Sticky Brownian Rounding and its Applications to

Constraint Satisfaction Problems

SEPEHR ABBASI-ZADEH, University of Toronto
NIKHIL BANSAL, TU Eindhoven, and Centrum Wiskunde & Informatica
GURU GURUGANESH, Google Research
ALEKSANDAR NIKOLOV, University of Toronto
ROY SCHWARTZ, Technion
MOHIT SINGH, Georgia Institute of Technology

Semidefinite programming is a powerful tool in the design and analysis of approximation algorithms for
combinatorial optimization problems. In particular, the random hyperplane rounding method of Goemans
and Williamson [31] has been extensively studied for more than two decades, resulting in various extensions
to the original technique and beautiful algorithms for a wide range of applications. Despite the fact that this
approach yields tight approximation guarantees for some problems, e.g., Max-Cut, for many others, e.g.,
Max-SAT and Max-DiCut, the tight approximation ratio is still unknown. One of the main reasons for this
is the fact that very few techniques for rounding semi-definite relaxations are known.
In this work, we present a new general and simple method for rounding semi-definite programs, based on
Brownian motion. Our approach is inspired by recent results in algorithmic discrepancy theory. We develop 33
and present tools for analyzing our new rounding algorithms, utilizing mathematical machinery from the
theory of Brownian motion, complex analysis, and partial differential equations. Focusing on constraint sat-
isfaction problems, we apply our method to several classical problems, including Max-Cut, Max-2SAT, and
Max-DiCut, and derive new algorithms that are competitive with the best known results. To illustrate the
versatility and general applicability of our approach, we give new approximation algorithms for the Max-Cut
problem with side constraints that crucially utilizes measure concentration results for the Sticky Brownian
Motion, a feature missing from hyperplane rounding and its generalizations.
CCS Concepts: • Theory of computation → Rounding techniques; Random walks and Markov chains;
Additional Key Words and Phrases: Semidefinite programming, Brownian motion, constraint satisfaction
problems

SA and AN were supported by an NSERC Discovery Grant (RGPIN-2016-06333). NB was supported by the NWO VICI grant
639.023.812. MS was supported by NSF grant CCF-BSF:AF1717947. Some of this work was carried out while AN and NB
were visitors at the Simons Institute program on Bridging Discrete and Continuous Optimization, partially supported by
NSF grant #CCF-1740425.
Authors’ addresses: S. Abbasi-Zadeh and A. Nikolov, Sandford Fleming 2301B, University of Toronto, 10 King’s
College Rd, Toronto, ON M5S 3G4, Canada; emails: [email protected], [email protected]; N. Bansal,
TU Eindhoven, and Centrum Wiskunde & Informatica, Science Park 123, 1098 XG Amsterdam, Netherlands; email:
[email protected]; G. Guruganesh, 1600 Amphitheatre Parkway Mountain View, CA 94043 Google Research, USA; email:
[email protected]; R. Schwartz, Taub Building (Room 521) Technion Israel Institute of Technology Haifa 3200003 Israel;
email: [email protected]; M. Singh, H. Milton Stewart School of Industrial and Systems Engineering. 755 Ferst
Drive, NW, Atlanta, GA 30332 Georgia Institute of Technology, USA; email: [email protected].
Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee
provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and
the full citation on the first page. Copyrights for components of this work owned by others than the author(s) must be
honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists,
requires prior specific permission and/or a fee. Request permissions from [email protected].
© 2022 Copyright held by the owner/author(s). Publication rights licensed to ACM.
1549-6325/2022/10-ART33 $15.00
https://doi.org/10.1145/3459096

ACM Transactions on Algorithms, Vol. 18, No. 4, Article 33. Publication date: October 2022.
33:2 S. Abbasi-Zadeh et al.

ACM Reference format:

Sepehr Abbasi-Zadeh, Nikhil Bansal, Guru Guruganesh, Aleksandar Nikolov, Roy Schwartz, and Mohit Singh.
2022. Sticky Brownian Rounding and its Applications to Constraint Satisfaction Problems. ACM Trans. Algo-
rithms 18, 4, Article 33 (October 2022), 50 pages.
https://doi.org/10.1145/3459096

1 INTRODUCTION
Semi-definite programming (SDP) is one of the most powerful tools in the design of approxima-
tion algorithms for combinatorial optimization problems. Semi-definite programs can be viewed as
relaxed quadratic programs whose variables are allowed to be vectors instead of scalars and scalar
multiplication is replaced by inner products between the vectors. The prominent approach when
designing SDP-based approximation algorithms is rounding: (1) An SDP relaxation is formulated
for the given problem, (2) The SDP relaxation is solved, and lastly, (3) The fractional solution for
the SDP relaxation is transformed into a feasible integral solution to the original problem, hence
the term rounding.
In their seminal work, Goemans and Williamson [31] presented an elegant and remarkably sim-
ple rounding method for SDPs: A uniformly random hyperplane (through the origin) is chosen,
and then each variable, which is a vector, is assigned to the side of the hyperplane it belongs to.
This (binary) assignment is used to round the vectors and output an integral solution. For example,
when considering Max-Cut, each side of the hyperplane corresponds to a different side of the cut.
Using the random hyperplane rounding, [31] gave the first non-trivial approximation guarantees
for fundamental problems such as Max-Cut, Max-2SAT, and Max-DiCut. Perhaps the most cele-
brated result of [31] is the 0.878 approximation for Max-Cut, which is known to be tight [39, 44]
assuming Khot’s Unique Games Conjecture [38]. Since then, the random hyperplane method has
inspired, for more than two decades now, a large body of research, both in approximation algo-
rithms and in hardness of approximation. In particular, many extensions and generalizations of
the random hyperplane rounding method have been proposed and applied to a wide range of ap-
plications, e.g., Max-DiCut and Max-2SAT [25, 40], Max-SAT [8, 13], Max-Bisection [11, 50],
Max-Agreement in correlation clustering [21], and the Cut-Norm of a matrix [3].
Despite this success and the significant work on variants and extensions of the random hyper-
plane method, the best possible approximation ratios for many fundamental problems still remain
elusive. Several such examples include Max-SAT, Max-Bisection, Max-2CSP, and Max-DiCut.
Perhaps the most crucial reason for the above failure is the fact that besides the random hyperplane
method and its variants, very few methods for rounding SDPs are known.
A sequence of papers by Austrin [10], Raghavendra [48], Raghavendra and Steurer [49] has
shown that SDP rounding algorithms that are based on the random hyperplane method and its
extensions nearly match the Unique Games hardness of any Max-CSP, as well as the integrality gap
of a natural family of SDP relaxations. However, the universal rounding proposed by Raghavendra
and Steurer is impractical, as it involves a brute-force search on a large constant-sized instance of
the problem. Moreover, their methods only allow computing an ε additive approximation to the
approximation ratio in time double-exponential in 1/ε.

1.1 Our Results and Techniques

Our main contributions are (1) to propose a new SDP rounding technique that is based on diffusion
processes, and, in particular, on Brownian motion; (2) to develop the needed tools for analyzing
our new SDP rounding technique by deploying a variety of mathematical techniques from proba-
bility theory, complex analysis, and partial differential equations (PDEs); (3) to show that this

ACM Transactions on Algorithms, Vol. 18, No. 4, Article 33. Publication date: October 2022.
Sticky Brownian Rounding and its Applications to Constraint Satisfaction Problems 33:3

rounding technique has useful concentration of measure properties, not present in random hyper-
plane based techniques, that can be used to obtain new approximation algorithms for a version of
the Max-Cut problem with multiple global side constraints.
Our method is inspired by the recent success of Brownian motion-based algorithms for construc-
tive discrepancy minimization, where it was used to give the first constructive proofs of some of the
most powerful results in discrepancy theory [14–16, 41]. The basic idea is to use the solution to the
semi-definite program to define the starting point and the covariance matrix of the diffusion pro-
cess, and let the process evolve until it reaches an integral solution. As the process is forced to stay
inside the cube [−1, 1]n (for Max-Cut) or [0, 1]n (for Max-2SAT and other problems), and to stick
to any face it reaches, we call the most basic version of our algorithm (without any enhancements)
the Sticky Brownian Motion rounding. The algorithm is defined more formally in Section 1.2.1.
Sticky Brownian Motion. Using the tools we introduce, we show that this algorithm is already
competitive with the state-of-the-art results for Max-Cut, Max-2SAT, and Max-DiCut.
Theorem 1. The basic Brownian rounding achieves an approximation ration of 0.861 for the Max-
Cut problem.√Moreover, when the Max-Cut instance has value 1−ε, Sticky Brownian Motion achieves
value 1 − Ω( ε).
In particular, using complex analysis and evaluating various elliptic integrals, we show that the
separation probability for any two unit vectors u and v separated by an angle θ , is given by a
certain hypergeometric function of θ (see Theorem 4 for details). This precise characterization of
the separation probability also proves that the Sticky Brownian Motion rounding is different from
the random hyperplane rounding. The overview of the analysis is in Sections 1.2.2 and 2 has the
details.
We can also analytically show the following upper bound for Max-2SAT.
Theorem 2. The Sticky Brownian Motion rounding achieves approximation ratio of at least 0.8749
for Max-2SAT.
While the complex analysis methods also give exact results for Max-2SAT, the explicit expres-
sions are much harder to obtain as one has to consider all possible starting points for the diffusion
process, while in the Max-Cut case, the process always starts at the origin. Because of this, in
order to prove Theorem 2 we introduce another method of analysis based on PDEs, and the maxi-
mum principle, which allows us to prove analytic bounds on PDE solutions. Moreover, numerically
solving the PDEs suggests the bound 0.921. The overview and details of the Max-2SAT analysis are,
respectively, in Sections 1.2.3 and 3. Section 5 has details about numerical calculations for various
problems.
For comparison, the best known approximation ratio for Max-Cut is the Goemans–
Williamson (GW) constant αGW ≈ 0.878, and the best known approximation ratio for Max-2SAT
is 0.94016 [40]. The result for Max-Cut instances of value 1 − ε is optimal up to constants [39],
assuming the Unique Games Conjecture.
We emphasize that our results above are achieved with a single algorithm “out of the box”,
without any additional engineering. While the analysis uses sophisticated mathematical tools, the
algorithm itself is simple, efficient, and straightforward to implement.
Extensions. Next, we consider two different modifications of Sticky Brownian Motion that
allow us to improve the approximation guarantees above, and show the flexibility of diffusion
based rounding algorithms. The first one is to smoothly slow down the process depending on
how far it is from the boundaries of the cube. As a proof of concept, we show, numerically, that a
simple modification of this kind matches the GW approximation of Max-Cut up to the first three

ACM Transactions on Algorithms, Vol. 18, No. 4, Article 33. Publication date: October 2022.
33:4 S. Abbasi-Zadeh et al.
Table 1. Approximation Ratios for Sticky Brownian Motion Rounding and
Sticky Brownian Motion with Slowdown

Algorithm Max-Cut Max-2SAT Max-DiCut 

Brownian Rounding 0.861 0.921 0.79
Brownian with Slowdown 0.878 † 0.929 0.81
† indicates that for Max-Cut, the approximation for the slowed down walk differs from
the GW bound only in the fourth decimal. For Max-DiCut, the  indicates that we only
consider the n + 1-dimensional walk.

digits after the decimal point. We also obtain significant improvements for other problems over
the vanilla method.
Second, we propose a variant of Sticky Brownian Motion running in n + 1 dimensions rather
than n dimensions, and we analyze it for the Max-DiCut problem. The extra dimension is used
to determine whether the nodes labeled +1 or those labeled −1 are put on the left side of the cut.
We show that this modification achieves an approximation ratio of 0.79 for Max-DiCut. Slowing
down the process further improves this approximation to 0.81. We give a summary of the obtained
results1 in Table 1. Both of these are natural methods to try. Intuition and details of the extensions
are given, respectively, in Sections 1.2.4 and 5.
Recent Progress. Very recently, in a beautiful result, Eldan and Naor [24] describe a slowdown
process that exactly achieves the GW bound of 0.878 for Max-Cut, answering an open question
posed in an earlier version of this paper. This shows that our rounding techniques are at least
as powerful as the classical random hyperplane rounding, and are potentially more general and
flexible.
In general, given the dearth of techniques for rounding semidefinite programs, we expect that
rounding methods based on diffusion processes, together with the analysis techniques introduced
in this article, will find broader use, and, perhaps lead to improved results for Max-CSP problems.
Applications. To further illustrate the versatility and general applicability of our approach, we
consider the Max-Cut with Side Constraints problem, abbreviated Max-Cut-SC, a generalization
of the Max-Bisection problem, which allows for multiple global constraints. In an instance of the
Max-Cut-SC problem, we are given an n-vertex graph G = (V , E), a collection F = {F 1 , . . . , Fk }
of subsets of V , and cardinality bounds b1 , . . . , bk ∈ N. The goal is to find a subset S ⊂ V that
maximizes the weight |δ (S )| of edges crossing the cut (S, V \ S ), subject to having |S ∩ Fi | = bi for
all i ∈ [k].
Since even checking whether there is a feasible solution is NP-hard [22], we aim for bi-criteria
approximation algorithms.2 We give the following result for the problem, using the Sticky Brown-
ian Motion as a building tool.
Theorem 3. There exists a O (npoly(log(k )/ε ) )-time algorithm that on input a satisfiable instance
G = (V , E), F , and b1 , . . . , bk , as defined above, outputs a (0.843 − ε, ε)-approximation with high
probability.
In the presence of a single side constraint, the problem is closely related to the Max-Bisection
problem [11, 50], and, more generally to Max-Cut with a cardinality constraint. While our
1 Our numerical results are not obtained via simulating the random algorithm but solving a discretized version of a PDE
that analyzes the performance of the algorithm. Error analysis of such a discretization can allow us to prove the correctness
of these bounds within a reasonable accuracy.
2 We say that a set S ⊂ V is an (α, ε )-approximation if  |S ∩ F | − b  ≤ εn for all i ∈ [k ], and |δ (S ) | ≥ α · |δ (T ) | for all
 i i
T ⊂ V such that |T ∩ F i | = bi for all i ∈ [k ].

ACM Transactions on Algorithms, Vol. 18, No. 4, Article 33. Publication date: October 2022.
Sticky Brownian Rounding and its Applications to Constraint Satisfaction Problems 33:5

methods use the stronger semi-definite programs considered in [50] and [11], the main new tech-
nical ingredient is showing that the Sticky Brownian Motion possesses concentration of measure
properties that allow us to approximately satisfy multiple constraints. By contrast, the hyperplane
rounding and its generalizations that have been applied previously to the Max-Cut and Max-
Bisection problems do not seem to allow for such strong concentration bounds. For this reason,
the rounding and analysis used in [50] only give an O (npoly(k /ε ) ) time algorithm for the Max-Cut-
SC problem, which is trivial for k = Ω(n), whereas our algorithm has non-trivial quasi-polynomial
running time even in this regime. We expect that this concentration of measure property will find
further applications, in particular, to constraint satisfaction problems with global constraints.
Remark. We can achieve better results using Sticky Brownian Motion with slowdown. In par-
ticular, in time O (npoly(log(k )/ε ) ) we can get a (0.858 − ε, ε)-approximation with high probability
for any satisfiable instance.3 However, we focus on the basic Sticky Brownian Motion algorithm
to simplify exposition. Note that due to the recent work by Austrin and Stanković [12], we know
that adding even a single global cardinality constraint to the Max-Cut problem makes it harder to
approximate. In particular, they show that subject to a single side constraint, Max-Cut is Unique
Games-hard to approximate within a factor of approximately 0.858. Thus, assuming the Unique
Games conjecture, our approximation factor for the Max-Cut-SC problem is close to optimal up
to small numerical errors. (We emphasize the possibility of numerical errors as both our result,
and the hardness result in [12] are based on numerical calculations.) Whether these techniques
can result in an optimal approximation ratio remains an interesting open question.

1.2 Overview
1.2.1 The Sticky Brownian Motion Algorithm. Let us describe our basic algorithm in some detail.
Recall that the GW- SDP for Max-Cut is equivalent to the following vector program: given a graph
G = (V , E), we write
 1 − wi · wj
max
2
(i, j ) ∈E
s.t . wi · wi = 1 ∀i ∈ V ,
where the variables wi range over n-dimensional real vectors (n = |V |). The Sticky Brownian
Motion rounding algorithm we propose maintains a sequence of random fractional solutions
X0 , . . . , XT such that X0 = 0 and XT ∈ {−1, +1}n is integral. Here, a vertex of the hypercube
{−1, +1}n is naturally identified with a cut, with vertices assigned +1 forming one side of the cut,
and the ones assigned −1 forming the other side.
Let At be the random set of coordinates of Xt −1 , which are not equal to −1 or +1; we call these
coordinates active. At each time step t = 1, . . . ,T , the algorithm picks ΔXt sampled from the
Gaussian distribution with mean 0 and covariance matrix Wt , where (Wt )i j = wi · wj if i, j ∈ At ,
and (Wt )i j = 0 otherwise. The algorithm then takes a small step in the direction of ΔXt , i.e., sets
Xt = Xt −1 +γ ΔXt for some small fixed real number γ (which can be as small a 1/poly(n)). If the ith
coordinate of Xt is very close to −1 or +1 for some i, then it is rounded to either −1 or +1, whichever
is closer. The parameters γ and T are chosen so that the fractional solutions Xt never leave the
cube [−1, 1]n , and so that the final solution XT is integral with high probability. As γ goes to 0, the
trajectory of the ith coordinate of Xt closely approximates a Brownian motion started at 0, and
stopped when it hits one of the boundary values {−1, +1}. Importantly, the trajectories of different
coordinates are correlated according to the SDP solution. A precise definition of the algorithm is
given in Section 2.2.
3 The 0.858-result is achieved using numerical simulation and likely doesn’t match the hardness result exactly.

ACM Transactions on Algorithms, Vol. 18, No. 4, Article 33. Publication date: October 2022.
33:6 S. Abbasi-Zadeh et al.

The algorithm for Max-2SAT (and Max-DiCut) is essentially the same, modulo using the co-
variance matrix from the appropriate standard SDP relaxation, and starting the process at the
marginals for the corresponding variables. We explain this in greater detail below.

1.2.2 Overview of the Analysis for Max-Cut. In order to analyze this algorithm, it is sufficient to
understand the probability that an edge (i, j) is cut as a function of the angle θ between the vectors
wi and wj . Thus, we can focus on the projection ((Xt )i , (Xt )j ) of Xt . We observe that ((Xt )i , (Xt )j )
behaves like a discretization of correlated two-dimensional Brownian motion started at (0, 0), until
the first time τ when it hits the boundary of the square [−1, 1]2 . After τ , ((Xt )i , (Xt )j ) behaves
like a discretization of a one-dimensional Brownian motion restricted to one of the sides of the
square. From now on we will treat the process as being continuous, and ignore the discretization,
which only adds an arbitrarily small error term in our analysis. It is convenient to apply a linear
transformation to the correlated Brownian motion ((Xt )i , (Xt )j ) so that it behaves like a standard
two-dimensional Brownian motion Bt started at (0, 0). We show that this linear transformation
maps the square [−1, 1]2 to a rhombus S centered at 0 with internal angle θ ; we can then think of
τ as the first time Bt hits the boundary of S. After time τ , the transformed process is distributed
like a one-dimensional Brownian motion on the side of the rhombus that was first hit. To analyze
this process, we need to understand the probability distribution of Bτ . The probability measure
associated with this distribution is known as the harmonic measure on the boundary ∂S of S, with
respect to the starting point 0. These transformations and connections are explained in detail in
Section 2.3.
The harmonic measure has been extensively studied in probability theory and analysis. The
simplest special case is the harmonic measure on the boundary of a disc centered at 0 with respect
to the starting point 0. Indeed, the central symmetry of the disc and the Brownian motion implies
that it is just the uniform measure. A central fact we use is that harmonic measure in two dimen-
sions is preserved under conformal (i.e., angle-preserving) maps. Moreover, such maps between
polygons and the unit disc have been constructed explicitly using complex analysis, and, in par-
ticular, are given by the Schwarz–Christoffel formula [2]. Thus, the Schwarz–Christoffel formula
gives us an explicit formulation of sampling from the harmonic measure on the boundary ∂S of
the rhombus: it is equivalent to sampling a uniformly random point on the boundary of the unit
disc D centered at the origin, and mapping this point via a conformal map F that sends D to S.
Using this formulation, in Section 2.4, we show how to write the probability of cutting the edge
(i, j) as an elliptic integral.
Calculating the exact value of elliptic integrals is a challenging problem. Nevertheless, by ex-
ploiting the symmetry in the Max-Cut objective, we relate our particular elliptic integral to in-
tegrals of the incomplete beta and hypergeometric functions. We further simplify these integrals
and bring them into a tractable form using several key identities from the theory of special func-
tions. Putting everything together, we get a precise closed form expression for the probability that
the Sticky Brownian Motion algorithm cuts a given edge in Theorem 4, and, as a consequence, we
obtain the claimed guarantees for Max-Cut in Theorems 1 and 7.

1.2.3 Overview of the Analysis for Max-2SAT. The algorithm for Max-2SAT is almost identical
to the Max-Cut algorithm, except that the SDP solution is asymmetric, in the following sense.
We can think of the SDP as describing the mean and covariance of a “pseudo-distribution” over
the assignments to the variables. In the case of Max-Cut, we could assume that, without loss
of generality, the mean of each variable (i.e., one-dimensional marginal) is 0 since S and S are
equivalent solutions. However, this is not the case for Max-2SAT. We use this information, and
instead of starting the diffusion process at the center of the cube, we start it at the point given

ACM Transactions on Algorithms, Vol. 18, No. 4, Article 33. Publication date: October 2022.
Sticky Brownian Rounding and its Applications to Constraint Satisfaction Problems 33:7

by the marginals. For convenience, and also respecting standard convention, we work in the cube
[0, 1]n rather than [−1, 1]n . Here, in the final solution XT , if (XT )i = 0 we set the ith variable to
true and if (XT )i = 1, we set it to false. We again analyze each clause C separately, which allows
us to focus on the diffusion process projected to the coordinates ((X t )i , (X t )j ), where i and j are
the variables appearing in C. However, the previous approach of using the Schwarz–Christoffel
formula to obtain precise bounds on the probability does not easily go through, since it relies
heavily on the symmetry of the starting point of the Brownian motion. It is not clear how to
extend the analysis when we change the starting point to a point other than the center, as the
corresponding elliptic integrals appear to be intractable.
Instead, we appeal to a classical connection between diffusion processes and partial differential
equations [47, Chapter 9]. Recall that we are focusing on a single clause C with variables i and j,
and the corresponding diffusion process ((X t )i , (X t )j ) in the unit square [0, 1]2 starting at a point
given by the marginals and stopped at the first time τ when it hits the boundary of the square; after
that time the process continues as a one-dimensional Brownian motion on the side of the square
it first hit. For simplicity let us assume that both variables appear un-negated in C. The probability
that C is satisfied then equals the probability that the process ends at one of the points (0, 1), (1, 0),
or (0, 0). Let u : [0, 1]2 → [0, 1] be the function, which assigns to (x, y) the probability that this
happens when the process is started at (x, y). Since on the boundary ∂[0, 1]2 of the square our
process is a one-dimensional martingale, the value of u (x, y) is easy to compute on ∂[0, 1]2 , and in
fact equals 1 −xy. Then, in the interior of the square, we have u (x, y) = E[u ((X τ )i , (X τ )j )]. It turns
out that this identifies u as the unique solution to an elliptic partial differential equation (PDE)
Lu = 0 with the Dirichlet boundary condition u (x, y) = 1 − xy ∀(x, y) ∈ ∂[0, 1]2 . In our case,
∂2u ∂2u
the operator L just corresponds to Laplace’s operator L[u] = ∂x 2 + ∂y 2 after applying a linear

transformation to the variables and the domain. This connection between our rounding algorithm
and PDEs is explained in Section 3.2.
Unfortunately, it is still not straightforward to solve the obtained PDE analytically. We deal with
this difficulty using two natural approaches. First, we use the maximum principle of elliptic PDE’s
[28], which allows us to bound the function u from below. In particular, if we can find a function д
such that д(x, y) ≤ u (x, y) = 1 −xy on the boundary of the square, and Lд ≥ 0 in the interior, then
the maximum principle tells us that д(x, y) ≤ u (x, y) for all x, y in the square. We exhibit simple
low-degree polynomials which satisfy the boundary conditions by design, and use the sum of
squares (SoS) proof system to certify non-negativity under the operator L. In Section 3.3, we use
this method to show that Sticky Brownian Motion rounding achieves approximation ratio at least
0.8749.
Our second approach is to solve the PDE numerically to a high degree of accuracy using finite
element methods. We use this approach in Section 5 to numerically obtain results showing a 0.921
approximation ratio for Max-2SAT.

1.2.4 Extensions of Sticky Brownian Motion.

Using different slowdown functions. Recall that in the Sticky Brownian Motion rounding each in-
crement is proportional to ΔXt sampled from a Gaussian distribution with mean 0 and covariance
matrix Wt . The covariance is derived from the SDP: for example, in the case of Max-Cut, it is
initially set to be the Gram matrix of the vectors produced by the SDP solution. Then, whenever a
coordinate (Xt )i reaches {−1, +1}, we simply zero-out the corresponding row and column of Wt .
This process can be easily modified by varying how the covariance matrix Wt evolves with time.
Instead of zeroing out rows and columns of Wt , we can smoothly scale them based on how far
(Xt −1 )i is from the boundary values {−1, 1}. A simple way to do this, in the case of the Max-Cut

ACM Transactions on Algorithms, Vol. 18, No. 4, Article 33. Publication date: October 2022.
33:8 S. Abbasi-Zadeh et al.

problem, is to set
(Wt )i j = (1 − (Xt −1 )i2 ) α /2 (1 − Xt −1 )j2 ) α /2 wi · wj ,
for a constant 0 ≤ α < 2. Effectively, this means that the process is slowed down smoothly as it
approaches the boundary of the cube [−1, +1]n .
Intuition: Until both endpoints are not fixed, each step of the Brownian motion preserves the
probability of cutting an edge (in expectation). Therefore, slowing down the process is an attempt
to ensure that all vertices will hit the boundary at the same time. This tries to ensure the probability
of cutting an edge to be closer to the actual expectation. For the specific slowdown function, the
slowdown corresponds to scaling the vectors appropriately to match their length in the original sdp
solution. In particular, the SDP vectorsvi can be written as vi = x i v 0 + 
w i where w i is orthogonal
projection to v 0⊥ . The length of w i is x i − x i2 in 0–1 sdp version and 1 − x i2 in the ±1 version.
Since the w i gives the correlations, this naturally suggests that choose slowdown as a function of

1 − x i2 .
This modified diffusion process, which we call Sticky Brownian Motion with Slowdown, still
converges to {−1, +1}n in finite time. Once again, the probability of cutting an edge (i, j) of our
input graph can be analyzed by focusing on the two-dimensional projection ((Xt )i , (Xt )j ) of Xt .
Moreover, we can still use the general connection between diffusion processes and PDE’s men-
tioned above. That is, if we write u (x, y) : [−1, 1]2 → [0, 1] for the probability that edge (i, j) is
cut if the process is started at (x, y), then u can be characterized as the solution of an elliptic PDE
1−xy
with boundary conditions u (x, y) = 2 ∀(x, y) ∈ ∂[−1, 1]2 . We solve this PDE numerically
using the finite element method to estimate the approximation ratio for a fixed value of the pa-
rameter α, and then we optimize over α. At the value α =1.61 our numerical solution shows an
approximation ratio that matches the GW approximation of Max-Cut up to the first three digits
after the decimal point. We also analyze an analogous algorithm for Max-2SAT and show that
for α =1.61 it achieves an approximation ratio of 0.929. The detailed analysis of the slowed down
Sticky Brownian Motion rounding is given in Section 5.
A higher-dimensional version. We also consider a higher-dimensional version of the Sticky Brow-
nian Motion rounding, in which the Brownian motion evolves in n + 1 dimensions rather than n.
This rounding is useful for asymmetric problems like Max-DiCut4 in which the SDP produces non-
uniform marginals, as we discussed above in the context of Max-2SAT. Such an SDP has a vector
w0 in addition to w1 , . . . , wn , and the marginals are given by w0 · wi . Now, rather than using the
marginals to obtain a different starting point, we consider the symmetric Sticky Brownian Motion
process starting from the center but using all the n +1 vectors w0 , . . . , wn . At the final step T of the
process, in the case of Max-DiCut, the variables whose value is equal to (XT )0 are assigned to the
left side of the cut, and the variables with the opposite value are assigned to the right side of the
cut. Thus, for an edge i → j to be cut, it must be the case that (XT )i = (XT )0 and (XT )j = 1 − (XT )0 .
While analyzing the probability that this happens is a question about Brownian motion in three
rather than two dimensions, we reduce it to a two-dimensional question via the inclusion-exclusion
principle. After this reduction, we can calculate the probability that an edge is cut by using the
exact formula proved earlier for the Max-Cut problem. Our analysis, which is given in Section 6,
shows that this (n + 1)-dimensional Sticky Brownian Motion achieves an approximation of 0.79
for Max-DiCut. Moreover, combining the two ideas, of changing the covariance matrix at each
step, as well as performing the n + 1-dimensional Sticky Brownian Motion, achieves a ratio of 0.81.

4 Theinput for Max-DiCut is a directed graph G = (V , E ), and the goal is to find a cut S ⊆ V that maximizes the number
of edges going from S to S .

ACM Transactions on Algorithms, Vol. 18, No. 4, Article 33. Publication date: October 2022.
Sticky Brownian Rounding and its Applications to Constraint Satisfaction Problems 33:9

1.2.5 Overview of the Analysis for Max-Cut-SC. The starting point for our algorithm for the
Max-Cut-SC problem is a stronger SDP relaxation derived using the SoS hierarchy. Similar re-
laxations were previously considered in [11, 50] for the Max-Bisection problem. In addition to
giving marginal values and a covariance matrix for a “pseudo-distribution” over feasible solutions,
the SoS SDP makes it possible to condition on small sets of variables. The global correlation round-
ing method [17, 32] allows us to choose variables to condition on so that, after the conditioning,
the covariance matrix has small entries on average. Differing from previous works [11, 50], we
then run the Sticky Brownian Motion rounding defined by the resulting marginals and covariance
matrix. We can analyze the weight of cut edges using the PDE approach outlined above. The main
new challenge is to bound the amount by which the side constraints are violated. To do so, we
show that Sticky Brownian Motion concentrates tightly around its mean, and, in particular, it sat-
isfies sub-Gaussian concentration in directions corresponding to sets of vertices. Since the mean
of the Sticky Brownian Motion is given by the marginals, which satisfy all side constraints, we
can bound how much constraints are violated via the concentration and a union bound. To show
this key concentration property, we use the fact that the covariance that defines the diffusion has
small entries, and that Brownian Motion is a martingale. Then the concentration inequality fol-
lows, roughly, from a continuous analogue of Azuma’s inequality. The detailed analysis is given in
Section 4. We again remark that such sub-Gaussian concentration bounds are not known to hold
for the random hyperplane rounding method or its generalizations as considered in [11, 50].

1.3 Related Work

In their seminal work, Goemans and Williamson [31] presented the random hyperplane rounding
method, which yielded an approximation of 0.878 for Max-Cut. For the closely related Max-DiCut
problem they presented an approximation of 0.796. This was subsequently improved in a sequence
of papers: Feige and Goemans [25] presented an approximation of 0.859; Matuura and Matsui
improved the factor to 0.863; and culminating in the work of Lewin et al. [40] who present the
current best known approximation of 0.874, getting close to the 0.878 approximation of [31] for
Max-Cut. Another fundamental and closely related problem is Max-Bisection. In their classic
work [27], Frieze and Jerrum present an approximation of 0.651 for this problem. Their result was
later improved to 0.699 by Ye [53], to 0.701 by Halperin and Zwick [34], and to 0.702 by Feige and
Langberg [26]. Using the SoS hierarchy, Raghavendra and Tan [50] gave a further improvement
to 0.85, and finally, Austrin et al. [11] presented an almost tight approximation of 0.8776. With
respect to hardness results, Håstad [35] proved a hardness of 16/17 for Max-Cut (which implies the
exact same hardness for Max-Bisection) and a hardness of 11/12 for Max-DiCut (both of these
hardness results are assuming P  NP). If one assumes the Unique Games Conjecture of Khot [38],
then it is known that the random hyperplane rounding algorithm of [31] is tight [39, 44]. Thus,
it is worth noting that though Max-Cut is settled conditional on the Unique Games conjecture,
both Max-DiCut and Max-Bisection still remain unresolved, even conditionally.
Another fundamental class of closely related problems are Max-SAT and its special cases Max-
k-SAT. For Max-2SAT Goemans and Williamson [31], using random hyperplane rounding, pre-
sented an approximation of 0.878. This was subsequently improved in a sequence of works: Feige
and Goemans [25] presented an approximation of 0.931; Matsui and Matuura [42] improved the
approximation factor to 0.935; and finally Lewin et al. [40] presented the current best known ap-
proximation of 0.94016. Regarding hardness results for Max-2SAT, assuming P  NP, Håstad [35]
presented a hardness of 21/22. Assuming the Unique Games Conjecture Austrin [9] presented a
(virtually) tight hardness of 0.94016, matching the algorithm of [40]. For Max-3SAT, Karloff and
Zwick [37] and Zwick [54] presented an approximation factor of 7/8 based on the random hyper-
plane method. The latter is known to be tight by the celebrated hardness result of Håstad [35].

ACM Transactions on Algorithms, Vol. 18, No. 4, Article 33. Publication date: October 2022.
33:10 S. Abbasi-Zadeh et al.

For Max-4SAT Halperin and Zwick [33] presented an (almost) tight approximation guarantee of
0.8721. When considering Max-SAT in its full generality, a sequence of works [7, 8, 13] slowly
improved the known approximation factor, where the current best one is achieved by Avidor et al.
[13] and equals 0.797.5 For the general case of Max-CSP a sequence of works [10, 48] culminated in
the work of Raghavendra and Steurer [49] who presented an algorithm that assuming the Unique
Games Conjecture matches the hardness result for any constraint satisfaction problem. However,
as previously mentioned, this universal rounding is impractical as it involves a brute-force solu-
tion to a large constant instance of the problem. Moreover, it only allows computing an ε additive
approximation to the approximation ratio in time double-exponential in 1/ε .
Many additional applications of random hyperplane rounding and its extensions exist. Some
well known examples include: 3-Coloring [5, 20, 36], Max-Agreement in correlation clustering
[21, 51], the maximization of quadratic forms [23], and the computation of the Cut-Norm [3].
Let us now briefly focus on the extensions and generalizations of random hyperplane round-
ing. The vast majority of the above mentioned works use different extensions of the basic random
hyperplane rounding. Some notable examples include: rotation of the vectors [3, 45, 53, 55], pro-
jections [23, 26, 46], and combining projections with clustering [10, 48, 49]. It is worth noting that
the above extensions and generalizations of the basic random hyperplane method are not the only
approaches known for rounding SDPs. The most notable example of the latter is the seminal work
of Arora et al. [6] for the Sparsest-CUT problem. Though the approach of [6] uses random projec-
tions, it is based on different mathematical tools, e.g., Lévy’s isoperimetric inequality. Moreover,
the algorithmic machinery that was developed since the work of [6] has found uses for minimiza-
tion problems, and in particular it is useful for minimization problems that relate to graph cuts and
clustering.
Brownian motion was first used for rounding SDPs in Bansal [14] in the context of constructive
discrepancy minimization. This approach has since proved itself very successful in this area, and
has led to new constructive proofs of several major results [15, 16, 41]. However, this line of work
has largely focused on improving logarithmic factors, and its methods are not precise enough to
analyze constant factor approximation ratios.

2 BROWNIAN ROUNDING FOR MAX-CUT VIA CONFORMAL MAPPINGS

2.1 Prerequisites: Definition of Brownian Motion
For completeness, we recall the definition of standard Brownian motion.
Definition 1. A stochastic process {Bt }t ≥0 taking values in Rn is called an n-dimensional Brow-
nian motion started at x ∈ Rn if
— B0 = x;
— the process has independent increments, i.e., for all N and all times 0 ≤ t 1 ≤ t 2 ≤ · · · ≤ t N ,
the increments Bt N − Bt N −1 , Bt N −1 − Bt N −2 , . . . , Bt2 − Bt1 are independent random variables;
— for all t ≥ 0 and all h > 0, the increment Bt +h − Bt is distributed as a Gaussian random
variable with mean 0 and covariance matrix equal to the scaled identity, h · In ;
— the function f (t ) = Bt is almost surely continuous.
The process {Bt }t ≥0 is called standard Brownian motion if x = 0.
The fact that this definition is not empty, i.e., that such a stochastic process exists, is non-trivial.
The first rigorous proof of this fact was given by Wiener [52]. We refer the reader to the book [43]
for a thorough introduction to Brownian motion and its properties.
5 Avidor et al. also present an algorithm with a conjectured approximation of 0.8434, refer to [13] for the exact details.

ACM Transactions on Algorithms, Vol. 18, No. 4, Article 33. Publication date: October 2022.
Sticky Brownian Rounding and its Applications to Constraint Satisfaction Problems 33:11

In this section, we use Max-Cut as a case study for the method of rounding a semi-definite
relaxation via Sticky Brownian Motion. Recall, in an instance of the Max-Cut problem we are
given a graph G = (V , E) with edge weights a : E → R+ and the goal is to find a subset S ⊂ V that

maximizes the total weight of edges crossing the cut (S, V \S ), i.e., a(δ (S )) := {u,v } ∈E:u ∈S,vS auv .
We first introduce the standard semi-definite relaxation for the problem and the sticky Brownian
rounding algorithm. To analyze the algorithm, we use the invariance of Brownian motion with
respect to conformal maps, along with several identities of special functions.

2.2 SDP Relaxation and Sticky Brownian Rounding Algorithm

Before we proceed, we recall again the SDP formulation for the Max-Cut problem, famously stud-
ied by Goemans and Williamson [31].
 (1 − wi · wj )
max a(e)
2
e=(i, j ) ∈E
s.t . wi · wi = 1 ∀i = 1, . . . , n.
We now describe the Sticky Brownian Motion rounding algorithm specialized to the Max-
Cut problem. Let W denote the positive semi-definite correlation matrix defined by the vectors
w1 , . . . , wn , i.e., for every 1 ≤ i, j ≤ n we have that: Wi, j = wi · wj . Given a solution W to the semi-
definite program, we perform the following rounding process: start at the origin and perform a
Brownian motion inside the [−1, 1]n hypercube whose correlations are governed by W. Addition-
ally, the random walk is sticky: once a coordinate reaches either −1 or +1 it is fixed and does not
change anymore.
Formally, we define a random process {Xt }t ≥0 as follows. We fix X0 = 0. Let {Bt }t ≥0 be standard
Brownian motion in Rn started at the origin,6 and let τ1 = inf {t : x0 + W1/2 Bt  [−1, 1]n } be the
first time x0 + W1/2 Bt exits the cube. With probability 1, you can assume that τ1 is also the first
time that the process lies on the boundary of the cube. Here W1/2 is the principle square root of W.
Then, for all 0 ≤ t ≤ τ1 we define
Xt = x0 + W /2 Bt .
1

This defines the process until the first time it hits a face of the cube. From this point on, we will
force it to stick to this face. Let At = {i : (Xt )i  ±1} be the active coordinates of the process at
time t, and let Ft = {x ∈ [−1, 1]n : x i = (Xt )i ∀i ∈ At } be the face of the cube on which Xt lies at
time t. With probability 1, Fτ1 has dimension n − 1. We define the covariance matrix (Wt )i j = Wi j
when i, j ∈ At , and (Wt )i j = 0 otherwise. Then, we take τ2 = inf {t ≥ τ1 : Xτ1 +Wτ/12 (Bt −Bτ1 )  Fτ1 }
1

to be the first time that Brownian motion started at Xτ1 with covariance given by Wτ1 exits the
face Fτ1 . Again, with probability 1, we can assume that this is also the first time the process lies on
the boundary of Fτ1 . For all τ1 < t ≤ τ2 we define
Xt = Xτ1 + Wτ/12 (Bt − Bτ1 ).
1

Again, with probability 1, dim Fτ2 = n − 2. The process is defined analogously from here on. In
general, τi = inf {t ≥ τi−1 : Xτi −1 + Wτ/i2−1 (Bt − Bτi −1 )  Fτi −1 } is (with probability (1) the first
1

time that the process hits a face of the cube of dimension n − i. Then for τi−1 < t ≤ τi we have
Xt = Xτi −1 + Wτ/i2−1 (Bt − Bτi −1 ). At time τn , Xτn ∈ {−1, 1}n , so the process remains fixed, i.e., for any
1

t ≥ τn , Xt = Xτn . The output of the algorithm then corresponds to a cut S ⊆ V defined as follows:
   
S = i ∈ V : Xτn i = 1 .
We say that a pair of nodes {i, j} is separated when |S ∩ {i, j}| = 1.
6 We will always assume that a standard Brownian motion starts at the origin. See Appendix 2.1 for a precise definition.

ACM Transactions on Algorithms, Vol. 18, No. 4, Article 33. Publication date: October 2022.
33:12 S. Abbasi-Zadeh et al.

Remark. While we have defined the algorithm as a continuous diffusion process, driven by Brow-
nian motion, a standard discretization will yield a polynomial time algorithm that achieves the
same guarantee up to an error that is polynomially small. Such a discretization was outlined in
the Introduction. An analysis of the error incurred by discretizing a continuous diffusion process
in this way can be found, for example, in [29] or the book [30]. More sophisticated discrete sim-
ulations of such diffusion processes are also available, and can lead to better time complexity as
a function of the error. One example is the Walk on Spheres algorithm analyzed by Binder and
Braverman [19]. This algorithm allows us to draw a sample Xτ from the continuous diffusion pro-
cess, stopped at a random time τ , such that Xτ is within distance ε from the boundary of the cube
[−1, 1]n . The time necessary to sample Xτ is polynomial in n and log(1/ε). We can then round Xτ
to the nearest point on the boundary of the cube, and continue the simulation starting from this
rounded point. It is straightforward to show, using the fact that the probability to cut an edge is
continuous in the starting point of our process, that if we set ε = o(n−1 ), then the approximation
ratio achieved by this simulation is within an o(1) factor from the one achieved by the continuous
process. In the rest of the article, we focus on the continuous process since our methods of analysis
are naturally amenable to it.

2.3 Analysis of the Algorithm

Our aim is to analyze the expected value of the cut output by the Sticky Brownian Motion round-
ing algorithm. Following Goemans and Williamson [31], we aim at bounding the probability an
edge is cut as compared to its contribution to the SDP objective. Theorem 4 below gives an exact
characterization of the probability of separating a pair of vertices {i, j} in terms of the gamma func-
tion and hypergeometric functions. We refer to Appendix A.1 for the definitions of these functions
and a detailed exposition of their basic properties.
Theorem 4. The probability that the Sticky Brownian Motion rounding algorithm will separate a
pair {i, j} of vertices for which θ = cos −1 (wi · wj ) equals
 1+a 1+a a 
Γ( a+1 ) 2 , 2 , 2
1 − 1−a 2a · 3 F 2 ;1 ,
2, 2 + 1
a a
Γ( 2 )Γ( 2 + 1) 2

where a = θ/π , Γ is the gamma function, and 3 F 2 is the hypergeometric function.

Theorem 1 will now follow from the following corollary of Theorem 4. The corollary follows
from numerical estimates of the gamma and hypergeometric functions.
Corollary 1. For any pair {i, j}, the probability that the pair {i, j} is separated is at least
1−w ·w
0.861· 2i j .
We now give an outline of the proof of Theorem 4. The plan is to first show that the desired
probability can be obtained by analyzing the two-dimensional standard Brownian motion starting
at the center of a rhombus. Moreover, the probability of separating i and j can be computed using
the distribution of the first point on the boundary that is hit by the Brownian motion. Conformal
mapping and, in particular, the Schwarz–Christoffel formula, allows us to obtain a precise expres-
sion for such a distribution and thus for the separation probability, as claimed in the theorem. We
now expand on the above plan.
First observe that to obtain the probability that i and j are separated, it is enough to consider
the two-dimensional process obtained by projecting to the ith and jth coordinates of the vector Xt .
Projecting the process onto these coordinates, we obtain a process X̃t ∈ R2 that can be equivalently
defined as follows. Let  
1 cos(θ )
W̃ = ,
cos(θ ) 1

ACM Transactions on Algorithms, Vol. 18, No. 4, Article 33. Publication date: October 2022.
Sticky Brownian Rounding and its Applications to Constraint Satisfaction Problems 33:13

where θ is the angle between wi and wj . Let Bt be standard Brownian motion in R2 started at
0, and let τ = inf {t : W̃1/2 Bt  [−1, 1]2 } be the first time the process hits the boundary of the
square. Then for all 0 ≤ t ≤ τ we define X̃t = W̃1/2 Bt . Any coordinate k for which (X̃τ )k ∈ {±1}
remains fixed from then on, i.e., for all t > τ , (X̃t )k = (X̃τ )k . The coordinate  that is not fixed
at time τ (one exists with probability 1) continues to perform one-dimensional Brownian motion
started from (X̃τ ) until it hits −1 or +1, at which point it also becomes fixed. Let σ be the time
this happens; it is easy to show that σ < ∞ with probability 1, and, moreover, E[σ ] < ∞. We say
that the process {X̃t }t ≥0 is absorbed at the vertex X̃σ ∈ {−1, 1}2 .
Observation 1. The probability that the algorithm separates vertices i and j equals
 
Pr {X̃t }t is absorbed in {(+1, −1), (−1, +1)} .

With an abuse of notation, we denote X̃t by Xt and W̃ by W for the rest of the section, which
is aimed at analyzing the above probability. We also denote by ρ = cos(θ ) the correlation between
the two coordinates of the random walk, and call the two-dimensional process just described a ρ-
correlated walk. It is easier to bound the probability that i and j are separated by transforming the
ρ-correlated walk inside [−1, 1]2 into a standard Brownian motion inside an appropriately scaled
rhombus. We do this by transforming {Xt }t ≥0 linearly into an auxiliary random process {Yt }t ≥0
which will be sticky inside a rhombus (see Figure 1(a)–(b)). Formally, given the random process
{Xt }t ≥0 , we consider the process Yt = O · W−1/2 · Xt , where O is a rotation matrix to be chosen
shortly. Recalling that for 0 ≤ t ≤ τ the process {Xt }0≤t ≤τ is distributed as {W1/2 Bt }0≤t ≤τ , we have
that, for all 0 ≤ t ≤ τ ,
Yt = O · Bt ≡ Bt .
Above ≡ denotes equality in distribution, and follows from the invariance of Brownian motion
under rotation. Applying OW−1/2 to the points inside [−1, 1]2 , we get a rhombus S with vertices
b1 , . . . , b4 , which are the images of the points (+1, −1), (+1, +1), (−1, +1), (−1, −1), respectively.
We choose O so that b1 lies on the positive x-axis and b2 on the positive y-axis. Since OW−1/2 is a
linear transformation, it maps the interior of [−1, 1]2 to the interior of S and the sides of [−1, 1]2
to the sides of S. We have then that τ is the first time Yt hits the boundary of S, and that after
this time Yt sticks to the side of S that it first hit and evolves as (a scaling of) one-dimensional
Brownian motion restricted to this side, and started at Yτ . The process then stops evolving at the
time σ when Yσ ∈ {b1 , . . . , b4 }. We say that {Yt }t ≥0 is absorbed at Yσ .
The following lemma, whose proof appears in the appendix, formalizes the main facts we use
about this transformation.
Lemma 1. Applying the transformation OW−1/2 to {Xt }t ≥0 , we get a new random process {Yt }t ≥0
which has the following properties:
(1) If Xt is in the interior/boundary/vertex of [−1, 1]2 then Yt is in the interior/boundary/vertex of
S, respectively.
(2) S is a rhombus whose internal angles at b1 and b3 are θ , and at b2 and b4 are π − θ . The vertex
b1 lies on the positive x-axis, and b2 , b3 , b4 are arranged counter-clockwise.
(3) The probability that the algorithm will separate the pair {i, j} is exactly
Pr[Yt is absorbed in b1 or b3 ].
In the following useful lemma we show that, in order to compute the probability that the process
{Yt }t ≥0 is absorbed in b1 or b3 , it suffices to determine the distribution of the first point Yτ on the
boundary ∂S that the process {Yt }t ≥0 hits. This distribution is a probability measure on ∂S known
in the literature as the harmonic measure (with respect to the starting point 0). We denote it by μ ∂S .
The statement of the lemma follows.

ACM Transactions on Algorithms, Vol. 18, No. 4, Article 33. Publication date: October 2022.
33:14 S. Abbasi-Zadeh et al.

Fig. 1. Figure (a) depicts {Xt }t ≥0 in the [−1, 1]2 square, Figure (b) depicts {Yt }t ≥0 in the rhombus S, and
Figure (c) depicts {Bt }t ≥0 in the unit disc D. The linear transformation W−1/2 transforms the [−1, 1]2 square
to S (Figure (a) to Figure (b)), whereas the conformal mapping Fθ transforms D to S (Figure (c) to Figure (b)).

Lemma 2.
b2
p − b1 
Pr[Yt is absorbed in b1 or b3 ] = 4 · 1− dμ ∂S (p).
b1 b2 − b1 
Proof. Since both S and Brownian motion are symmetric with respect to reflection around the
coordinate axes, we see that μ ∂S is the same as we go from b1 to b2 or b4 , and as we go from b3 to
b2 or b4 . Therefore,
Pr[pair {i, j} is separated] = 4 · Pr[pair {i, j} is separated | Yτ lies on the segment [b1 , b2 ]].
The process {Yt }τ ≤t ≤σ is a one-dimensional martingale, so E[Yσ |Yτ ] = Yτ by the optional
stopping theorem [43, Proposition 2.4.2]. If we also condition on Yτ ∈ [b1 , b2 ], we have that
Yσ ∈ {b1 , b2 }. An easy calculation then shows that the probability of being absorbed in b1 con-
ditional on Yτ and on the event Yτ ∈ [b1 , b2 ] is exactly Y τ −b 2  Yτ −b1 
b2 −b1  = 1 − b2 −b1  . Then,
 
Yτ − b1  b2
p − b1 
Pr[pair {i, j} is separated | Yτ ∈ [b1 , b2 ]] = E 1 − = 1− dμ ∂S (p).
b2 − b1  b1 b2 − b1 
This proves the lemma. 

To obtain the harmonic measure directly for the rhombus S we appeal to conformal mappings.
We use the fact that the harmonic measure can be defined for any simply connected region U in the
plane with 0 in its interior. More precisely, let Bt be standard two-dimensional Brownian motion
started at 0, and τ (U ) = inf {t : Bt  U } be the first time it hits the boundary of U . Then μ ∂U
denotes the probability measure induced by the distribution of Bτ (U ) , and is called the harmonic
measure on ∂U (with respect to 0). When U is the unit disc centered at 0, the harmonic measure
is uniform on its boundary because Brownian motion is invariant under rotation. Then the main
idea is to use conformal maps to relate harmonic measures on the different domains, namely, the
disc and our rhombus S.

2.4 Conformal Mapping

Before we proceed further, it is best to transition to the language of complex numbers and identify
R2 with the complex plane C. A complex function F : U → V where U , V ⊆ C is conformal if
it is holomorphic (i.e. complex differentiable) and its derivative f  (x )  0 for all x ∈ U . The key
fact we use about conformal maps is that they preserve harmonic measure. Below we present this
theorem from Mörters and Peres [43] specialized to our setting. In what follows, D will be the unit
disc in C centered at 0.

ACM Transactions on Algorithms, Vol. 18, No. 4, Article 33. Publication date: October 2022.
Sticky Brownian Rounding and its Applications to Constraint Satisfaction Problems 33:15

Theorem 5 ([43, p. 204, Theorem 7.23]). Suppose Fθ is a conformal map from the unit disk D to
S. Let μ ∂D and μ ∂S be the harmonic measures with respect to 0. Then μ ∂D ◦ Fθ−1 = μ ∂S .
Thus the above theorem implies that in our setting, the probability that a standard Brownian
motion will first hit any segment S of the boundary of D is the same as the probability of the
standard Brownian motion first hitting its image under Fθ , i.e., Fθ (S ) in ∂S.
To complete the picture, the Schwarz–Christoffel formula gives a conformal mapping from the
unit disc D to S that we utilize.
Theorem 6 ([2, Theorem 5, Section 2.2.2]). Define the function Fθ (ω) by
ω ω
Fθ (ω) = fθ (s)ds = (1 − s) −(1−θ /π ) (1 + s) −(1−θ /π ) (s − i) −θ /π (s + i) −θ /π ds.
s=0 s=0
Then, for some real number c > 0, cFθ (ω) is a conformal map from the unit-disk D to the rhombus S.
The conformal map has some important properties, which will aid us in calculating the prob-
abilities. We collect them in the following lemma, which follows from standard properties of the
Schwarz–Christoffel integral [2], and is easily verified.
Lemma 3. The conformal map cFθ (ω) has the following properties:
(1) The four points located at {1, i, −1, −i} map to the four vertices {b1 , . . . , b4 } of the rhombus S,
respectively.
(2) The origin maps to the origin.
(3) The boundary of the unit-disk D maps to the boundary of S. Furthermore, the points in the arc
from 1 to i map to the segment [b1 , b2 ].
Define the function r : [0, π /2] → R as r (ϕ) := |Fθ (e iϕ ) − Fθ (1)|.
Lemma 4. The probability that vertices {i, j} are separated, given that the angle between wi and
wj is θ , is
π /2
2 r (ϕ)
1− dϕ.
π 0 r (π /2)
Proof. Rewriting the expression in Lemma 2 in complex number notation, we have
b2
|z − b1 | b2
|z − cFθ (1)|
Pr[{i, j} separated] = 4 · 1− dμ ∂S (z) = 4 · 1− dμ ∂S (z).
b1 |b2 − b1 | b1 c |Fθ (i) − Fθ (1)|
Since the conformal map Fθ preserves the harmonic measure between the rhombus S and the
unit-disk D (see Theorem 5) and by Lemma 3, the segment from b1 to b2 is the image of the arc
from 1 to i under cFθ , we can rewrite the above as
π /2
|cFθ (e iϕ ) − cFθ (1)|
=4· 1− dμ ∂D (e iϕ ).
0 c |Fθ (i) − Fθ (1)|
The harmonic measure μ ∂D on the unit-disk is uniform due to the rotational symmetry of Brow-
nian motion.
π /2
|cFθ (e iϕ ) − cFθ (1)| dϕ
=4· 1− .
0 c |Fθ (i) − Fθ (1)| 2π
Simplifying the above, we see that the right hand side above equals
π /2 π /2
2 |Fθ (e iϕ ) − Fθ (1)| 2 r (ϕ)
· 1− dϕ = · 1− dϕ.
π 0 |F θ (i) − F θ (1)| π 0 r (π /2)
This completes the proof. 

ACM Transactions on Algorithms, Vol. 18, No. 4, Article 33. Publication date: October 2022.
33:16 S. Abbasi-Zadeh et al.

To calculate the approximation ratio exactly, we will make use of the theory of special functions.
While these calculations are technical, they are not trivial. To aid the reader, we give a brief primer
in Appendix A.1 and refer them to the work of Andrews et al. [4], Beals and Wong [18] for a more
thorough introduction.
The proof of Theorem 4 will follow from the following key claims whose proofs appear in the
appendix. Letting a = θ/π and b = 1 − a, we have
Claim 1.
1
r (ϕ) = β sin2 ϕ (a/2, b/2),
4
when ϕ ∈ [0, π /2].

Claim 2.  1+a 
π /2
β (a/2 + 1/2, 1/2) 2 , 2 , 2
1+a a
4· r (ϕ)dϕ = · 3 F2 ;1 .
2, 2 + 1
a a a
0

2.5 Asymptotic Calculation for θ Close to π

We consider the case when the angle θ = (1 − ϵ ) · π as ϵ → 0. The hyperplane-rounding algorithm
separates such an edge by θ/π , and hence has a separation probability of 1 − ϵ. We show a similar
asymptotic behaviour for the Brownian rounding algorithm, albeit with slightly worse constants.
We defer the proof to the appendix.
Theorem 7. Given an edge {i, j} with cos −1 (wTi wj ) = θ = (1 − ϵ )π , the Sticky Brownian Motion
rounding will cut the edge with probability at least 1 − ( π4 ϵ + O (ϵ 2 )).

3 BROWNIAN ROUNDING FOR MAX-2SAT VIA PARTIAL DIFFERENTIAL

EQUATIONS
In this section, we use Max-2SAT as a case study for extending the Sticky Brownian Motion round-
ing method to other constraint satisfaction problems besides Max-Cut. In the Max-2SAT problem
we are given n variables z 1 , . . . , zn and m clauses C 1 , . . . , Cm , where the j th clause is of the form
y j1 ∨ y j2 (y j is a literal of z j , i.e., z j or z j ). The goal is to assign to each variable zi a value of true or
false so as to maximize the number of satisfied clauses.

3.1 Semi-definite Relaxation and Brownian Rounding Algorithm

The standard SDP relaxation (see [54]) used for Max-2SAT is the following:

m
max 1 − vj1 · vj2 ,
j=1
s.t . v0 · v0 = 1, (1)
v0 · vi = vi · vi ∀i = −n, . . . , n, (2)
vi · v−i = 0 ∀i = 1, . . . , n, (3)
v0 · (vi + v−i ) = 1 ∀i = 1, . . . , n, (4)
1 ≥ v0 · vi + vj · v0 − vi · vj ∀i, j = −n, . . . , n, (5)
vi · v0 ≥ vi · vj ∀i, j = −n, . . . , n, (6)
vi · vj ≥ 0 ∀i, j = −n. . . . , n. (7)

ACM Transactions on Algorithms, Vol. 18, No. 4, Article 33. Publication date: October 2022.
Sticky Brownian Rounding and its Applications to Constraint Satisfaction Problems 33:17

In the above, v0 is a unit vector that denotes the false assignment (constraint 1), whereas a zero
vector denotes the true assignment. We use the standard notation that vi denotes the literal zi and
v−i denotes the literal z i . Therefore, vi · v−i = 0 for every i = 1, . . . .n (constraints 3 and 4) since
zi needs to be either true or false. The remainder of the constraints (constraints 5, 6 and 7) are
equivalent to the 22 triangle inequalities over all triples of vectors that include v0 .
When trying to generalize the Brownian rounding algorithm for Max-Cut presented in Sec-
tion 2 to Max-2SAT, there is a problem: unlike Max-Cut the Max-2SAT problem is not symmetric.
Specifically, for Max-Cut both S and S are equivalent solutions having the same objective value.
However, for Max-2SAT an assignment to the variables z 1 = α 1 , . . . , zn = α n is not equivalent to
the assignment z 1 = α 1 , . . . , zn = α n (here α i ∈ {0, 1} and α i = 1 ⊕ α i ). For example, if vi · v0 = 1
then we would like the Brownian rounding algorithm to always assign zi to false. The Brownian
rounding for Max-Cut cannot handle such a requirement. In order to tackle the above problem
we incorporate v0 into both the starting point of the Brownian motion and the covariance matrix.
Let us now formally define the Brownian rounding algorithm for Max-2SAT. For simplicity of
presentation denote for every i = 1, . . . , n by x i the marginal value of zi , formally: x i := vi · v0 .
Additionally, let wi be the (unique) unit vector in the  direction of the projection of vi to the sub-
space orthogonal to v0 , i.e., wi satisfies vi = x i v0 + x i − x i2 wi .7 Similarly to Max-Cut, our Sticky
Brownian Motion rounding algorithm performs a random walk in Rn , where the i th coordinate
corresponds to the variable zi . For simplicity of presentation, the random walk is defined in [0, 1]n
as opposed to [±1]n , where 1 denotes false and 0 denotes true.8 Unlike Max-Cut, the starting point
X0 is not the center of the cube. Instead, we use the marginals, and set (X0 )i := x i . The covariance
matrix W is defined by Wi, j := wi · wj for every i, j = 1, . . . , n, and similarly to Max-Cut, let
W1/2 be the principle square root of W. Letting {Bt }t ≥0 denote standard Brownian motion in Rn ,
we define τ1 = inf {t : W1/2 Bt + X0  [0, 1]n } to be the first time the process hits the boundary of
[0, 1]n . Then, for all times 0 ≤ t ≤ τ1 , the process Xt is defined as
Xt = W /2 Bt + X0 .
1

After time τ1 , we force Xt to stick to the face F 1 hit at time τ1 : i.e., if (Xτ1 )i ∈ {0, 1}, then we fix it
forever, by zeroing out the ith row and column of the covariance matrix of W for all future time
steps. The rest of the process is defined analogously to the one for Max-Cut: whenever Xt hits a
lower dimensional face of [0, 1]n , it is forced to stick to it until finally a vertex is reached, at which
point Xt stops changing. We use τi for the first time that Xt hits a face of dimension n − i; then,
Xτn ∈ {0, 1}n .
The output of the algorithm corresponds to the collection of the variables assigned a value of
true T ⊆ {1, . . . , n}:
 
T = {i : Xτn i = 0},
whereas implicitly the collection of variables assigned a value of false are {i : (Xτn )i = 1}.

3.2 Analysis of the Algorithm

Our goal is to analyze the expected value of the assignment produced by the Sticky Brownian
Motion rounding algorithm. Similarly to previous work, we aim at giving a lower bound on the
probability that a fixed clause C is satisfied. Unfortunately, the conformal mapping approach de-
scribed in Section 2 does not seem to be easily applicable to the extended Sticky Brownian Motion

is easy to see that x −i = 1 − x i and w−i = −wi for every i = 1, . . . , n.

7 It
8 We note that the Brownian rounding algorithm for Max-2SAT can be equivalently defined in [−1, 1]n , however, this will
incur some overhead in the notations which we would like to avoid.

ACM Transactions on Algorithms, Vol. 18, No. 4, Article 33. Publication date: October 2022.
33:18 S. Abbasi-Zadeh et al.

rounding described above for Max-2SAT, because our calculations for Max-Cut relied heavily on
the symmetry of the starting point of the random walk. We propose a different method for an-
alyzing the Brownian rounding algorithm that is based on partial differential equations and the
maximum principle. We prove analytically the following theorem which gives a guarantee on the
performance of the algorithm. We also note that numerical calculations show that the algorithm
in fact achieves the better approximation ratio of 0.921 (see Section 5 for details).
Theorem 8. The Sticky Brownian Motion rounding algorithm for Max-2SAT achieves an approxi-
mation of at least 0.8749.
3.2.1 Analysis via Partial Differential Equations and Maximum Principle. As mentioned above,
our analysis focuses on the probability that a single clause C with variables {zi , z j } is satisfied.
We assume the variables are not negated. This is without loss of generality as the algorithm and
analysis are identical in the case of negated variable. This is easy to see as v−i = (1 − x )v0 −
√
x − x 2 wi is defined almost identially to vi .
√ we denote by x the marginal value of zi and by y the marginal value
For simplicity of notation
of z j . Thus, vi = xv0 + x − x 2 wi and vj = yv0 + y − y 2 wj . Projecting the random process
{X}t ≥0 on the i and j coordinates of the random process, we obtain a new process {X̃t }t ≥0 where
X̃0 = (x, y). Let
 
1 cos(θ )
W̃ = ,
cos(θ ) 1
where θ is the angle between wi and wj . Then X̃t = X̃0 + W̃1/2 Bt for all 0 ≤ t ≤ τ , where
τ = inf {t : X̃0 + W̃1/2 Bt  [0, 1]2 } is the first time the process hits the boundary of the square.
After time τ , the process X̃t performs a one-dimensional standard Brownian motion on the first
side of the square it has hit, until it hits a vertex at some time σ . After time σ the process stays
fixed. Almost surely σ < ∞, and, moreover, it is easy to show that Eσ < ∞. We say that {X̃t }t ≥0
is absorbed at X̃σ ∈ {0, 1}2 .
Observation 2. The probability that the algorithm satisfies the clause {zi , z j } equals
Pr X̃σ is absorbed in {(0, 0), (0, 1), (1, 0)} .
We abuse notation slightly and denote X̃t by Xt and W̃ by W for the rest of the section which
is aimed at analyzing the above probability. We also denote ρ = cos(θ ).
Our next step is fixing θ and analyzing the probability of satisfying the clause for all possible
values of marginals x and y. Indeed, for different x and y but the same θ , the analysis only needs to
consider the same random process with a different starting point. Observe that not all such x, y are
necessarily feasible for the SDP: we characterize which ones are feasible for a given θ in Lemma 7.
But considering all x, y allows us to handle the probability in Observation 2 analytically.
For any 0 ≤ x ≤ 1, 0 ≤ y ≤ 1, let u (x, y) denote the probability of starting the random walk at
the point (x, y) and ending at one of the corners (0, 0), (0, 1), or (1, 0). This captures the probability
of a clause being satisfied when the walk begins with marginals (x, y) (and angle θ ). We can easily
calculate this probability exactly when either x or y are in the set {0, 1}. We obtain the following
easy lemma whose proof appears in the appendix.
Lemma 5. For ϕ (x, y) = 1 − xy, we have
u (x) = ϕ (x) for all x ∈ ∂[0, 1]2 . (8)
Moreover, for all x in the interior of the square [0, 1]2 , u (x) = Ex [ϕ (Xτ )], where Ex denotes expectation
with respect to starting the process at X0 = x.

ACM Transactions on Algorithms, Vol. 18, No. 4, Article 33. Publication date: October 2022.
Sticky Brownian Rounding and its Applications to Constraint Satisfaction Problems 33:19

Next we use the fact that Brownian motion gives a solution to the Dirichlet boundary problem.
While Brownian motion gives a solution to Laplace’s equation ([43] chapter 3), since our random
process is a diffusion process, we need a slightly more general result.9 We state the following result
from [47], specialized to our setting, that basically states that given a diffusion process in [0, 1]2
and a function ϕ on the boundary, the extension of the function defined on the interior by the
expected value of the function at the first hitting point on the boundary is characterized by an
elliptic partial differential equation.
Theorem 9 ([47] Theorem 9.2.14). Let D = (0, 1) 2 ⊆ R2 , Σ ∈ R2×2 and let a 11 , a 12 , a 21 , a 22 be
defined as follows:  
a 11 a 12 1
= ΣΣ .
a 21 a 22 2
For any x ∈ D, consider the process Xt = X0 + ΣBt where Bt is standard Brownian motion in R2 .
Let τ = inf {t : Xt  D}. Given a bounded continuous function ϕ : ∂D → R, define the function
u : D → R such that

u (x) = Ex ϕ (Xτ ) ,
where Ex denotes the expected value when X0 = x ∈ R2 .For example, u (x) is the expected value
of ϕ when first hitting ∂D conditioned on starting at point x. Consider the uniformly elliptic partial
differential operator L in D defined by

2
∂2
L= ai j .
i, j=1
∂x i ∂x j

Then u ∈ C 2 (D) is the unique solution to the partial differential equation10 :

Lu = 0 in D
lim u (x) = ϕ (y) for all y ∈ ∂D.
x→y
x∈D

We instantiate our differential equation by choosing Σ = W1/2 and thus ai j are the entries of
W. It is important to note that all ai j ’s are independent of the starting point x ∈ [0, 1]2 . Thus, we
obtain that u is the unique function satisfying the following partial differential equation:
∂ 2u ∂ 2u ∂ 2u
+ 2 + 2ρ =0 ∀(x, y) ∈ Int[0, 1]2 ,
∂x 2 ∂y ∂x ∂y
u (x, y) = 1 − xy ∀(x, y) ∈ ∂[0, 1]2 .
Above, and in the rest of the article, we use Int D to denote the interior of a set D, and ∂D to denote
its boundary.
It remains to solve the above PDE that will allow us to calculate u (x, y) and give the probability
of satisfying the clause.

3.3 Maximum Principle

Finding closed form solutions to general PDE’s is challenging and, there is no guarantee any so-
lution would be expressible in terms of simple functions. However, to find a good approximation
ratio, it suffices for us to find good lower-bounds on the probability of satisfying the clause. For ex-
ample, we need to give a lower bound on the function u (x, y) from the previous section over those
(x, y) that are feasible. Since the PDE’s generated by our algorithm are elliptic (a particular kind of
9 This result can also be derived from Theorem 3.12 in [43] after applying a linear transformation to the variables.
10 u ∈ C k (D ) means that u has a continuous k th derivative over D, and u ∈ C 0 (D ) means that u is continuous.

ACM Transactions on Algorithms, Vol. 18, No. 4, Article 33. Publication date: October 2022.
33:20 S. Abbasi-Zadeh et al.

PDE), we will use a property of elliptic PDE’s which will allow us to produce good lower-bounds
on the solution at any given point. More precisely, we use the following theorem from Gilbarg and
Trudinger [28].
Let L denote the operator
 ∂2
L := ai j ,
ij
∂i ∂j

and we say that L is an elliptic operator if the coefficient matrix A = [ai j ]i, j is positive semi-
definite.
We restate a version of Theorem 3.1 in Gilbarg and Trudinger [28] that shows how the maximum
principle can be used to obtain lower bounds on u (x, y). Here, D̄ denotes the closure of D.

Theorem 10 (Maximum Principle). Let L be elliptic on a bounded domain D and suppose

L[д](x ) ≥ 0 ∀x ∈ D for some д ∈ C 2 (D) ∩ C 0 (D̄). Then the maximum of д on D is achieved
on ∂D, that is,
sup д(x ) = sup д(x ).
x ∈D x ∈∂D

Theorem 10 has the following corollary that allows us to obtain lower bounds on u (x, y).

Corollary 2. Let L be elliptic on a bounded domain D and for some u, д ∈ C 2 (D) ∩ C 0 (D̄).
(1) L[д](x ) ≥ L[u](x ) ∀x ∈ D,
(2) д(x ) ≤ u (x ) ∀x ∈ ∂D,
then, д(x ) ≤ u (x )∀x ∈ D.

We refer the reader to [28] for a formal proof. Recall that u satisfies the system L[u] = 0. Thus,
it is enough to construct candidate functions д : [0, 1]2 → R such that

∂ 2д ∂ 2д ∂ 2д
+ 2 + 2ρ ≥0 ∀(x, y) ∈ Int[0, 1]2 , (9)
∂x 2 ∂y ∂x ∂y
д(x, y) ≤ (1 − xy) ∀(x, y) ∈ ∂[0, 1]2 . (10)

Then we obtain that д(x, y) ≤ u (x, y) for all (x, y) ∈ [0, 1]2 . In what follows we construct many dif-
ferent such functions each of which works for a different range of the parameter θ (equivalently, ρ).

3.4 Candidate Functions for Maximum Principle

We now construct feasible candidates to the maximum principle as described in Corollary 2. We
define the following functions:
√
(1) д1 (x, y) = 1 − xy − cos(θ ) x − x 2 y − y 2 .
(2) д2 (x, y) = 1 − xy − 2 cos(θ )(x − x 2 )(y − y 2 ).
(3) д3 (x, y) = 1 − xy − 12 (1 + 5 cos(θ ))(x − x 2 )(y − y 2 )(x + y)(2 − x − y).
The following lemma shows that the above functions satisfy the conditions required for the
application of the maximum principle (its proof appears in the appendix).

Lemma 6. Each of д1 , д2 , д3 satisfies the boundary conditions, i.e., дi (x, y) = u (x, y) for all x, y ∈
∂[0, 1]2 and for all values θ . Moreover, we have the following for each (x, y) ∈ [0, 1]2 :

ACM Transactions on Algorithms, Vol. 18, No. 4, Article 33. Publication date: October 2022.
Sticky Brownian Rounding and its Applications to Constraint Satisfaction Problems 33:21

(1) If 1 ≥ cos(θ ) ≥ 0, then Lд1 ≥ 0.

(2) If 0 ≥ cos(θ ) ≥ − 12 , then Lд2 ≥ 0.
(3) If − 12 ≥ cos(θ ) ≥ −1, then Lд3 ≥ 0.
While some of these proofs are based on simple inequalities, proving others requires us to use
SoS expressions. For example, to show Lд3 ≥ 0, we consider Lд3 = p(x, y, cos(θ )) as a polynomial
in x, y and cos(θ ). Replacing z = cos(θ ), our aim is to show p(x, y, z) ≥ 0 if 0 ≤ x, y ≤ 1 and
−1 ≤ z ≤ − 12 . Equivalently, we need to show p(x, y, z) ≥ 0 whenever r 1 (x, y, z) := x − x 2 ≥ 0,
r 2 (x, y, z) := y − y 2 ≥ 0 and r 3 (x, y, z) := −(z + 12 ) ≥ 0 and r 4 (x, y, z) := (z + 1) ≥ 0. We show
this by obtaining polynomials qi (x, y, z) for i = 0, 1, 2, 3, 4 such that each qi is a SoS polynomial of
fixed degree and we have

4
p(x, y, z) = q 0 (x, y, z) + qi (x, y, z)r i (x, y, z).
i=1
Observe that the above polynomial equality proves the desired result by evaluating the RHS for
every 0 ≤ x, y ≤ 1 and −1/2 ≥ z ≥ −1. Clearly, the RHS is non-negative: each qi is non-negative
since it is a SoS, and each r i is non-negative in the region we care about, by construction. We
mention that we obtain these proofs via solving a semi-definite program of fixed degree (at most
6) for each of the qi polynomials (missing details appear in the appendix).
Let us now focus on the approximation guarantee that can be proved using the above functions
д1 , д2 , and д3 . The following lemma compares the lower bounds on the probability of satisfying a
clause, as given by д1 , д2 , and д3 , to the SDP objective. Recall that the contribution
√ of any clause
with marginals x and y and angle θ to the SDP’s objective is given by: 1−xy−cos(θ ) x − x 2 y − y 2 .
We denote this contribution by SDP(x, y, θ ). It is important to note that not all triples (x, y, θ ) are
feasible (recall that θ is the angle between wi and wj ), due to the triangle inequalities in the SDP.
This is summarized in the following lemma.
Lemma 7. Let x, y, θ be as defined by a feasible pair of vectors vi and v j . Then they must satisfy
the following constraints:
(1) 0 ≤ x ≤ 1, 0≤ y ≤ 1, 0 ≤ θ ≤ π .
xy
(2) cos(θ ) ≥ − (1−x )(1−y ) .

(1−x )(1−y )
(3) cos(θ ) ≥ − xy .
Finally, we prove the following lemma which proves an approximation guarantee of 0.8749 for
Max-2SAT via the PDE and the maximum principle approach. As before, these proofs rely on
explicitly obtaining SoS proofs as discussed above. We remark that these proofs essentially aim to
obtain 78 = 0.875-approximation but errors of the order 10−5 allow us to obtain a slightly worse
bound using this methods. The details appear in the appendix.
Lemma 8. Consider any feasible triple (x, y, θ ) satisfying the condition in Lemma 7. We have the
following.
(1) If 1 ≥ cos(θ ) ≥ 0, then д1 (x, y) ≥ 1 · SDP(x, y, θ ).
(2) If 0 ≥ cos(θ ) ≥ − 12 , then д2 (x, y) ≥0.8749·SDP(x, y, θ ).
(3) If − 12 ≥ cos(θ ) ≥ −1, then д3 (x, y) ≥0.8749·SDP(x, y, θ ).

4 MAX-CUT WITH SIDE CONSTRAINTS (MAX-CUT-SC)

In this section, we describe how to apply the Sticky Brownian Motion rounding and the framework
of Raghavendra and Tan [50] to the Max-Cut-SC problem in order to give a bi-criteria approxi-
mation algorithm whose running time is non-trivial even when the the number of constraints is
ACM Transactions on Algorithms, Vol. 18, No. 4, Article 33. Publication date: October 2022.
33:22 S. Abbasi-Zadeh et al.

small. The main novelty is that our framework can handle all these constraints “out of the box”,
while it is not clear how to analyze the basic random hyperplane rounding algorithm or any other
algorithm.

4.1 Problem Definition and Basics

Let us recall the relevant notation and definitions. An instance of the Max-Cut-SC problem is
given by an n-vertex graph G = (V , E) with edge weights a : E → R+ , as well as a collection
F = {F 1 , . . . , Fk } of subsets of V , and cardinality bounds b1 , . . . , bk ∈ N. For ease of notation, we

will assume that V = {1, . . . , n}. Moreover, we denote the total edge weight by a(E) = e ∈E a(e).
The goal in the Max-Cut-SC problem is to find a subset S ⊂ V that maximizes the weight a(δ (S ))
of edges crossing the cut (S, V \ S ), subject to having |S ∩ Fi | = bi for all i ∈ [k]. These cardinality
constraints may not be simultaneously satisfiable, and moreover, when k grows with n, checking
satisfiability is NP-hard [22]. For these reasons, we allow for approximately feasible solutions. We
will say that a set of vertices S ⊆ V is an (α, ε)-approximation to the Max-Cut-SC problem if
 |S ∩ F | − b  ≤ εn for all i ∈ [k], and a(δ (S )) ≥ α · a(δ (T )) for all T ⊂ V such that |T ∩ F | = b
 i i i i
for all i ∈ [k]. In the remainder of this section we assume that the instance given by G, F , and b
is satisfiable, i.e., that there exists a set of vertices T such that |T ∩ Fi | = bi for all i ∈ [k]. Our
algorithm may fail if this assumption is not satisfied. If this happens, then the algorithm will certify
that the instance was not satisfiable.
We start with a simple baseline approximation algorithm, based on independent rounding. The
algorithm outputs an approximately feasible solution which cuts a constant fraction of the total
edge weight. For this reason, it achieves a good bi-criteria approximation when the value of the
optimal solution OPT is much smaller than εa(E). This allows us to focus on the case in which
OPT is bigger than εa(E) for our main rounding algorithm. The proof of the lemma, which follows
from standard arguments, appears in the appendix.
/ε )
Lemma 9. Suppose that n ≥ 2 ln(8k ε2
and ε ≤ 12 . There exists a polynomial time algorithm that on
input a satisfiable instance G = (V , E), F , and b1 , . . . , bk , as defined above, outputs a set S ⊆ V such
that, with high probability, a(δ (S )) ≥ ε2 a(E), and  |S ∩ Fi | − bi  ≤ εn for all i ∈ [k].

4.2 Sum of Squares Relaxation

Our main approximation algorithm is based on a semidefinite relaxation, and the sticky Brownian
motion. Let us suppose that we are given the optimal objective value OPT of a feasible solution:
this assumption can be removed by doing binary search for OPT. We can then model the problem
of finding an optimal feasible solution by the quadratic program

a(e)(x i − x j ) 2 ≥ OPT,
e=(i, j ) ∈E

s.t. x j = bi ∀i = 1, . . . , k,
j ∈F i
x j (1 − x j ) = 0 ∀j = 1, . . . , k.
Let us denote this quadratic feasibility problem by Q. The SoS (Lasserre) hierarchy gives a semi-
definite program that relaxes Q. We denote by SoS (Q ) the solutions to the level- SoS relaxations
of Q. Any solution in SoS (Q ) can be represented as a collection of vectors V = {vS : S ⊆ [n], 0 ≤
|S | ≤ }. To avoid overly cluttered notation, we write vi for v {i } ; we also write v0 for v∅ . We need
the following properties of V, valid as long as  ≥ 2.
(1) v0 · v0 = 1.

ACM Transactions on Algorithms, Vol. 18, No. 4, Article 33. Publication date: October 2022.
Sticky Brownian Rounding and its Applications to Constraint Satisfaction Problems 33:23

(2) vS · vT = vS  · vT  for any S, S ,T ,T  such that S ∪ T = S  ∪ T  and |S ∪ T | ≤ k. In particular,

vi · vi = vi · v0 for any i.
(3) For any i and j the following inequalities hold:
1 ≥ v0 · vi + vj · v0 − vi · vj , (11)
vi · v0 ≥ vi · vj , (12)
vi · vj ≥ 0, (13)

(4) e=(i, j ) ∈E a(e)vi − vj  2 ≥ OPT,
(5) For any i, there exist two solutions V i→0 and V i→1 in SoS−1 (Q ) such that, if we denote the
vectors in V i→0 by vS0 , and the vectors in V i→1 by vS1 , we have
vS · v0 = (1 − vi · v0 )vS0 · v00 + (vi · v0 )vS1 · v10 .
(6) For any Ft t = 1 . . . k, the vectors satisfy

vi  2 ≤ bi .
i ∈F t

Moreover, a solution V ∈ SoS can be computed in time polynomial in n  .

Intuitively, we think of V as describing a pseudo-distribution over solutions to Q, and we inter-
pret vS · vT as the pseudo-probability that all variables in S ∪ T are set to one, or, equivalently, as

the pseudo-expectation of i ∈S ∪T x i . Usually we cannot expect that there is any true distribution
giving these probabilities. Nevertheless, the pseudo-probabilities and pseudo-expectations satisfy
some of the properties of actual probabilities. For example, the transformation from V to V i→b
corresponds to conditioning x i to b.
We will denote by x S = vS · v0 the marginal value of set S. In particular, we will work with
the single-variable marginals x i = x {i } = vi · v0 , and will denote x = (x 1 , . . . , x n ). As before,
it will be convenient to work with the component of vi which is orthogonal to v0 . We define
wi = vi − x i v0 , and wi =  w1i  wi . Note that, by the Pythagorean theorem,  wi  2 = x i − x i2 , and

vi = x i v0 + x i − x i2 wi . We define the matrices W and W by Wi, j := wi · wj and Wi, j := wi · wj .
We can think of W as the covariance matrix of the pseudodistribution corresponding to the SDP
solution. The following lemma, due to Barak, Raghavendra, and Steurer [17], and, independently,
to Guruswami and Sinop [32], shows that any pseudodistribution can be conditioned so that the
covariances are small on average.
Lemma 10. For any ε 0 ≤ 1, and any V ∈ SoS (Q ), where  ≥ 1
ε 04
+ 2, there exists an efficiently
computable V  ∈ SoS−1/ε04 (Q ), such that

n 
n
Wi,2 j ≤ ε 04n2 , (14)
i=1 j=1

where W is defined with respect to V . In particular, V  can be computed by conditioning V on 1

ε 04
variables.

4.3 Rounding Algorithm

For our algorithm, we first solve a semidefinite program to compute a solution in SoS (Q ), to
which we apply Lemma 10 with parameter ε 0 , which we will choose later. In order to be able to
apply the lemma, we choose  =  ε14  + 2. The rounding algorithm itself is similar to the one we
0
used for Max-2SAT. We perform a Sticky Brownian Motion with initial covariance W, starting at
the initial point X0 = x, i.e., at the marginals given by the SDP solution. As variables hit 0 or 1,
ACM Transactions on Algorithms, Vol. 18, No. 4, Article 33. Publication date: October 2022.
33:24 S. Abbasi-Zadeh et al.

we freeze them, and delete the corresponding row and column of the covariance matrix. The main
difference from the Max-2SAT rounding is that we stop the process at time τ , where τ is another
parameter that we will choose later. Then, independently for each i = 1, . . . , n, we include vertex
i in the final solution S with probability (Xτ )i , and output S.
The key property of this rounding that allows us to handle a large number of global constraints

is that, for any Fi ∈ F , the value j ∈Fi (Xτ )j that the fractional solution assigns to the set Fi sat-

isfies a sub-Gaussian concentration bound around bi . Note that j ∈Fi (Xt )j is a martingale with
expectation equal to bi . Moreover, by Lemma 10, the entries of the covariance matrix W are small
on average, which allows us to also bound the entries of the covariance matrix W, and, as a con-
sequence, bound how fast the variance of the martingale increases with time. The reason we stop
the walk at time τ is to make sure the variance does not grow too large: This freedom, allowed
by the Sticky Brownian Motion rounding, is important for our analysis. The variance bound then

implies the sub-Gaussian concentration of j ∈Fi (Xτ )j around its mean bi , and using this concen-
tration we can show that no constraint is violated by too much. This argument crucially uses the
fact that our rounding is a random walk with small increments, and we do not expect similarly
strong concentration results for the random hyperplane rounding or its variants.
The analysis of the objective function, as usual, reduces to analyzing the probability that we cut
an edge. However, because we start the Sticky Brownian Motion at x, which may not be equal to 0,
our analysis from Section 2 is not sufficient. Instead, we use the PDE-based analysis from Section 3,
which easily extends to the Max-Cut objective. One detail to take care of is that, because we stop
the walk early, edges incident on vertices that have not reached 0 or 1 by time τ may be cut
with much smaller probability than their contribution to the SDP objective. To deal with this, we
choose the time τ when we stop the walk large enough, so that any vertex has probability at least
1 − poly(ε) to have reached {0, 1} by time τ . We show that this happens for τ = Θ(log(1/ε)). This

value of τ is small enough so that we can usefully bound the variance of j ∈Fi (Xτ )i and prove the
sub-Gaussian concentration we mentioned above.
Let us recall some notation that will be useful in our analysis. We will use τi for the first time t
that Xt hits a face of [0, 1]n of dimension n−i; then, Xτn ∈ {0, 1}n . We also use Wt for the covariance
used at time step t, which is equal to W with rows and columns indexed by {i : (Xt )i ∈ {0, 1}}
zeroed out.
As discussed, our analysis relies on a martingale concentration inequality, and the following
lemma, which is proved with the methods we used above for the Max-2SAT problem. A proof
sketch can be found in the appendix.
Lemma 11. For the SDP solution V and the Sticky Brownian Motion Xt described above, and for
any pair {i, j} of vertices
Pr[(Xτn )i  (Xτn )j ] ≥ 0.843 · vi − vj  2 .
The next lemma shows that the probability that any coordinate is fixed by time t drops expo-
nentially with t. We use this fact to argue that by time τ = Θ(log(1/ε)) the endpoints of any edge
have probability at least 1 − poly(ε) to be fixed, and, therefore, edges are cut with approximately
the same probability as if we did not stop the random walk early, which allows us to use Lemma 11.
The proof of this lemma, which is likely well-known, appears in the appendix.
Lemma 12. For any i, and any integer t ≥ 0, Pr[∀s ≤ t : 0 < (Xs )i < 1] < 4−t .
The following concentration inequality is our other key lemma. The statement is complicated by
the technical issue that the concentration properties of the random walk depend on the covariance
matrix W, while Lemma 10 bounds the entries of W. When x i (1 − x i ) or x j (1 − x j ) is small, Wi, j
can be much smaller than Wi, j . Because of this, we only prove our concentration bound for sets

ACM Transactions on Algorithms, Vol. 18, No. 4, Article 33. Publication date: October 2022.
Sticky Brownian Rounding and its Applications to Constraint Satisfaction Problems 33:25

of vertices i for which x i (1 − x i ) is sufficiently large. For those i for which x i (1 − x i ) is small, we
will instead use the fact that such x i are already nearly integral to prove a simpler concentration
bound.
Lemma 13. Let ε 0 , ε 1 ∈ [0, 1], and n ≥ 8τε1ε 2 . Define V>ε1 = {i : 2x i (1 − x i ) > ε 1 }. For any set
0
F ⊆ V>ε1 , and any t ≥ 0, the random set S output by the rounding algorithm satisfies
⎡⎢   ⎤⎥  
⎢   ⎥ ε 1t 2
Pr ⎢ |F ∩ S | − x i  ≥ tε 0n⎥ ≤ 4 exp − .
⎢⎣ i ∈F
 ⎥⎦ 4τ
We give the proof of Lemma 13 after we finish the proof of Theorem 3, restated below for
convenience.
Theorem 3. There exists a O (npoly(log(k )/ε ) )-time algorithm that on input a satisfiable instance
G = (V , E), F , and b1 , . . . , bk , as defined above, outputs a (0.843 − ε, ε)-approximation with high
probability.

Proof. The algorithm outputs either the set S output by the Sticky Brownian Rounding de-
scribed above, or the one guaranteed by Lemma 9, depending on which one achieves a cut of
larger total weight. If OPT ≤ ε2 a(E), then Lemma 9 achieves the approximation we are aiming for.
Therefore, for the rest of the proof, we may assume that OPT ≥ ε2 a(E), and that the algorithm
outputs the set S computed by the Sticky Brownian Rounding. Then, it is enough to guarantee
that, with high probability,
ε2
a(δ (S )) ≥ 0.843 · OPT − a(E). (15)
2
Let us set ε 1 = ε 2ε 0 , and define, as above, V>ε1 = {i : 2x i (1 − x i ) > ε 1 } and let V ≤ε1 = {i : 2x i (1 −
x i ) ≤ ε 1 }. Let Y be the indicator vector of the set S output by the algorithm. Observe that, for each

i, since Yi is a Bernoulli random variable with expectation x i , we have E[ i ∈V≤ε1 |Yi − x i |] ≤ ε 1n,
and, therefore,
⎡⎢  ⎤⎥
⎢ ε2
Pr ⎢⎢ |Yi − x i | ≥ 16ε 0n⎥⎥⎥ ≤ .
⎢⎣i ∈V≤ε1 ⎥⎦ 16
Then, for any Fi ∈ F , by Lemma 13 applied to Fi ∩ V>ε1 , we have
⎡⎢  ⎤
 32k ⎥ ε2
⎢ 
Pr ⎢⎢ |Fi ∩ V>ε1 ∩ S | − 
x i  ≥
4τ
ln 2 ε 0n⎥⎥⎥ ≤ .
⎢⎣ i ∈F ∩V>ε1
ε1 ε ⎥⎦ 16k
ε2
Therefore, with probability at least 1 − 8, for all i ∈ [k] we have
        
 |Fi ∩ S | −  
x i  ≤  |Fi ∩ V ≤ε0 ∩ S | −  
x i  +  |Fi ∩ V>ε0 ∩ S | − x i 
 i ∈F

   i ∈F ∩V≤ε0     i ∈F ∩V>ε0 

   
≤ 
|Yi − x i | +  |Fi ∩ V>ε0 ∩ S | − x i 
i ∈F ∩V≤ε0  i ∈F ∩V>ε0 

≤ 16 + ln 2  ε 0n.
4τ 32k
ε 2ε 0 ε
 

ACM Transactions on Algorithms, Vol. 18, No. 4, Article 33. Publication date: October 2022.
33:26 S. Abbasi-Zadeh et al.
2
This means that, with probability at least 1 − ε8 , S satisfies all the constraints up to additive error
εn, as long as
⎧
⎪ε ε4 ⎫
⎪
ε 0 ≤ min ⎨
⎪ 32 ,  ⎬.
⎪
⎩ 4 τ ln ε 2 ⎭
32k
√
It remains to argue about the objective function. For τ ≥ log2 2 ε 2 , Lemma 12 implies that, for
any vertex i, Pr[(Xτ )i  {0, 1}] ≤ 4−τ ≤ ε8 . By Lemma 11, any pair of vertices {i, j} is separated
2

with probability
Pr[(Xτn )i  (Xτn )j ] ≥ 0.843 · vi − vj  2 ,
where we recall that, for edge e = (i, j), a(e)vi − vj  2 is the contribution of e to the objective
value. Then,
Pr[(Xτ )i  (Xτ )j ] ≥ Pr[(Xτn )i  (Xτn )j , (Xτ )i = (Xτn )i , (Xτ )j = (Xτn )j ]
= Pr[(Xτn )i  (Xτn )j , (Xτ )i ∈ {0, 1}, (Xτ )j ∈ {0, 1}]
ε2
≥ Pr[(Xτn )i  (Xτn )j ] −
4
ε2
≥ 0.843 · vi − vj 22 − .
4
ε2
Therefore, E[a(δ (S ))] ≥ 0.843 · OPT − 4 a(E). By Markov’s inequality applied to a(E) − a(δ (S )),
 ε2  1+ ε2
ε2
4
Pr a(δ (S )] < 0.843 · OPT − a(E) < ε2
< 1− .
2 1+ 2
5
3 2
In conclusion, we have that with probability at least 40 ε , constraint (15) is satisfied, and all global
constraints are satisfied up to an additive error of εn. The probability can be made arbitrarily close
to 1 by repeating the entire algorithm O (ε −2 ) times. To complete the proof of the theorem, we can
verify that the running time is dominated by the time required to find a solution in SoS (Q ), which
is polynomial in n  , where  = O (ε 0−4 ) = poly(log(k )/ε). 
We finish this section with the proof of Lemma 13.
Proof of Lemma 13. Since each i is included in S independently with probability (Xτ )i , by Ho-
effding’s inequality we have
⎡⎢  ⎤⎥  
 ε 1t 2
Pr ⎢⎢ |F ∩ S | − (Xτ )i  ≥ tε 1n⎥⎥ ≤ 2e −2ε1 t n ≤ 2 exp −
2 2
,
⎢⎣ i ∈F
 ⎥⎦ 4τ
where the final inequality follows by our assumption on n. Therefore, it is enough to establish
⎡⎢ ⎤⎥  
⎢   ⎥ ε 1t 2
Pr ⎢ (Xτ )i − x i  ≥ tε 1n⎥ ≤ 2 exp − . (16)
⎢⎣ i ∈F  ⎥⎦ 4τ
Suppose y ∈ {0, 1}n is the indicator vector of F so that

y (Xτ − X0 ) = y (Xτ − x) = ((Xτ )i − x i ).
i ∈F
A standard calculation using Itô’s lemma (see Exercise 4.4. in [47]) shows that, for any λ ≥ 0, the
random process  
t
 λ2 
Yt = exp λy (Xt − x) − (y Ws y)ds ,
2 0

ACM Transactions on Algorithms, Vol. 18, No. 4, Article 33. Publication date: October 2022.
Sticky Brownian Rounding and its Applications to Constraint Satisfaction Problems 33:27

is a martingale with starting state Y0 = 1. Since, for any s, Ws equals W with some rows and
columns zeroed out, we have that W−Ws is positive semidefinite, and y Ws y ≤ y Wy. Therefore,
  
λ2
E exp λy (Xτ − x) − τ y Wy ≤ E[Yτ ] = 1.
2
Rearranging, this gives us that, for all λ ≥ 0,
 (X −x) 2 y Wy/2 2 y Wy/2
E[e λy τ
] ≤ E[e τ λ ] ≤ eτ λ . (17)
We can bound y Wy using Cauchy–Schwarz, the assumption that 2x i (1 − x i ) > ε 1 for all i ∈ F ,
and constraint (14):

y Wy = Wi, j
i ∈F j ∈F
1/2

≤ |F |  Wi,2 j 
 i ∈F j ∈F 
1/2
 Wi,2 j
= |F |  
x i x j (1 − x i )(1 − x j )
 i ∈F j ∈F 
1/2
2n   2 
<  W 
ε 1 i ∈F j ∈F i, j
 
2ε 02n2
≤ .
ε1
 (X −x) 2 ε 2 n 2 /ε
Plugging back into constraint (17), we get E[e λy τ ] ≤ eτ λ 0 1 . The standard exponential
moment argument then implies constraint (16). 

5 BROWNIAN ROUNDING WITH SLOWDOWN

As noted in Section 2, the Sticky Brownian rounding algorithm does not achieve the optimal value
for the Max-Cut problem. A natural question is to ask if we can modify the algorithm to achieve
the optimal constant. In this section, we will show that a simple modification achieves this ratio
up to at least three decimals. Our results are computer-assisted as we solve partial differential
equations using finite element methods. These improvement indicate that variants of the Brownian
Rounding approach offer a direction to obtain optimal SDP rounding algorithms for Max-Cut
problem as well as other CSP problems.
In the sticky Brownian motion, the covariance matrix Wt is a constant, until some vertex’s
marginals (Xt )i becomes ±1. At that point, we abruptly zero the ith row and column. In this section,
we analyze the algorithm where we gradually dampen the step size of the Brownian motion as it
approaches the boundary of the hypercube, until it becomes 0 at the boundary. We call this process
a “Sticky Brownian Motion with Slowdown.”
Let (Xt )i denote the marginal value of vertex i at time t. Initially (X0 )i = 0. First, we describe the
discrete algorithm which will provide intuition but will also be useful to those uncomfortable with
Brownian motion and diffusion processes. At each time step, we will take a step whose length is
scaled by a factor of (1 − (Xt )i2 ) α for some constant α > 0. In particular, the marginals will evolve
according to the equation:
√
(Xt +dt )i = (Xt )i + (1 − (Xt )i ) 2 ) α /2 · (wi · Gt ) · dt . (18)

ACM Transactions on Algorithms, Vol. 18, No. 4, Article 33. Publication date: October 2022.
33:28 S. Abbasi-Zadeh et al.

where Gt is distributed according to an n-dimensional Gaussian and dt is a small discrete step by

which we advance the time variable. When Xt is sufficiently close to −1 or +1, we round it to the
nearest one of the two: from then on it will stay fixed because of the definition of the process,
i.e., we will have (Xs )i = (Xt )i for all s > t.
More formally, Xt is defined as an Itô diffusion process which satisfies the stochastic differential
equation
dXt = A(Xt ) · W /2 · dBt ,
1
(19)
where Bt is the standard Brownian motion in Rn
and A(Xt ) is the diagonal matrix with entries
[A(Xt )]ii = (1 − (Xt )i2 ) α /2 . Since this process is continuous, it becomes naturally sticky when
some coordinate (Xt )i reaches {−1, 1}.
Once again, it suffices to restrict our attention to the two-dimensional case where we analyze
the probability of cutting an edge (i, j) and we will assume that
 
1 cos(θ )
W̃ = ,
cos(θ ) 1
where θ is the angle between wi and wj .
Let τ be the first time when Xt hits the boundary ∂[−1, 1]2 . Since the walk slows down as it
approaches the boundary, it is worth asking if E[τ ] is finite. In Lemma 21, we show that E[τ ] is
finite for constant α.
Let u (x, y) denote the probability of the Sticky Brownian Walk algorithm starting at (x, y) cut-
ting an edge, i.e., the walk is absorbed in either (+1, −1) or (−1, +1). It is easy to give a precise
formula for u at the boundary as the algorithm simplifies to a one-dimensional walk. Thus, u (x, y)
satisfies the boundary condition ϕ (x, y) = (1−xy)/2 (which can obtained by, e.g., optimal stopping
theorem). for all points (x, y) ∈ ∂[−1, 1]2 . For a given (x, y) ∈ Int[−1, 1]2 , we can say
u (x, y) = E (x,y ) [ϕ (X̃τ (i), X̃τ (j))],
where E (x,y ) denotes the expectation of diffusion process that begins at (x, y). Informally, u (x, y) is
the expected value of ϕ when first hitting ∂[−1, 1]2 conditioned on starting at point (x, y). Observe
that the probability that the algorithm will cut an edge is given by u (0, 0).
The key fact about u (x, y) that we use is that it is the unique solution to a Dirichlet Problem,
formalized in Lemma 14 below.
Lemma 14. Let L α denote the operator
∂2 2 α /2 2 α /2 ∂
2
2 α ∂
2
L α = (1 − x 2 ) α + 2 cos(θ )(1 − x ) (1 − y ) + (1 − y ) ,
∂x 2 ∂x ∂y ∂y 2
then the function u (x, y) is the unique solution to the Dirichlet Problem:
L α [u](x, y) = 0 ∀(x, y) ∈ Int([−1, 1]2 )
lim u (x, y) = ϕ (x̃, ỹ) ∀(x̃, ỹ) ∈ ∂[−1, 1]2 .
(x,y )→(x̃, ỹ ),
(x,y ) ∈Int([−1,1]2 )

The proof is largely similar to the one described in Lemma 5 and Theorem 9 with two caveats:
(1) In Theorem 9, we use a fixed Σ. However we can handle the general case, where Σ is allowed
to depend on the diffusion prcoess(i.e. Σ(Xt )), by appealing to general Theorem 9.2.14 in [47].
(2) To apply Theorem 9.2.14 from [47], we need the resulting matrix Σ(Xt )Σ(Xt )  to have eigen-
values bounded away from zero. In our case, Σ(Xt )Σ(Xt )  can have zero rows and columns
on the boundary. To avoid this, we simply restrict our domain to be the hypercube scaled by

ACM Transactions on Algorithms, Vol. 18, No. 4, Article 33. Publication date: October 2022.
Sticky Brownian Rounding and its Applications to Constraint Satisfaction Problems 33:29
Table 2. Approximation ratio of Sticky
Brownian Motion rounding with
Slowdown for Max-Cut and Max-2SAT

α Max-Cut Max-2SAT
0 0.861 0.921
1 0.874 0.927
1.61 0.878 0.929
For Max-2SAT, the best value 0.929 is
achieved at the marginals (0.36, 0.38), and
the angle between the vectors is (0.697π ).

a small value [−1 + δ, 1 − δ ]. This is sufficient since our discrete algorithm will only run in
this region.
Numerical Results. The Dirichlet problem is parameterized by two variables: the slowdown pa-
rameter α and the angle between the vectors θ . We can numerically solve the above equation using
existing solvers for any given fixed α and angle θ ∈ [0, π ]. We solve these problems for a variety
of α between 0 and 2 and all values of θ in [0, π ] discretized to a granularity of 0.02.11
We observe that as we increase α from 0 to 2, the approximation ratio peaks around α ≈ 1.61
for all values of θ . In particular, when α = 1.61, the approximation ratio is 0.878, which matches
the integrality gap for this relaxation up to three decimal points.
The Brownian rounding with slowdown is a well-defined algorithm for any 2-CSP. We investi-
gate 3 different values of slowdown parameter, i.e., α, and show their relative approximation ratios.
We show that with a slowdown of 1.61 we achieve an approximation ratio of 0.929 for Max-2SAT.
We list these values below in Table 2.
For the Max-Cut problem, since we start the walk at the point (0, 0), we only need to investigate
the performance of the rounding for all possible angles between two unit vectors which range in
[0, θ ] (Figure 2). In particular, we are able to achieve values that are comparable to the GW bound.

6 HIGHER-DIMENSIONAL BROWNIAN ROUNDING

Our motivating example for considering the higher-dimension Brownian rounding is the Max-
DiCut problem: given a directed graph G = (V , E) equipped with non-negative edge weights
a : E → R+ we wish to find a cut S ⊆ V that maximizes the total weight of edges going out of S.
The standard semi-definite relaxation for Max-DiCut is the following:
 (w0 + wi ) · (w0 − wj )
max ae · ,
4
e=(i→j ) ∈E
s.t . wi · wi = 1 ∀i = 0, 1, . . . , n,
wi − wj  + wj − wk  ≥ wi − wk 
2 2 2
∀i, j, k = 0, 1, . . . , n.

In the above, the unit vector w0 denotes the cut S, whereas −w0 denotes S. We also include the
triangle inequalities which are valid for any valid relaxation. These inequalities first appeared
in [25], while the SDP relaxation was present in earlier works such as [31].
The sticky Brownian rounding algorithm for Max-DiCut fails to give a good performance
guarantee. In particular, consider the configuration where w 1 = 0.2w 0 + (1 − 0.22 )u and

11 Our code, containing the details of the implementation, is available at [1].

ACM Transactions on Algorithms, Vol. 18, No. 4, Article 33. Publication date: October 2022.
33:30 S. Abbasi-Zadeh et al.

w 2 = 0.8w 0 + (1 − 0.82 )v where u and v vectors orthogonal to w 0 with uT v = cos(2.48).

For this configuration, the probability that the random walk separates u from v is at most 0.60
times the SDP contribution for this pair. Thus we design a high-dimensional variant of the al-
gorithm that incorporates the inherent asymmetry of the problem. Let us now describe the
high-dimensional Brownian rounding algorithm. It is similar to the original Brownian rounding

ACM Transactions on Algorithms, Vol. 18, No. 4, Article 33. Publication date: October 2022.
Sticky Brownian Rounding and its Applications to Constraint Satisfaction Problems 33:31

Fig. 2. Comparing the performance of three values of the slowdown parameter for the Max-Cut problem.

algorithm given for Max-Cut, except that it evolves in Rn+1 with one additional dimension for w0 .
Let W ∈ R (n+1)×(n+1) denote the positive semi-definite correlation matrix defined by the vectors
w0 , w1 , . . . , wn , i.e., for every 0 ≤ i, j ≤ n we have that: Wi, j = wi · wj . The algorithm starts at the
origin and perform a sticky Brownian motion inside the [±1]n+1 hypercube whose correlations
are governed by W.
As before, we achieve this by defining a random process {Xt }t ≥0 as follows:

Xt = W /2 Bt ,
1

where {Bt }t ≥0 is the standard Brownian motion in Rn+1 starting at the origin and W1/2 is the square
root matrix of W. Additionally, we force {Xt }t ≥0 to stick to the boundary of the [±1]n+1 hypercube,
i.e., once a coordinate of Xt equals either 1 or −1 it is fixed and remains unchanged indefinitely.
This description can be formalized the same way we did for the Max-Cut problem. Below we use
σ for the time at which Xσ ∈ {−1, 1}n+1 , which has finite expectation.
Unlike the Brownian rounding algorithm for Max-Cut, we need to take into consideration the
value w0 was rounded to, i.e., (Xσ )0 , since the zero coordinate indicates S. Formally, the output
S ⊆ V is defined as follows:
 
S = i ∈ V : (Xσ ) i = (Xσ ) 0 .

To simplify the rest of the presentation, let us denote Z i := (Xσ )i for every i = 0, 1, . . . , n.
The event that an edge (i → j) ∈ E is an outgoing edge from S, i.e., i ∈ S and j ∈ S, involves three
random variables: Z i , Z j , and Z 0 . Formally, the above event happens if and only if Z i = Z 0 and
Z j  Z 0 . We now show how events on any triplet of the random variables Z 0 , Z 1 , . . . , Z n can be pre-
cisely calculated. To simplify the presentation, denote the following for every i, j, k = 0, 1, 2, . . . , n

ACM Transactions on Algorithms, Vol. 18, No. 4, Article 33. Publication date: October 2022.
33:32 S. Abbasi-Zadeh et al.

and α, β, γ ∈ {±1}:
pi (α )  Pr[Z i = α],
pi j (α, β )  Pr[Z i = α, Z j = β],
pi jk (α, β, γ )  Pr[Z i = α, Z j = β, Z k = γ ].
Observation 3. The following two hold:
(1) pi (α ) = pi (−α ), pi j (α, β ) = pi j (−α, −β ), and pi jk (α, β, γ ) = pi jk (−α, −β, −γ ) for every i, j, k =
0, 1, 2, . . . , n and α, β, γ ∈ {±1}.
(2) pi (α ) = 1/2 for every i = 0, 1, 2, . . . , n and α ∈ {±1}.
The proof of Observation 3 follows immediately from symmetry.
The following lemmas proves that every conjunction event that depends on three variables from
Z 0 , Z 1 , Z 2 , . . . , Z n can be precisely calculated.
Lemma 15. For every i, j, k = 0, 1, 2, . . . , n and α, β, γ ∈ {±1}:
1 1
pi jk (α, β, γ ) = pi j (α, β ) + pik (α, γ ) + p jk (β, γ ) − .
2 2
Proof.
1 − pi jk (α, β, γ ) = 1 − pi jk (−α, −β, −γ )
= Pr Z i = α ∨ Z j = β ∨ Z k = γ
= pi (α ) + p j (β ) + pk (γ ) − pi j (α, β ) − pik (α, γ ) − p jk (β, γ ) + pi jk (α, β, γ ).
The first equality follows from property (1) of Observation 3. The second equality follows from
De-Morgan’s law. The third equality follows from the inclusion and exclusion principle. Isolating
pi jk (α, β, γ ) above and using property (2) of Observation 3 concludes the proof. 
Let us now consider the case study problem Max-DiCut. One can verify that an edge (i → j) ∈ E
is a forward edge crossing the cut S if and only if the following event happens: {Z i = Z 0  Z j }
(recall that Z 0 indicates S). Thus, the value of the Brownian rounding algorithm, when considering
only the edge (i → j), equals p0i j (1, 1, −1) + p0i j (−1, −1, 1). Lemma 15 above shows that if one
knows the values of pi j (α, β ) for every i, j = 0, 1 . . . , n and α, β ∈ {±1}, then p0i j (1, 1, −1) and
p0i j (−1, −1, 1) can be calculated (thus deriving the exact probability that (i → j) is a forward edge
crossing S).
How can we calculate pi j (α, β ) for every i, j = 0, 1 . . . , n and α, β ∈ {±1}? Fix some i, j, α, and
β. We note that Theorem 4 can be used to calculate pi j (α, β ). The reason is that: (1) Theorem 4
provides the value of pi j (−1, 1) + pi j (1, −1); (2) pi j (−1, −1) + pi j (−1, 1) + pi j (1, −1) + pi j (1, 1) = 1;
and (3) pi j (−1, −1) = pi j (1, 1) and pi j (−1, 1) = pi j (1, −1) from symmetry. We conclude that using
Theorem 4 we can exactly calculate the probability that (i → j) is a forward edge crossing S, and
obtain that this probability equals:
p0i j (1, 1, −1) + p0i j (−1, −1, 1).
Simplifying this using Lemma 15, we get
1
(p0j + pi j − p0i ),
2
where pi j is the probability that i and j are separated as given by Theorem 4.
Similarly to Max-2SAT, not all triplets of angles {θ 0i , θ 0j , θ i j } are possible due to the triangle
inequality constraints (here θ i j indicates the angle between wi and wj ). Let us denote by F the

ACM Transactions on Algorithms, Vol. 18, No. 4, Article 33. Publication date: October 2022.
Sticky Brownian Rounding and its Applications to Constraint Satisfaction Problems 33:33

collection of all possible triplet of angles for the Max-DiCut problem. Then, we can lower bound
the approximation guarantee of the Brownian rounding algorithm as follows:
⎧ ⎫
2 (p 0j + pi j − p 0i )
1
⎪
⎨ ⎪
⎬.
min
(θ 0i ,θ 0j ,θ i j ) ∈ F ⎪ 1 (1 − cos(θ 0j ) + cos(θ 0i ) − cos(θ i j )) ⎪
⎩ 4 ⎭
This results in the following theorem.
Theorem 11. The high-dimensional Brownian rounding algorithm achieves an approximation ra-
tio of 0.79 for the Max-DiCut problem.
The ratio of 0.79 is achieved at the point w i = 0.86w 0 + (1 − 0.862 )u and w j = 0.08w 0 +
(1 − 0.082 )v and the angle between u and v is cos−1 (2.1708).
We also remark that we can introduce slowdown (as discussed in Section 5 to the high dimen-
sional Brownian rounding algorithm. Numerically, we show that this improves the performance
to 0.81-approximation.

APPENDICES
A OMITTED PROOFS FROM SECTION 2
We start with a brief primer about special functions with an emphasis on the lemmas and identities
that will be useful for our analysis. We recommend the excellent introductions in Andrews et al.
[4], Beals and Wong [18] for a thorough introduction.

A.1 Special Functions: A Primer

While there is no common definition of special functions, three basic functions, Γ, β and the hy-
pergeometric functions p Fq show up in nearly all treatments of the subject. We will define them
and some useful relationships between them.
Definition 2 (Gamma Function). The gamma function is defined as
∞
Γ(z) := x z−1e −x dx,
0
for all complex numbers z with non-negative real part, and analytically extended to all z 
0, −1, −2, . . ..
Fact 1. Recall that the gamma function satisfies the recurrence Γ(z + 1) = zΓ(z) and it follows
easily from the definition that Γ(1) = 1. In particular, when n is a positive integer, Γ(n + 1) = n!
Definition 3 (Beta Function). The beta function β (a, b) is defined for complex numbers a and b
with Re(a) > 0, Re(b) > 0 by
1
β (a, b) = s a−1 (1 − s)b−1ds.
0
Clearly, β (a, b) = β (b, a). Setting s = u/(u + 1) gives the following alternate form.
∞  1  a+b
β (a, b) = u a−1 du.
0 1 +u

Lemma 16 (Theorem 1.1.4 in [4]). The beta function can be expressed in terms of the gamma
function using the following identity:
Γ(a)Γ(b)
β (a, b) = .
Γ(a + b)
We will use the following very useful fact.
ACM Transactions on Algorithms, Vol. 18, No. 4, Article 33. Publication date: October 2022.
33:34 S. Abbasi-Zadeh et al.

Lemma 17 (Exercise 2.2 in [18]).

π /2
1
sina−1 θ cosb−1 θdθ = β (a/2, b/2).
0 2
The next family of functions we utilize are the hypergeometric functions.
 
a, ..., ap
Definition 4 (Hypergeometric Function). The hypergeometric function p Fq b1, ..., bq
;z is defined
1
as  
a 1 , . . . , ap ∞
(a 1 )n . . . (ap )n z n
p Fq ; z := ,
b 1 , . . . , bq n=0
(b1 )n · (bq )n n!
where the Pochhammer symbol (rising factorial) is defined inductively as
(a)n := a(a + 1) . . . (a + n − 1) and (a)0 = 1.
A very simple but useful way to write the binomial theorem using the Pochhammer symbol is
∞
(a)n n
(1 − x ) −a = x .
n=0
n!
The Pochhammer symbol also satisfies the formula (a)n = Γ(a + n)/Γ(a).
A useful connection between the hypergoemetric function 2 F 1 and the gamma function is given
in the following lemma.
Lemma 18 ((Euler’s Integral Representation) (Theorem 2.2.1 in [4])). If Re(c) > Re(b) > 0
then   1
a, b Γ(c)
2 F1 ;x = t b−1 (1 − s) c−b−1 (1 − xs) −a ds,
c Γ(b)Γ(c − b) 0
where we assume that (1 − xs) −a takes its principal value.
Definition 5. The incomplete beta integral is defined as
x
β x (a, b) = t a−1 (1 − t )b−1dt,
0
and is well-defined for Re(a) > 0 and x  [1, ∞).
Lemma 17 easily extends to the incomplete beta integral too, as captured in the following lemma.
Lemma 19.
ϕ
1
sina−1 θ cosb−1 θdθ = β sin2 ϕ (a/2, b/2).
0 2
√
Proof. Let sin θ = t, then cos θ = 1 − t 2 and (cos θ )dθ = dt, and we get
ϕ sin(ϕ )
sina−1 θ cosb−1 θdθ = t a−1 (1 − t 2 ) (b−2)/2dt .
0 0
Setting s = t 2 gives
sin2 (ϕ )
(1/2)s (a−2)/2 (1 − s) (b−2)/2ds = (1/2)β sin2 ϕ (a/2, b/2).
0
This completes the proof. 
The following identity relates the incomplete beta integral to hypergeometric functions.

ACM Transactions on Algorithms, Vol. 18, No. 4, Article 33. Publication date: October 2022.
Sticky Brownian Rounding and its Applications to Constraint Satisfaction Problems 33:35

Lemma 20 ((Gauss’s Identity) (Exercise 8.7 in [18])).

 
x
xa a, 1 − b
β x (a, b) = t a−1 (1 − t )b−1dt = · 2 F1 ;x .
0 a a+1
Proof. It is natural to substitute t = sx, as we can now integrate from s = 0 to 1. This gives
1 1
x a−1s a−1 (1 − sx )b−1xds = x a s a−1 (1 − sx )b−1ds.
0 0
Using the integral form given in Lemma 18 with 1 − b in the place of a, a in the place of b, and
a + 1 in the place of c, we get that the integral equals
   
Γ(a)Γ(1) 1 − b, a xa 1 − b, a
xa · 2 F1 ;x = · 2 F1 ;x .
Γ(a + 1) a+1 a a+1
By the symmetry in the definition of 2 F 1 with respect to the first two arguments, the result
follows. 

A.2 Proof of Theorem 4

First, we will prove the claim, which expresses the function r (ϕ) in terms of the incomplete beta
function.
Claim 1.
1
r (ϕ) = β sin2 ϕ (a/2, b/2),
4
when ϕ ∈ [0, π /2].
Proof. Recall r (ϕ) = |Fθ (e iϕ )−Fθ (1)|. Furthermore, from Lemma 3, we know that the conformal
map maps the arc from 1 to i to an edge on the rhombus S. Hence, we can write r (ϕ) as an integral
ϕ ϕ
r (ϕ) = |Fθ (e iψ )|dψ = | fθ (e iψ )|dψ .
0 0

Expanding fθ , and substituting a = θ

π and b = 1 − a, we have
ϕ
= |(1 − e 2iψ ) a−1 · (1 + e 2iψ )b−1 |dψ .
0
Expanding this in terms of trigonometric functions, and simplifying using double angle formulas,
we get
ϕ
= |(−2i · e iψ · sinψ ) a−1 · (2 · e iψ · cosψ )b−1 |dψ
0
ϕ
= |2a+b−2 | · | − ie iψ (a−1) | · | sinψ a−1 | · |e iψ (b−1) | · | cosψ b−1 |dψ .
0

Since | − ie iψ (a−1) | = |e iψ (b−1) | = 1 and the remaining terms are positive, we drop the norms.
ϕ
1
= (sinψ ) a−1 (cosψ )b−1dψ
0 2
1
= β sin2 ϕ (a/2, b/2) by Lemma 19. 
4
By substituting ϕ = π /2 we immediately get the following corollary:
Corollary 3. The length of the side of rhombus is given by r = r (π /2) = 1/4 · β (a/2, b/2).

ACM Transactions on Algorithms, Vol. 18, No. 4, Article 33. Publication date: October 2022.
33:36 S. Abbasi-Zadeh et al.

The claim below will characterize the integral of the incomplete beta function which will be
important for us later.
Claim 2.
π /2  1+a 
β (a/2 + 1/2, 1/2) 2 , 2 , 2
1+a a
4· r (ϕ)dϕ = · 3 F2 ;1 .
2, 2 + 1
a a a
0

Proof. By Lemma 19, the left hand side equals

π /2  
a b
β sin2 ϕ , dϕ
0 2 2
π /2 2 sin2 ϕ ⎡⎢ a
a/2
⎤
⎢ 2 , 1 − 2 ; sin2 ϕ ⎥⎥⎥ dϕ
b
= 2 F1 ⎢ By Lemma 20
0 a ⎢⎣ a2 + 1 ⎥⎦
π /2  
2, 2
a a+1
2(sin ϕ) a
= 2 F1 a ; sin2 ϕ dϕ Substituting b = 1 − a
2 +1
0 a
2 π /2   (a/2)n (a/2 + 1/2)n (sin ϕ) 2n+a 
∞
= dϕ Expand using Definition 4
a 0 n=0
(a/2 + 1)n n!
2   π /2  (a/2) (a/2 + 1/2)
∞
n n
= (sin ϕ) 2n+a dϕ (*)
a n=0 0 (a/2 + 1)n · n!

We take a brief interlude to analyze the integral in the parenthesis above:

π /2
1
(sin ϕ) 2n+a dϕ = β (n + a/2 + 1/2, 1/2) By Lemma 17
0 2
Γ(1/2) Γ(n + a/2 + 1/2)
= By Lemma 16
2 Γ(n + a/2 + 1)
Γ(1/2) (a/2 + 1/2)n Γ(a/2 + 1/2)
=
2 (a/2 + 1)n Γ(a/2 + 1)
β (a/2 + 1/2, 1/2) (a/2 + 1/2)n
= .
2 (a/2 + 1)n
Going back and substituting the above result into the Equation (*), we get

β (a/2 + 1/2, 1/2)  (a/2)n (a/2 + 1/2)n (a/2 + 1/2)n 

∞
=
a n=0 n!(a/2 + 1)n (a/2 + 1)n 
 1+a 1+a a 
β (a/2 + 1/2, 1/2) , ,
= · 3 F2 a 2 2 2
;1 . 
a 2 + 1, a
2 + 1
Armed with Claim 2 and Corollary 3, we can prove Theorem 4.
Theorem 4. The probability that the Sticky Brownian Motion rounding algorithm will separate a
pair {i, j} of vertices for which θ = cos −1 (wi · wj ) equals
 1+a 1+a a 
Γ( a+1 ) 2 , 2 , 2
1 − 1−a 2a · 3 F 2 ;1 ,
2, 2 + 1
a a
Γ( 2 )Γ( 2 + 1) 2

where a = θ/π , Γ is the gamma function, and 3 F 2 is the hypergeometric function.

ACM Transactions on Algorithms, Vol. 18, No. 4, Article 33. Publication date: October 2022.
Sticky Brownian Rounding and its Applications to Constraint Satisfaction Problems 33:37

Proof. Substituting r = r (π /2) below, by Lemma 4 we have that the probability of separating
the vertices is
π /2
2 r (ϕ)
1− dϕ,
π ϕ=0 r
or equivalently, the probability of not separating them is
π /2
2 r (ϕ)
dϕ.
π ϕ=0 r
Expanding r using Claim 1, we get that this equals
π /2
2 4
r (ϕ)dϕ.
π β (a/2, b/2) ϕ=0

Expanding the right hand side integral using Claim 2, we get

 
β (a/2 + 1/2, 1/2) 2 , 2 , 2
1+a 1+a a
2 1
= · · 3 F2 ; 1 .
2 + 1, 2 + 1
π β (a/2, b/2) a a a

Using Lemma 16 and the fact that Γ(1/2) 2 = π we can simplify this to
 1+a 1+a a 
Γ(a + 1/2) , 2 , 2
= · 3 F2 a 2 ;1 . 
Γ(1/2 − a/2)Γ(a/2 + 1) 2 2 + 1, 2 + 1
a

A.3 Proof of Theorem 7

First, we rewrite r in a form that will be useful later.
Claim 3.
π /2 π /2
2 r (ϕ)dϕ = ϕ (sin ϕ)b−1 (cos ϕ) a−1dϕ.
0 0
Proof. The left hand side equation can be written as
π /2 π /2 π /2  ϕ 
1
2 r (ϕ)dϕ = β sin2 ϕ (a/2, b/2)dϕ = (sinψ ) a−1 (cosψ )b−1dψ dϕ.
0 0 2 0 0
 
Applying integration by parts: pdq = [pq] − qdp with q = ϕ − π /2 and p =
ϕ
0
(sinψ a−1 )(cosψ )b−1dψ gives

 ϕ  π /2 π /2
d ϕ
(ϕ − π /2) (sinψ a−1 )(cosψ )b−1dψ + (π /2 − ϕ) (sinψ a−1 )(cosψ )b−1dψdϕ.
0 0 0 dϕ 0
The first term is 0, and using the fundamental theorem of calculus, the second term is
π /2
(π /2 − ϕ)(sin ϕ) a−1 (cos ϕ)b−1dϕ.
0
Substituting ϕ for π /2 − ϕ gives
π /2
ϕ (sin ϕ)b−1 (cos ϕ) a−1dϕ. 
0

ACM Transactions on Algorithms, Vol. 18, No. 4, Article 33. Publication date: October 2022.
33:38 S. Abbasi-Zadeh et al.

Next, we claim that

Claim 4. When θ = (1 − ϵ )π , we can say
π /2
2 r (ϕ)dϕ ≤ 2 · (1 + O (ϵ log(1/ϵ ))).
0
Proof. Using Claim 3, we can write
π /2 π /2
2 r (ϕ)dϕ = ϕ (sin ϕ)b−1 (cos ϕ) a−1dϕ
0 0
π /2
ϕ
=
(tan ϕ) ϵ dϕ.
sin ϕ 0
Since x
sin(x ) ≤ 2 for 0 ≤ x ≤ π/2, to prove the claim it suffices to show that
π /2
((tan ϕ) ϵ − 1) dϕ = O (ϵ log(1/ϵ )).
0
Let ϕ 0 = arctan(1/ϵ ). We will break the above integral into two parts and deal with each separately:
π /2 ϕ0 π /2
((tan ϕ) ϵ − 1) dϕ = ((tan ϕ) ϵ − 1) + ((tan ϕ) ϵ − 1) .
0 0 ϕ0
Case 1 for ϕ ≤ ϕ 0 ,
 1 ϵ
(tan ϕ) ϵ ≤ = exp(ϵ log(1/ϵ )) = 1 + O (ϵ log(1/ϵ )),
ϵ
so
ϕ0
((tan ϕ) ϵ − 1) dϕ = O (ϵ log(1/ϵ )).
0
Case 2 For ϕ > ϕ 0 ,
π /2 π /2
((tan ϕ) ϵ − 1) dϕ ≤ 1/(cos ϕ) ϵ dϕ
ϕ0 ϕ0
π /2−ϕ 0
= (1/ sin ϕ) ϵ dϕ Since sin(x ) = cos(π /2 − x )
0
π /2−ϕ 0
≤ (2/ϕ) ϵ dϕ Since 1 ≤ x/sin(x ) ≤ 2
0
(π /2 − ϕ 0 ) 1−ϵ
≤ 2ϵ
1−ϵ
≤ (π /2 − ϕ 0 )(1 + O (ϵ )).
Finally, we note that π /2 − ϕ 0 ≤ tan(π /2 − ϕ 0 ) = 1/ tan(ϕ 0 ) = ϵ. 
Theorem 7. Given an edge {i, j} with cos −1 (wTi wj )
= θ = (1 − ϵ )π , the Sticky Brownian Motion
rounding will cut the edge with probability at least 1 − ( π4 ϵ + O (ϵ 2 )).
Proof. Let a = 1 − ϵ and b = ϵ.
As discussed in Lemma 4, the non-separation probability is
π /2
2
r (ϕ)dϕ,
πr 0
 π /2
where r = r (π /2). So we will compute the asymptotics of r := r (π /2) and 0
2 · r (ϕ)dϕ as ε → 0.

ACM Transactions on Algorithms, Vol. 18, No. 4, Article 33. Publication date: October 2022.
Sticky Brownian Rounding and its Applications to Constraint Satisfaction Problems 33:39

First we compute the asymptotics of r as ε → 0. Recall that

r = (1/4)β (a/2, b/2) By Corollary 3

Γ((1 − ϵ )/2)Γ(ϵ/2)
= By Lemma 16
4Γ(1/2)
Γ((1 − ϵ )/2)Γ(1 + ϵ/2) ϵ 2 ϵ
= Using Γ( ) = Γ(1 + ).
2ϵ Γ(1/2) 2 ϵ 2
Using the standard fact that Γ(z + ϵ ) = Γ(z)(1 + O (ϵ )) for fixed z > 0
1
= (Γ(1/2) + O (ϵ ))(Γ(1) + O (ϵ ))
2ϵ Γ(1/2)
1
= + O (1),
(2ϵ )
which implies that
1/r = 2ϵ + O (ϵ ) 2 .
 π /2
Using Claim 4, we know that 0 2 · r (ϕ)dϕ is at most 2 · (1 + O (ϵ log(1/ϵ )). Combining the
two, we get the probability of non-separation is ϵ π4 + O (ϵ 2 ) ≈ 1.27ϵ + O (ϵ 2 ). 

A.4 Other Missing Proofs

Lemma 1. Applying the transformation OW−1/2 to {Xt }t ≥0 , we get a new random process {Yt }t ≥0
which has the following properties:
(1) If Xt is in the interior/boundary/vertex of [−1, 1]2 then Yt is in the interior/boundary/vertex of
S, respectively.
(2) S is a rhombus whose internal angles at b1 and b3 are θ , and at b2 and b4 are π − θ . The vertex
b1 lies on the positive x-axis, and b2 , b3 , b4 are arranged counter-clockwise.
(3) The probability that the algorithm will separate the pair {i, j} is exactly
Pr[Yt is absorbed in b1 or b3 ].

Proof. Part 1 is immediate from the continuity and linearity of the map O · W−1/2 .
To prove part 2, observe that the W 1/2 is given explicitly by the matrix
 
1 1 cos( θ
) + sin( θ
) cos( θ
) − sin( θ
)
W2 = √ · 2 2 2 2 .
2 cos( 2 ) − sin( 2 ) cos( 2 ) + sin( 2 )
θ θ θ θ

Taking, its inverse, we get the matrix

 
1 sec( θ2 ) + csc( θ2 ) sec( θ2 ) − csc( θ2 )
W− 2 = √ ·
1
.
8 sec( 2 ) − csc( 2 ) sec( θ2 ) + csc( θ2 )
θ θ

1
Since W− 2 [−1, 1]2 is the image of a parallelogram, it must also be a parallelogram. Moreover, one
can directly check that the diagonals are orthogonal to each other, so it must be a rhombus. It is
easy to calculate the angle between the sides and see that it is exactly θ at the image of (1, −1) and
π − θ at the image of (1, 1).
Then part 3 follows from the previous parts: if Xt is one a side or a vertex of [−1, 1], then Yt is
on the corresponding side or vertex of S. 

ACM Transactions on Algorithms, Vol. 18, No. 4, Article 33. Publication date: October 2022.
33:40 S. Abbasi-Zadeh et al.

B OMITTED PROOFS FROM SECTION 3

Recall that, for 0 ≤ x ≤ 1, 0 ≤ y ≤ 1, the function u (x, y) denotes the probability of a clause
being satisfied when the random walk walk begins with marginals (x, y) and angle θ . Equivalently,
u (x, y) is the probability that the walk, started at (x, y), ends at one of the corners (0, 0), (0, 1) or
(1, 0).
Lemma 5. For ϕ (x, y) = 1 − xy, we have
u (x) = ϕ (x) for all x ∈ ∂[0, 1]2 . (8)
Moreover, for all x in the interior of the square [0, 1]2 , u (x) = Ex [ϕ (Xτ )], where Ex denotes expectation
with respect to starting the process at X0 = x.

Proof. Recall that τ is the first time when Xt hits the boundary of [0, 1]2 , and σ is the first time
when Xt hits a vertex of [0, 1]2 . The function ϕ evaluates to 1 at the vertices (0, 0), (0, 1), and (1, 0),
and to 0 at (1, 1). Therefore, by definition, u (x) = Ex [ϕ (Xσ )].
Let us first consider the case when x is on the boundary of [0, 1]2 . Then one of the coordinates
of Xt remains fixed for the entire process. Since ϕ is affine in each of its arguments, and Xt is a
martingale, by the optional stopping theorem,
∀x ∈ ∂[0, 1]2 : u (x) = Ex [ϕ (Xσ )] = ϕ (Ex [Xσ ]) = ϕ (x).
When x is in the interior of [0, 1]2 , we have, by the law of total expectation,
∀x ∈ Int[0, 1]2 : u (x) = Ex [ϕ (Xσ )] = Ex [EXτ [ϕ (Xσ )]] = Ex [ϕ (Xτ )].
The final equality follows by the special case when the starting point of the random walk is on the
boundary of [0, 1]2 . This proves the lemma. 

Before we give an outline of the proofs of Lemmas 6 and 8, we prove Lemma 7. This will allow
us to give a completely analytic proof of 34 .
Lemma 7. Let x, y, θ be as defined by a feasible pair of vectors vi and v j . Then they must satisfy
the following constraints:
(1) 0 ≤ x ≤ 1, 0≤ y ≤ 1, 0 ≤ θ ≤ π .
xy
(2) cos(θ ) ≥ − (1−x )(1−y ) .

(1−x )(1−y )
(3) cos(θ ) ≥ − xy .

Proof. Clearly, the first set of the

√ constraints are obvious. We focus on the second and the third
constraint. Recall that vi = xv0 + x − x 2 wi and vj = yv0 + y − y 2 wj where wi and wj are unit
vectors orthogonal to v0 with cos(θ ) = wj · wj . Thus we have
√
vi · vj = xy + cos(θ ) x − x 2 y − y 2 .
But then we have the following valid constraint from the SDP:
vi · vj ≥ 0,
which implies that

xy
cos(θ ) ≥ − ,
(1 − x )(1 − y)
proving the second inequality.
ACM Transactions on Algorithms, Vol. 18, No. 4, Article 33. Publication date: October 2022.
Sticky Brownian Rounding and its Applications to Constraint Satisfaction Problems 33:41
√
For the other inequality, observe that we have v−i = (1 −x )v0 − x − x 2 wi and v−j = (1 −y)v0 −
y − y 2 wj . Then we have
√
v−i · v−j = (1 − x )(1 − y) + cos(θ ) x − x 2 y − y 2 .
But then we have the following valid constraint from the SDP:
v−i · v−j ≥ 0,
which implies that

(1 − x )(1 − y)
cos(θ ) ≥ − ,
xy
proving the third inequality. 
To ease the remainder of the presentation we first prove that the Brownian rounding algorithm
achieves an approximation of 3/4 for Max-2SAT via the maximum principle. In order to achieve
that we use the following two functions for different ranges of θ .
√
— д1 (x, y) = 1 − xy − cos(θ ) x − x 2 y − y 2 .
— f (x, y) = 1 − xy.
First consider the case when 0 ≤ θ ≤ π2 . In this case, we show д1 satisfies the requirement of the
Corollary 2 as well as give an approximation factor of 1. The last fact is trivially true since д1 is
exactly the SDP objective.
For conditions of the Corollary 2, we need to show that
∂ 2д1 ∂ 2д1 ∂ 2д1
+ + 2 cos(θ ) ≥0 ∀(x, y) ∈ Int[0, 1]2 ,
∂x 2 ∂y 2 ∂x ∂y
д1 (x, y) ≤ (1 − xy) ∀(x, y) ∈ ∂[0, 1]2 .
Since (x − x 2 )(y − y 2 ) = 0 on ∂[0, 1]2 , we obtain that д1 (x, y) = 1 − xy on ∂[0, 1]2 as required. It
remains to show that
∂2 ∂2 ∂2
д1 (x, y) + д1 (x, y) + 2 cos θ д1 (x, y) ≥ 0,
∂x 2 ∂y 2 ∂x ∂y
for all (x, y) ∈ (0, 1) 2 . Consider
∂2 ∂2 ∂2
h(x, y) := д1 (x, y) + 2 д1 (x, y) + 2 cos θ д1 (x, y).
∂x 2 ∂y ∂x ∂y
To show h is non-negative, we do the following change of variables in x = (1+sin(a)) 2 and y =
(1+sin(b ))
2 for some |a|, |b | ≤ π
2 . Such a and b exist since 0 ≤ x, y ≤ 1. Now simplifying, we obtain:
 
1 + sin(a) 1 + sin(b)
h ,
2 2
 2
= 2 cos(θ ) sec3 (a) sec3 (b) cos 2 (a) − cos2 (b)

+ 2 cos2 (a) cos2 (b) (1 − cos(θ ) sin(a) sin(b) − cos(a) cos(b)) .

Since |a|, |b | ≤ π
2 and 0 ≤ θ ≤ π
2, we have that sec(a), sec(b), cos(θ ) ≥ 0. Thus, it enough to show
that
1 − cos(θ ) sin(a) sin(b) − cos(a) cos(b) ≥ 0.

ACM Transactions on Algorithms, Vol. 18, No. 4, Article 33. Publication date: October 2022.
33:42 S. Abbasi-Zadeh et al.

Since the above expression is linear in cos(θ ), it is enough to check for extreme values of cos(θ )
which takes value between 0 and 1. It is clearly true when cos(θ ) = 0. For cos(θ ) = 1, it equals
1 − cos(a − b) and is thus non-negative.
Now consider −1 ≤ cos(θ ) ≤ 0. We show that f (x, y) = 1 − xy satisfies the condition of
Corollary 2 and is at least 34 the value of the SDP objective for all feasible (x, y, θ ). First let us focus
on the condition of Corollary 2. Clearly, the boundary conditions are satisfied by construction.
∂ 2 f (x,y ) ∂ 2 f (x,y ) ∂ 2 f (x,y )
Note that ∂x 2 = 0, ∂y 2 = 0, and that ∂x ∂y = −1. Thus,

L f = −2 cos(θ ) ≥ 0,
since cos(θ ) ≤ 0 as desired.
It remains to show that f provides an approximation
√ guarantee of 3/4 in case cos(θ ) < 0. Re-
call that SDP (x, y, θ ) = 1 − xy − cos(θ ) x − x 2 y − y 2 is the contribution of a clause to the
SDP’s objective whose two variables zi and z j have marginal values of x and y respectively and
that cos(θ ) = wi · wj . We prove the following claim, which would imply that we obtain a 34 -
approximation.
Claim 5. For any x, y, θ that satisfy the feasibility conditions in Lemma 7 and cos(θ ) < 0, we have
3
д(x, y) ≥ SDP (x, y, θ ).
4
Proof. From Lemma 7, we have
 
⎧
⎪ ⎫
⎪ xy (1 − x )(1 − y) ⎪
⎪
− cos(θ ) ≤ min ⎨
⎪ , ⎬.
⎪
⎪ (1 − x )(1 − y) xy ⎪
⎩ ⎭
Observe that we have д(x, y) ≥ 4 SDP (x, y, θ ), if
3


(1 − xy) ≥ −3 cos(θ ) (x − x 2 )(y − y 2 ).

First, suppose xy ≤ 14 . Then

  
xy
−3 cos(θ ) (x − x 2 )(y − y 2 ) ≤ 3 · (x − x 2 )(y − y 2 ) = 3xy
(1 − x )(1 − y)
≤ 1 − xy.
Else, if xy ≥ 1
4, then we have

 
(1 − x )(1 − y)
1 − xy + 3 cos(θ ) (x − x 2 )(y − y 2 ) ≥ 1 − xy − 3 · (x − x 2 )(y − y 2 )
xy
= −2 + 3x + 3y − 4xy.
Over all 1 ≥ x ≥ 0, 1 ≥ y ≥ 0 with fixed xy, the quantity 2+3x +3y −4xy is minimized when x = y.
Since xy ≥ 14 , we must have x ≥ 12 . But then it becomes −2(1 − 3x + 2x 2 ) = −2(1 − 2x )(1 − x ) ≥ 0
since 12 ≤ x ≤ 1. This proves the 34 -approximation. 
We now give a brief outline of the proof of Lemmas 6 and 8. The complete proofs involve long
sum of square expression that are available at [1].
Lemma 6. Each of д1 , д2 , д3 satisfies the boundary conditions, i.e., дi (x, y) = u (x, y) for all x, y ∈
∂[0, 1]2 and for all values θ . Moreover, we have the following for each (x, y) ∈ [0, 1]2 :

ACM Transactions on Algorithms, Vol. 18, No. 4, Article 33. Publication date: October 2022.
Sticky Brownian Rounding and its Applications to Constraint Satisfaction Problems 33:43

(1) If 1 ≥ cos(θ ) ≥ 0, then Lд1 ≥ 0.

(2) If 0 ≥ cos(θ ) ≥ − 12 , then Lд2 ≥ 0.
(3) If − 12 ≥ cos(θ ) ≥ −1, then Lд3 ≥ 0.

Proof.
Feasibility of д1 (x, y). We already showed in the above proof of 34 -approximation.
Feasibility of д2 (x, y). Now we consider д2 (x, y) = 1−xy−2 cos(θ )(x −x 2 )(y−y 2 ). Since (x −x 2 )(y−
y 2 ) = 0 on ∂[0, 1]2 , we obtain that д2 (x, y) = 1 − xy on ∂[0, 1]2 as required. It remains to show that

∂2 ∂2 ∂2
Lд2 = д2 (x, y) + 2 д2 (x, y) + 2 cos θ д2 (x, y) ≥ 0
∂x 2 ∂y ∂x ∂y
for all (x, y) ∈ (0, 1) 2 for any 0 ≥ cos(θ ) ≥ − 12 . A simple calculation allows us to obtain that

Lд2 = −2 cos(θ ) 1 + 2x 2 + 2y 2 + 2 cos(θ ) − 2x − 2y − 4y cos(θ ) − 4x cos(θ ) + 8xy cos(θ ) .

Since −2 cos(θ ) > 0, it is enough to show that for any 0 ≤ x ≤ 1 and 0 ≤ y ≤ 1,
h(x, y) = 1 + 2x 2 + 2y 2 + 2 cos(θ ) − 2x − 2y − 4y cos(θ ) − 4x cos(θ ) + 8xy cos(θ ) ≥ 0.
We now prove the above inequality. Since the above expression is linear in cos(θ ), for any fixed
x, y the minimum appears at either cos(θ ) = 0 or cos(θ ) = − 12 . First consider cos(θ ) = 0. In this
case, we obtain
1 1
h(x, y) = 1 + 2x 2 + 2y 2 − 2x − 2y = (1 − 2x ) 2 + (1 − 2y) 2 ≥ 0,
2 2
as required.
Now if cos(θ ) = − 12 , we obtain
h(x, y) = 2x 2 + 2y 2 − 4xy = 2(x − y) 2 ≥ 0
as required. This proves Lд2 ≥ 0.
Feasibility of д3 (x, y). Now we consider д3 (x, y) = 1−xy − 12 (1+5 cos(θ ))(x −x 2 )(y −y 2 )(x +y)(2−
x − y) on ∂[0, 1]2 , we obtain that д2 (x, y) = 1 − xy on ∂[0, 1]2 as required. It remains to show that

∂2 ∂2 ∂2
Lд3 = д3 (x, y) + д3 (x, y) + 2 cos θ д3 (x, y) ≥ 0,
∂x 2 ∂y 2 ∂x ∂y
for all (x, y) ∈ (0, 1) 2 for any − 12 ≥ cos(θ ) ≥ −1.
To show Lд3 ≥ 0, we consider Lд3 = p(x, y, cos(θ )) as a polynomial in x, y and cos(θ ). Replac-
ing z = cos(θ ), our aim is to show p(x, y, z) ≥ 0 if 0 ≤ x, y ≤ 1 and − 12 ≤ z ≤ −1. Equivalently,
we need to show p(x, y, z) ≥ 0 whenever r 1 (x, y, z) := x − x 2 ≥ 0, r 2 (x, y, z) := y − y 2 ≥ 0 and
r 3 (x, y, z) := −(z + 12 ) ≥ 0 and r 4 (x, y, z) := (z + 1) ≥ 0. This we show by obtaining polynomials
qi (x, y, z) for i = 0, 1, 2, 3, 4 such that qi is a sum of square polynomial of fixed degree and we have

4
p(x, y, z) = q 0 (x, y, z) + qi (x, y, z)r i (x, y, z).
i=1

Observe that above polynomial inequality shows the desired inequality. Indeed evaluate the
above identity for any 0 ≤ x, y ≤ 1 and − 12 ≥ z ≥ −1. Clearly, the RHS is non-negative. Each qi
is non-negative since it is a SoS and each r i is non-negative by construction. We mention that
we obtain these proofs via solving a semi-definite program of fixed degree (4) for each of qis. We
also remark that these SoS expressions are obtained with a small error of order δ < 10−5 . This,
formally, implies that the approximation factors of slightly worse than 78 . 
ACM Transactions on Algorithms, Vol. 18, No. 4, Article 33. Publication date: October 2022.
33:44 S. Abbasi-Zadeh et al.

Lemma 8. Consider any feasible triple (x, y, θ ) satisfying the condition in Lemma 7. We have the
following.
(1) If 1 ≥ cos(θ ) ≥ 0, then д1 (x, y) ≥ 1 · SDP(x, y, θ ).
(2) If 0 ≥ cos(θ ) ≥ − 12 , then д2 (x, y) ≥0.8749·SDP(x, y, θ ).
(3) If − 12 ≥ cos(θ ) ≥ −1, then д3 (x, y) ≥0.8749·SDP(x, y, θ ).

Proof. We prove the three inequalities. We also remark that the SoS expressions below are
obtained with a small error of order δ < 10−5 . This, formally, implies that the approximation
factors of slightly worse than 78 .
(1) If 1 ≥ cos(θ ) ≥ 0, then д1 (x, y) ≥ 1 · SDP (x, y, θ ). Observe that д1 (x, y) = SDP (x, y, θ ) and
inequality holds.
(2) If 0 ≥ cos(θ ) ≥ − 12 , then д2 (x, y) ≥ 7/8 · SDP (x, y, θ ). We need to show that
√
1 − xy − 2 cos(θ )(x − x 2 )(y − y 2 ) ≥ 0.8749 · 1 − xy − cos(θ ) x − x 2 y − y 2 ,
which holds if
√
1 − xy − 16 cos(θ )(x − x 2 )(y − y 2 ) ≥ −7 cos(θ ) x − x 2 y − y 2 .
Since both sides are non-negative (1 − xy ≥ 0 and cos(θ ) ≤ 0), it is enough to show
2
1 − xy − 16 cos(θ )(x − x 2 )(y − y 2 ) − 49 cos2 (θ )(x − x 2 )(y − y 2 ) ≥ 0,

subject to r 1 (x, y, cos(θ )) := x − x 2 ≥ 0, r 2 (x, y, cos(θ )) := y − y 2 ≥ 0, r 3 (x, y, cos(θ )) :=

− cos(θ ) ≥ 0, r 4 (x, y, cos(θ )) := xy − (1 − x )(1 −y) cos2 (θ ) ≥ 0, r 5 (x, y, cos(θ )) := (1 − x )(1 −
y) − xy cos2 (θ ) ≥ 0 where the last two constraints follow from Lemma 7. Thus again, we
construct SoS polynomials qi (x, y, cos(θ )) for 0 ≤ i ≤ 5 such that
2
1 − xy − 16 cos(θ )(x − x 2 )(y − y 2 ) − 49 cos2 (θ )(x − x 2 )(y − y 2 )

5
= q 0 (x, y, cos(θ )) + qi (x, y, cos(θ ))r i (x, y, cos(θ )).
i=1

(3) If − 12 ≥ cos(θ ) ≥ −1, then д3 (x, y) ≥ 78 SDP (x, y, θ ). The similar argument as above allows
us to obtain SoS proofs. We omit the details. 

C OMITTED PROOFS FROM SECTION 4

Baseline Approximation.
/ε )
Lemma 9. Suppose that n ≥ 2 ln(8k ε2
and ε ≤ 12 . There exists a polynomial time algorithm that on
input a satisfiable instance G = (V , E), F , and b1 , . . . , bk , as defined above, outputs a set S ⊆ V such
that, with high probability, a(δ (S )) ≥ ε2 a(E), and  |S ∩ Fi | − bi  ≤ εn for all i ∈ [k].

Proof. If the constraints specified by F and b are satisfiable, then surely the following linear
program also has a solution.

x j = bi ∀i = 1, . . . , k,
j ∈F i
0 ≤ xj ≤ 1 ∀j = 1, . . . , k.

ACM Transactions on Algorithms, Vol. 18, No. 4, Article 33. Publication date: October 2022.
Sticky Brownian Rounding and its Applications to Constraint Satisfaction Problems 33:45

We compute a solution x ∈ Rn to the program, and form a vector y ∈ Rn by defining y j =

(1 − ε)x j + ε2 for all j ∈ [n]. The vector y still satisfies the constraints approximately, i.e. for all
i ∈ [k] we have
  
  εn .
 x j − b i  ≤ (20)
∈F  2
 j i 
We now apply standard randomized rounding to y: we form a set S by independently including
any j ∈ [n] in S with probability y j . By (20), and a Hoeffding and a union bound,
⎡⎢   ⎤⎥
Pr ⎢⎢∃i :  x j − bi  > εn⎥⎥ ≤ 2ke −ε n/2 .
2

⎢⎣ j ∈F  ⎥⎦
i

By the assumption we made on n, the right hand side is at most ε4 .

Next we analyze the weight of the cut edges a(δ (S )). Any edge e = (i, j) has probability
yi + y j − yi y j ≥ 2ε (1 − ε) ≥ ε.
to be cut. Therefore, E[a(δ (S ))] ≥ εa(E). By Markov’s inequality applied to a(E) − a(δ (S )),
 ε 1−ε ε
Pr a(δ (S )) < < ε ≤ 1− .
2 1− 2 2
Therefore, the probability that S satisfies every constraint up to an additive error of εn, and
a(δ (S )) ≥ ε2 a(E) is at least ε4 . We get the high probability guarantee by repeating the entire round-
ing procedure a sufficient number of times. 
Approximation Ratio Analysis.
Lemma 11. For the SDP solution V and the Sticky Brownian Motion Xt described above, and for
any pair {i, j} of vertices
Pr[(Xτn )i  (Xτn )j ] ≥ 0.843 · vi − vj  2 .
Proof. Let us denote by θ i j the angle between the unit vectors wi and wj , i.e. θ i j =

arccos(wi , wj ). Recall that, for any i, vi  2 = x i , and vi = x i v0 + x i2 − x i wi , where v0 and
wi are orthogonal to each other. Therefore, for any pair {i, j}, vi − vj  is characterized entirely
by the triplet (x i , x j , θ ), and is equal to

vi − vj  2 = x i + x j − 2 cos(θ i j ) x i (1 − x i )x j (1 − x j ). (21)
We will refer to triplets (x, y, θ ) as configurations, and will denote the expression on the right hand
side of (21) with x i = x, x j = y, and θ i j = θ by SDP(x, y, θ ).
To calculate Pr[(Xτn )i  (Xτn )j )], we use the techniques introduced in section 3. More con-
cretely, let
uθ (x, y) = Pr (Xτn )i  (Xτn )j ) | ((X0 )i , (X0 )j ) = (x, y) .
As shown in Section 3, the function uθ is the unique solution to the partial differential equations
∂ 2uθ ∂ 2uθ ∂ 2uθ
+ + 2 cos(θ ) =0 ∀(x, y) ∈ Int[0, 1]2 , (22)
∂x 2 ∂y 2 ∂x ∂y
uθ (x, y) = x + y − 2xy ∀(x, y) ∈ ∂[0, 1]2 . (23)
The above system is a Dirichlet problem and can be solved numerically for any configuration
(x, y, θ ).

ACM Transactions on Algorithms, Vol. 18, No. 4, Article 33. Publication date: October 2022.
33:46 S. Abbasi-Zadeh et al.

To calculate the worst case approximation ratio of the Sticky Brownian Motion algorithm, it
u θ (x,y )
) . However, just taking a minimum over all (x, y, θ ) ∈ [0, 1] ×
suffices to evaluate minx,y,θ SDP(x,y,θ 2

[0, π ] is too pessimistic, since there are many configurations (x, y, θ ), which never arise as solutions
to SoS for any  ≥ 2. It is therefore necessary to consider only configurations that may arise as
solutions to some instance. In particular, we know that any vectors v0 , vi , vj in the SDP solution
satisfy the triangle inequalities (11)–(13). Translating these inequalities to inequalities involving
x i , x j and θ i j gives
 
(1 − x ) · (1 − x ) (x i · x j )
cos(θ ) ≥ max −  ,
i j
,−
xi · x j (1 − x i )(1 − x j )
 
 
x i · (1 − x j ) (1 − x i ) · x j ) 
cos(θ ) ≤ min  , .
(1 − x i ) · x j x i · (1 − x j )
 
To compute the worst case approximation ratio, we use numerical methods. In particular, we solve
the Dirichlet problem (22)–(23) for all configurations (x, y, θ ) satisfying the inequalities above,
with a granularity of 0.02 in each coordinate. This numerical computation shows that the ratio
u θ (x,y )
SDP(x,y,θ ) is at least 0.843 for all valid configurations. 
Performing a similar calculation as the one performed in the previous proof with a slowed-down
brownian motion, we get a ratio of 0.858 instead of 0.843, which is achieved at the marginals
(0.38, 0.36) and angle 0.69 · π . The code to verify this analysis is provided in [1].
Hitting Time Analysis.
Lemma 12. For any i, and any integer t ≥ 0, Pr[∀s ≤ t : 0 < (Xs )i < 1] < 4−t .
Proof. We first make some observations about Brownian motion in R. Let Z t be a standard
one-dimensional Brownian motion started in Z 0 = z ∈ [0, 1], and let σ = inf {t : Z t ∈ {0, 1}} be
the first time Z t exits the interval [0, 1]. By Theorem 2.49. in [43], E[σ ] = z(1 − z) ≤ 14 . Therefore,
by Markov’s inequality, Pr[σ > 1] < 14 . Now observe that, by the Markov property of Brownian
motion, for any integer t ≥ 0 we have

t −1
Pr[σ > t] = Pr ∀s ∈ [r , r + 1] : 0 < Z s < 1 | 0 < Z r < 1 .
r =0
But, conditional on Z r , the process {Z s }s ≥r is a Brownian motion started at Z r , and, as we observed
above, each of the conditional probabilities on the right hand side above is bounded by 14 . Therefore,
we have Pr[σ > t] < 4−t .
To prove the lemma, we just notice that, until the first time σi when (Xt )i reaches {0, 1}, it is
distributed like a one-dimensional Brownian motion started at x i . This follows because, at any
t < σi , the variance per step of (Xt )i is (Wt )i,i = Wi,i = 1. Then, by observations above, Pr[σi >
t] ≤ 4−t . 

D OMITTED PROOFS FROM SECTION 5

Hitting Times Analysis.
Claim 6. For any α ∈ (1, 2), and a δ > 0, we can bound
 ⎡ 1, α ⎤ 
x 2 · 2 F 1 ⎢⎢⎢ 2 ; x 2 ⎥⎥
⎥⎥ −
1 − (1 − x 2 ) 1−α 
 ≤ Kδ,
 ⎢⎣ 32 2(α − 1)
⎦ 
for all x ∈ [−1 + δ, 1 − δ ] where Kδ = (1−(1−δ ) 2 ) α · 2(α −1) .
2 1

ACM Transactions on Algorithms, Vol. 18, No. 4, Article 33. Publication date: October 2022.
Sticky Brownian Rounding and its Applications to Constraint Satisfaction Problems 33:47

Proof. Restating the above equation we get

⎡⎢ 1 , α ⎤⎥
1 − (1 − x 2 ) 1−α
x 2 · 2 F 1 ⎢⎢ 2 3 ; x 2 ⎥⎥ −
⎢⎣ 2 ⎥⎦ 2(α − 1)
⎡⎢ 1 , α ⎤⎥
(1 − x 2 ) 1
= x 2 · 2 F 1 ⎢⎢ 2 3 ; x 2 ⎥⎥ + −
⎢⎣ 2 ⎥⎦ 2(α − 1)(1 − x 2 ) α 2(α − 1)
 ∞
(α )t · x 2t (1 − x 2 ) 1
= x2 · + −
t =0
(2t + 1)t! 2(α − 1)(1 − x 2 )α 2(α − 1)
∞
(α )t · x 2t (1 − x 2 )  (α )t (−1) t x 2t
∞
1
= x2 · + −
t =0
(2t + 1)t! 2(α − 1) t =0
t! 2(α − 1)
∞
(α )t · x 2t x2 (1 − x 2 )(−1) t 1
=· + − .
t =0
t! 2t + 1 2(α − 1) 2(α − 1)
Taking absolute values and using the fact that |x | ≤ 1,
∞
(α )t · x 2t 1 1
≤ · +
t =0
t! (α − 1) 2(α − 1)
1 1
=2· · .
(1 − x )
2 α 2(α − 1)
Simplifying we get the bound on Kδ ≤ 2 · 1
(1−x 2 ) α
· 1
2(α −1) . 

Lemma 21. The expected hitting time E[τ ] for the diffusion process defined for Brownian Walk
Algorithm with Slowdown when the starting point X0 ∈ [−1 + δ, 1 − δ ]n and α ∈ (1, 2) is a bounded
above by Kδ = (1−(1−δ
2
· 1 .
) 2 ) α 2(α −1)
log(δ )
For α = 1, the expected hitting time is bounded by 2 · Kδ , where Kδ ≤ 2 · δ .
While the hitting time is only defined for the points away from the boundary, this is the region
where the discrete algorithm runs. Therefore, this is sufficient for the analysis of our algorithm.

Proof. Without loss of generality, we assume the number of dimensions is 1. In the one-
dimensional walk, the diffusion process satisfies the stochastic differential equation:
dXt = (1 − Xt2 ) α /2dBt . (24)
To bound the hitting time, we use Dynkin’s equation to compute stopping times, which we present
below specialised to the diffusion process at Equation (24).
Dynkin’s Equation (Theorem 7.4.1 in [47]) Let f ∈ C 2 ([−1 + δ, 1 − δ ]). Suppose μ is a finite
stopping time, then
 μ  
∂2
Ex [f (Xμ )] = f (x ) + Ex (1 − x 2 ) α · 2 f (Xs ) ds .
0 ∂x
Let f (x ) denote the function
⎡⎢ 1 , α ⎤⎥
1 − (1 − x 2 ) 1−α
f (x ) = x 2 · 2 F 1 ⎢⎢ 2 3 ; x 2 ⎥⎥ − .
⎢⎣ 2 ⎥⎦ 2(α − 1)

ACM Transactions on Algorithms, Vol. 18, No. 4, Article 33. Publication date: October 2022.
33:48 S. Abbasi-Zadeh et al.

The above function f has the following two properties:

∂2 f
(1) f satisfies12 the differential equation (1 − x 2 ) α ∂x 2 = 1.
(2) By claim 6 f (x ) is bounded in the interval [−1 + δ, 1 − δ ] by a constant. Let Kδ =
maxx ∈Rδ | f (x )|.
Let μ j = min(j, τ ) and applying Dynkin’s equation we get that
Kδ ≥ Ex [f (Xμ j )]
 μj 
= f (x ) + Ex
1ds
0
= f (x ) + Ex [μ j ].
Simplifying the above, we get
2 · Kδ ≥ Ex [μ j ].
Since we know that Ex [τ ] = limj→∞ Ex [μ j ] almost surely, we can bound 2Kδ ≥ E[τ ].
The above function is only well defined when α > 1. For the case α = 1, we simply change f to
be
1 
f (x ) = (1 + x ) log(1 + x ) + (1 − x ) log(1 − x )
2
log(δ )
with Kδ = 2 δ and the argument goes through verbatim. 

ACKNOWLEDGMENTS
The authors are grateful to Assaf Naor and Ronen Eldan for sharing their manuscript with us. SA
would like to thank Gantumur Tsogtgerel for bringing the Maximum Principle to our attention,
and Christina C. Christara for helping us compare various numerical PDE solvers. SA and AN
thank Allan Borodin for useful discussions during the initial stages of this research. GG would like
to thank Anupam Gupta and Ian Tice for useful discussion, and for also pointing out the Maximum
Principle.

REFERENCES
[1] Sepehr Abbasi-Zadeh, Nikhil Bansal, Guru Guruganesh, Aleksandar Nikolov, Roy Schwartz, and Mohit Singh. 2018.
Code for PDE Solvability and Sum of Square Proofs. Retrieved from https://github.com/sabbasizadeh/brownian-
rounding.
[2] Lars V. Ahlfors. 1966. Complex Analysis: An Introduction to the Theory of Analytic Functions of One Complex Variable
(second ed.). McGraw-Hill Book Company.
[3] Noga Alon and Assaf Naor. 2006. Approximating the cut-norm via Grothendieck’s inequality. SIAM Journal on Com-
puting 35, 4 (2006), 787–803.
[4] George E. Andrews, Richard Askey, and Ranjan Roy. 1999. Special Functions. Cambridge University Press, Cambridge.
[5] Sanjeev Arora, Eden Chlamtac, and Moses Charikar. 2006. New approximation guarantee for chromatic number. In
Proceedings of the Symposium on Theory of Computing. 215–224.
[6] Sanjeev Arora, Satish Rao, and Umesh Vazirani. 2009. Expander flows, geometric embeddings and graph partitioning.
Journal of ACM 56, 2 (April 2009), 5:1–5:37.
[7] Takao Asano. 2006. An improved analysis of goemans and Williamson’s LP-relaxation for MAX SAT. Theoretical
Computer Science 354, 3 (2006), 339–353.
[8] Takao Asano and David P. Williamson. 2002. Improved approximation algorithms for MAX SAT. Journal of Algorithms
42, 1 (2002), 173–202.
[9] Per Austrin. 2007. Balanced Max 2-SAT might not be the hardest. In Proceedings of the ACM Symposium on Theory of
Computing. 189–197.

12 We verify this using Mathematica.

ACM Transactions on Algorithms, Vol. 18, No. 4, Article 33. Publication date: October 2022.
Sticky Brownian Rounding and its Applications to Constraint Satisfaction Problems 33:49

[10] Per Austrin. 2010. Towards sharp inapproximability for any 2-CSP. SIAM Journal on Computing 39, 6 (2010), 2430–
2463.
[11] Per Austrin, Siavosh Benabbas, and Konstantinos Georgiou. 2016. Better balance by being biased: A 0.8776-
Approximation for Max Bisection. ACM Transactions on Algorithms 13, 1 (2016), 2:1–2:27.
[12] Per Austrin and Aleksa Stankovic. 2019. Global cardinality constraints make approximating some Max-2-CSPs harder.
In Proceedings of the APPROX-RANDOM (LIPIcs). 24:1–24:17.
[13] Adi Avidor, Ido Berkovitch, and Uri Zwick. 2005. Improved approximation algorithms for Max NAE-SAT and Max
SAT. In Proceedings of the Workshop on Approximation and Online Algorithms. 27–40.
[14] Nikhil Bansal. 2010. Constructive algorithms for discrepancy minimization. In Proceedings of the Foundations of Com-
puter Science. 3–10.
[15] Nikhil Bansal, Daniel Dadush, and Shashwat Garg. 2016. An algorithm for Komlós conjecture matching Banaszczyk’s
bound. In Foundations of Computer Science FOCS. 788–799.
[16] Nikhil Bansal, Daniel Dadush, Shashwat Garg, and Shachar Lovett. 2018. The Gram-Schmidt walk: A cure for the
banaszczyk blues. In Proceedings of the Symposium on Theory of Computing, STOC. 587–597.
[17] Boaz Barak, Prasad Raghavendra, and David Steurer. 2011. Rounding semidefinite programming hierarchies via global
correlation. In Proceedings of the Foundations of Computer Science. 472–481.
[18] Richard Beals and Roderick Wong. 2010. Special Functions: A Graduate Text, Vol. 126. Cambridge University Press.
[19] Ilia Binder and Mark Braverman. 2012. The rate of convergence of the walk on spheres algorithm. Geometric And
Functional Analysis 22, 3 (2012), 558–587.
[20] Avrim Blum and David Karger. 1997. An Õ (n 3/14 )-coloring algorithm for 3-colorable graphs. Information Processing
Letters 61, 1 (1997), 49–53.
[21] Moses Charikar, Venkatesan Guruswami, and Anthony Wirth. 2005. Clustering with qualitative information. Journal
of Computer and System Sciences 71, 3 (2005), 360–383.
[22] Moses Charikar, Alantha Newman, and Aleksandar Nikolov. 2011. Tight hardness results for minimizing discrepancy.
In Proceedings of the Symposium on Discrete Algorithms. 1607–1614.
[23] Moses Charikar and Anthony Wirth. 2004. Maximizing quadratic programs: Extending Grothendieck’s inequality. In
Proceedings of the Foundations of Computer Science. 54–60.
[24] Ronen Eldan and Assaf Naor. 2019. Krivine diffusions attain the Goemans-Williamson approximation ratio.
arXiv:1906.10615. Retrieved from http://arxiv.org/abs/1906.10615.
[25] Uriel Feige and Michel Goemans. 1995. Approximating the value of two power proof systems, with applications to
Max 2-SAT and Max DiCut. In Proceedings 3rd Israel Symposium on the Theory of Computing and Systems. IEEE, 0182.
[26] Uriel Feige and Michael Langberg. 2001. The RPR2 rounding technique for semidefinite programs. In Proceedings of
the International Colloquium on Automata, Languages, and Programming. 213–224.
[27] A. Frieze and M. Jerrum. 1997. Improved approximation algorithms for MAX k-CUT and MAX BISECTION. Algorith-
mica 18, 1 (01 May 1997), 67–81.
[28] David Gilbarg and Neil S. Trudinger. 2015. Elliptic Partial Differential Equations of Second Order. Springer.
[29] Emmanuel Gobet. 2000. Weak approximation of killed diffusion using Euler schemes. Stochastic Processes and Their
Applications 87, 2 (2000), 167–197.
[30] Emmanuel Gobet. 2016. Monte-Carlo Methods and Stochastic Processes. CRC Press.
[31] Michel X. Goemans and David P. Williamson. 1995. Improved approximation algorithms for maximum cut and satis-
fiability problems using semidefinite programming. Journal of the ACM 42, 6 (1995), 1115–1145.
[32] Venkatesan Guruswami and Ali Kemal Sinop. 2011. Lasserre hierarchy, higher eigenvalues, and approximation
schemes for graph partitioning and quadratic integer programming with PSD objectives. In Proceedings of the Foun-
dations of Computer Science. 482–491.
[33] Eran Halperin and Uri Zwick. 1999. Approximation algorithms for MAX 4-SAT and rounding procedures for semidef-
inite programs. In Proceedings of the Integer Programming and Combinatorial Optimization. 202–217.
[34] Eran Halperin and Uri Zwick. 2002. A unified framework for obtaining improved approximation algorithms for max-
imum graph bisection problems. Random Structures and Algorithms 20, 3 (May 2002), 382–402.
[35] Johan Håstad. 2001. Some optimal inapproximability results. Journal of the ACM 48, 4 (July 2001), 798–859.
[36] David Karger, Rajeev Motwani, and Madhu Sudan. 1998. Approximate graph coloring by semidefinite programming.
Journal of ACM 45, 2 (1998), 246–265.
[37] Howard Karloff and Uri Zwick. 1997. A 7/8-approximation algorithm for MAX 3SAT? In Proceedings of the Foundations
of Computer Science. 406–415.
[38] Subhash Khot. 2002. On the power of unique 2-prover 1-round games. In Proceedings of the Symposium on Theory of
Computing. 767–775.
[39] Subhash Khot, Guy Kindler, Elchanan Mossel, and Ryan O’Donnell. 2007. Optimal inapproximability results for MAX-
CUT and other 2-variable CSPs? SIAM Journal on Computing 37, 1 (2007), 319–357.

ACM Transactions on Algorithms, Vol. 18, No. 4, Article 33. Publication date: October 2022.
33:50 S. Abbasi-Zadeh et al.

[40] Michael Lewin, Dror Livnat, and Uri Zwick. 2002. Improved rounding techniques for the MAX 2-SAT and MAX DI-
CUT problems. In Proceedings of the Integer Programming and Combinatorial Optimization. 67–82.
[41] Shachar Lovett and Raghu Meka. 2015. Constructive discrepancy minimization by walking on the edges. SIAM Journal
on Computing 44, 5 (2015), 1573–1582.
[42] Tomomi Matsui and Shiro Matuura. 2001. 0.935-approximation randomized algorithm for Max 2-SAT and its deran-
domization. Department of Mathematical Engineering and Information Physics, University of Tokyo (Technical Report
METR 2001-03) (2001).
[43] Peter Mörters and Yuval Peres. 2010. Brownian Motion. Cambridge Series in Statistical and Probabilistic Mathematics.
[44] Elchanan Mossel, Ryan O’Donnell, and Krzysztof Oleszkiewicz. 2005. Noise stability of functions with low influences:
Invariance and optimality. In Proceedings of the Foundations of Computer Science. 21–30.
[45] Y. Nesterov. 1998. Semidefinite relaxation and nonconvex quadratic optimization. Optimization Methods and Software
9, 1–3 (1998), 141–160.
[46] Ryan O’Donnell and Yi Wu. 2008. An optimal Sdp algorithm for Max-Cut, and equally optimal long code tests. In
Proceedings of the Symposium on Theory of Computing. 335–344.
[47] Bernt Øksendal. 1995. Stochastic Differential Equations (fourth ed.). Springer. An introduction with applications.
[48] Prasad Raghavendra. 2008. Optimal algorithms and inapproximability results for every CSP? In Proceedings of the 40th
Annual ACM Symposium on Theory of Computing. ACM, 245–254.
[49] Prasad Raghavendra and David Steurer. 2009. How to round any CSP. In Proceedings of the Foundations of Computer
Science. 586–594.
[50] Prasad Raghavendra and Ning Tan. 2012. Approximating CSPs with global cardinality constraints using SDP hierar-
chies. In Proceedings of the Symposium on Discrete Algorithms. SIAM, 373–387.
[51] Chaitanya Swamy. 2004. Correlation clustering: Maximizing agreements via semidefinite programming. In Proceedings
of the Symposium on Discrete Algorithms. 526–527.
[52] Norbert Wiener. 1923. Differential-Space. Journal of Mathematics and Physics 2, 1–4 (1923), 131–174. DOI:https://doi.
org/10.1002/sapm192321131
[53] Yinyu Ye. 2001. A .699-approximation algorithm for Max-Bisection. Mathematical Programming 90, 1 (01 Mar 2001),
101–111.
[54] Uri Zwick. 1998. Approximation algorithms for constraint satisfaction problems involving at most three variables per
constraint. In Proceedings of the Symposium on Discrete Algorithms. 201–210.
[55] Uri Zwick. 1999. Outward rotations: A tool for rounding solutions of semidefinite programming relaxations, with
applications to MAX CUT and other problems. In Proceedings of the Symposium on Theory of Computing. ACM,
679–687.

Received January 2020; revised December 2020; accepted March 2021

ACM Transactions on Algorithms, Vol. 18, No. 4, Article 33. Publication date: October 2022.