18

[GRS05] Frederic Green, Amitabha Roy, and Howard Straubing, Bounds on an exponential sum
arising in Boolean circuit complexity, C. R. Math. Acad. Sci. Paris 341 (2005), no. 5,
279–282. MR 2166139 (2006e:11117) 11

[GT09a] Ben Green and Terence Tao, The distribution of polynomials over finite fields, with
applications to the Gowers norms, Contrib. Discrete Math. 4 (2009), no. 2, 1–36. MR
2592422 (2011d:11022) 8

[GT09b] , Freiman’s theorem in finite fields via extremal set theory, Combin. Probab.
Comput. 18 (2009), no. 3, 335–355. MR 2501431 (2010f:11176) 18, 19

[Guy86] Richard K. Guy, Unsolved Problems: Any Answers Anent These Analytical Enigmas?,
Amer. Math. Monthly 93 (1986), no. 4, 279–281. MR 1540838 11

[Hat12] Hamed Hatami, A structure theorem for Boolean functions with small total influences,
Ann. of Math. (2) 176 (2012), no. 1, 509–533. MR 2925389 7

[Hei12] Steven Heilman, Euclidean partitions that optimize an ornstein-uhlenbeck quadratic

form, Submitted, arXiv:1211.7138v1, 2012. 20, 22

31
[HK92] Ron Holzman and Daniel J. Kleitman, On the product of sign vectors and unit vectors,
Combinatorica 12 (1992), no. 3, 303–316. MR 1195893 (94e:60007) 12

[HKM12] Prahladh Harsha, Adam Klivans, and Raghu Meka, An invariance principle for poly-
topes, J. ACM 59 (2012), no. 6, Art. 29, 25. MR 3008401 25

[HL11] Hamed Hatami and Shachar Lovett, Correlation testing for affine invariant properties
on Fnp in the high error regime, Proc. 43rd Annual ACM Symposium on the Theory of
Computing, 2011, pp. 187–194. 8

[HMN14] S. Heilman, E. Mossel, and J. Neeman, Standard Simplices and Pluralities are Not the
Most Noise Stable, Arxiv:1403.0885, 2014. 20

[HMO14] S. Heilman, E. Mossel, and K. Oleszkiewicz, Strong contraction and influences in tail
spaces, 2014. 6

[HST13] P. Hatami, S. Sachdeva, and M. Tulsiani, An Arithmetic Analogue of Fox’s Triangle

Removal Argument, ArXiv e-prints (2013). 18

[IM12] Marcus Isaksson and Elchanan Mossel, Maximally stable Gaussian partitions with dis-
crete applications, Israel J. Math. 189 (2012), 347–396. MR 2931402 20, 22

[JT10] J. Robert Johnson and John Talbot, Vertex Turán problems in the hypercube, J. Com-
bin. Theory Ser. A 117 (2010), no. 4, 454–465. MR 2592894 (2011h:05127) 8

[Kan] Daniel Kane, The average sensitivity of an intersection of half spaces, unpublished,
arXiv:1309.2987. 25

[Kan11a] Daniel M. Kane, The Gaussian surface area and noise sensitivity of degree-D polyno-
mial threshold functions, Comput. Complexity 20 (2011), no. 2, 389–412. MR 2822876
(2012f:68155) 24

[Kan11b] , A small PRG for polynomial threshold functions of Gaussians, 2011 IEEE
52nd Annual Symposium on Foundations of Computer Science—FOCS 2011, IEEE
Computer Soc., Los Alamitos, CA, 2011, pp. 257–266. MR 2932701 13

[Kan13] D.M. Kane, The correct exponent for the gotsman-linial conjecture, Computational
Complexity (CCC), 2013 IEEE Conference on, 2013, pp. 56–64. 24

[KK07] Jeff Kahn and Gil Kalai, Thresholds and expectation thresholds, Combin. Probab. Com-
put. 16 (2007), no. 3, 495–502. MR 2312440 (2008g:60025) 7

[KKL88] J. Kahn, G. Kalai, and N. Linial, The influence of variables on boolean functions, Pro-
ceedings of the 29th Annual Symposium on Foundations of Computer Science (Wash-
ington, DC, USA), SFCS ’88, IEEE Computer Society, 1988, pp. 68–80. 26

[KKMO07] Subhash Khot, Guy Kindler, Elchanan Mossel, and Ryan O’Donnell, Optimal inap-
proximability results for MAX-CUT and other 2-variable CSPs?, SIAM J. Comput. 37
(2007), no. 1, 319–357. MR 2306295 (2008d:68035) 20, 27

32
[KKO] Guy Kindler, Naomi Kirshner, and Ryan O’Donnell, Gaussian noise sensitivity and
Fourier tails. 10

[KLW10] Adam Klivans, Homin Lee, and Andrew Wan, Mansour’s Conjecture is true for random
DNF formulas, Proceedings of the 23rd Annual Conference on Learning Theory, 2010,
pp. 368–380. 24

[KN09] Subhash Khot and Assaf Naor, Approximate kernel clustering, Mathematika 55 (2009),
no. 1-2, 129–165. MR 2573605 (2011c:68166) 19

[KN13] , Sharp kernel clustering algorithms and their associated grothendieck inequali-
ties, Random Structures & Algorithms 42 (2013), no. 3, 269–300. 19

[KO12] Guy Kindler and Ryan O’Donnell, Gaussian noise sensitivity and fourier tails, Com-
putational Complexity (CCC), 2012 IEEE 27th Annual Conference on, IEEE, 2012,
pp. 137–147. 11

[KOS04] Adam R. Klivans, Ryan O’Donnell, and Rocco A. Servedio, Learning intersections and
thresholds of halfspaces, J. Comput. System Sci. 68 (2004), no. 4, 808–840. MR 2059646
(2005g:68060) 25

[KS08] Tali Kaufman and Madhu Sudan, Algebraic property testing: the role of invariance,
STOC’08, ACM, New York, 2008, pp. 403–412. MR 2582904 (2011c:68076) 12

[KSV09] D. Král, O. Serra, and L. Vena, A combinatorial proof of the removal lemma for groups,
Journal of Combinatorial Theory, Series A 116 (2009), no. 4, 971–978. 18

[KSV12] Daniel Král,’ Oriol Serra, and Lluı́s Vena, A removal lemma for systems of linear equa-
tions over finite fields, Israel J. Math. 187 (2012), 193–207. MR 2891704 13

[LM07] Michael T. Lacey and William McClain, On an argument of Shkredov on two-

dimensional corners, Online J. Anal. Comb. (2007), no. 2, Art. 2, 21. MR 2289954
(2007m:11037) 9

[LMS11] Shachar Lovett, Roy Meshulam, and Alex Samorodnitsky, Inverse conjecture for the
Gowers norm is false, Theory Comput. 7 (2011), 131–145. MR 2862496 8

[Man94] Yishay Mansour, Learning Boolean functions via the Fourier Transform, Theoretical
Advances in Neural Computation and Learning (Vwani Roychowdhury, Kai-Yeung Siu,
and Alon Orlitsky, eds.), Kluwer Academic Publishers, 1994, pp. 391–424. 24

[Man95] , An O(nlog log n ) learning algorithm for DNF under the uniform distribution,
J. Comput. System Sci. 50 (1995), no. 3, part 3, 543–550, Fifth Annual Workshop
on Computational Learning Theory (COLT) (Pittsburgh, PA, 1992). MR 1339562
(96h:68171) 24

[MN12] Elchanan Mossel and Joe Neeman, Robust optimality of gaussian noise stability,
Preprint, arXiv:1210.4126, 2012. 20, 22

33
[MNS+ 13] Frederic Magniez, Ashwin Nayak, Miklos Santha, Jonah Sherman, Gabor Tardos, and
David Xiao, Improved bounds for the randomized decision tree complexity of recursive
majority, Preprint, arXiv:1309.7565, 2013. 11

[MO05] Elchanan Mossel and Ryan O’Donnell, Coin flipping from a cosmic source: on error
correction of truly random bits, Random Structures Algorithms 26 (2005), no. 4, 418–
436. MR 2139875 (2006c:60012) 25

[MOO10] Elchanan Mossel, Ryan O’Donnell, and Krzysztof Oleszkiewicz, Noise stability of func-
tions with low influences: invariance and optimality, Ann. of Math. (2) 171 (2010),
no. 1, 295–341. MR 2630040 (2012a:60091) 22

[Mos10] Elchanan Mossel, Gaussian bounds for noise correlation of functions, Geom. Funct.
Anal. 19 (2010), no. 6, 1713–1756. MR 2594620 (2011b:60080) 25

[Naz03] Fedor Nazarov, On the maximal perimeter of a convex set in Rn with respect to a
Gaussian measure, Geometric aspects of functional analysis, Lecture Notes in Math.,
vol. 1807, Springer, Berlin, 2003, pp. 169–187. MR 2083397 (2005h:28032) 25

[NS94] Noam Nisan and Márió Szegedy, On the degree of Boolean functions as real polynomials,
Comput. Complexity 4 (1994), no. 4, 301–313, Special issue on circuit complexity
(Barbados, 1992). MR 1313531 (96d:68096) 23

[O’D12] Ryan O’Donnell, Open problems in analysis of boolean functions, Preprint,

arXiv:1204.6447v1, 2012. 2, 20, 24

[O’D14] , Analysis of boolean functions, Cambridge University Press, 2014. 2, 4

[OS07] Ryan O’Donnell and Rocco A. Servedio, Learning monotone decision trees in poly-
nomial time, SIAM J. Comput. 37 (2007), no. 3, 827–844 (electronic). MR 2341918
(2008k:68045) 26

[OSSS05] Ryan O’Donnell, Michael Saks, Oded Schramm, and Rocco A. Servedio, Every deci-
sion tree has an influential variable, Foundations of Computer Science, IEEE Annual
Symposium on 0 (2005), 31–39. 6, 12

[OW12] Ryan O’Donnell and John Wright, A new point of NP-hardness for unique games
[extended abstract], STOC’12—Proceedings of the 2012 ACM Symposium on Theory
of Computing, ACM, New York, 2012, pp. 289–306. MR 2961512 25

[OWZ11] Ryan O’Donnell, Yi Wu, and Yuan Zhou, Optimal lower bounds for locality sensi-
tive hashing (except when q is tiny), Proceedings of the 2nd Annual Symposium on
Innovations in Computer Science, 2011. 24

[Per04] Yuval Peres, Noise stability of weighted majority, Technical Report, arXiv:0412377,
2004. 22

[Raz87] A. A. Razborov, Lower bounds on the dimension of schemes of bounded depth in a

complete basis containing the logical addition function, Mat. Zametki 41 (1987), no. 4,
598–607, 623. MR 897705 (89h:03110) 11

34
[Ruz99] Imre Z. Ruzsa, An analog of Freiman’s theorem in groups, Astérisque (1999), no. 258,
xv, 323–326, Structure theory of set addition. MR 1701207 (2000h:11111) 18

[San11] Tom Sanders, Green’s sumset problem at density one half, Acta Arith. 146 (2011),
no. 1, 91–101. MR 2741192 (2011k:11017) 18

[San12] , On the Bogolyubov-Ruzsa lemma, Anal. PDE 5 (2012), no. 3, 627–655. MR

2994508 19

[Sha09] Asaf Shapira, Green’s conjecture and testing linear-invariant properties, STOC’09—
Proceedings of the 2009 ACM International Symposium on Theory of Computing,
ACM, New York, 2009, pp. 159–166. MR 2780061 (2012f:68190) 13

[Shk06] I. D. Shkredov, On a problem of Gowers, Izv. Ross. Akad. Nauk Ser. Mat. 70 (2006),
no. 2, 179–221. MR 2223244 (2007e:11030) 9

[Smo87] Roman Smolensky, Algebraic methods in the theory of lower bounds for Boolean circuit
complexity, Proceedings of the 19th Annual ACM Symposium on Theory of Computing,
1987, pp. 77–82. 11

[Sze89] Mario Szegedy, Algebraic methods in lower bounds for computational models with lim-
ited communication, PhD thesis, University of Chicago, 1989. 23

[Tal89] Michel Talagrand, A conjecture on convolution operators, and a non-Dunford-Pettis

operator on L1 , Israel J. Math. 68 (1989), no. 1, 82–88. MR 1035882 (91b:47078) 23

[Tal94] M. Talagrand, Constructions of majorizing measures, Bernoulli processes and cotype,

Geom. Funct. Anal. 4 (1994), no. 6, 660–717. MR 1302336 (95j:60008) 12

[Tao13] Terence Tao, Mixing for progressions in nonabelian groups, Forum Math. Sigma 1
(2013), e2 (40 pages). MR 3143714 9

[Tok05] Norihide Tokushige, Intersecting families — uniform versus weighted, Ryukyu Math.
J. (2005), no. 17, 89–103. 17

[VW08] Emanuele Viola and Avi Wigderson, Norms, XOR lemmas, and lower bounds for poly-
nomials and protocols, Theory Comput. 4 (2008), 137–168. MR 2480238 (2009m:68106)
11

[Yan07] Ke Yang, On the (im)possibility of non-interactive correlation distillation, Theoret.

Comput. Sci. 382 (2007), no. 2, 157–166. MR 2352111 (2009b:68011) 25

35