Degrees of Freedom for Critical Random 2-SAT

∗, 1 †, 1, 2
Andreas Basse-O’Connor, and Mette Skjøtt,
1
Department of Mathematics, Aarhus University, Denmark
2
Kvantify ApS, DK-2300 Copenhagen S, Denmark
arXiv:2505.15940v1 [math.PR] 21 May 2025

Abstract
The random k-SAT problem serves as a model that represents the ’typical’ k-SAT instances.
This model is thought to undergo a phase transition as the clause density changes, and it
is believed that the random k-SAT problem is primarily difficult to solve near this critical
phase. In this paper, we introduce a weak formulation of degrees of freedom√ for random k-SAT
problems and demonstrate that the critical random 2-SAT problem has 3 n degrees of freedom.
This quantity represents the maximum number of variables that can √ be assigned truth values
without affecting the formula’s satisfiability. Notably, the value of 3 n differs significantly from
the degrees of freedom in random 2-SAT √ problems sampled below the satisfiability threshold,
where the corresponding value equals n. Thus, our result underscores the significant shift in
structural properties and variable dependency as satisfiability problems approach criticality.

∗
basse@math.au.dk
†
metteskjott@math.au.dk

1
1 Introduction
1.1 Background and motivation
The Boolean satisfiability problem (SAT) is a highly studied topic in computer science, notable for
being the first problem proven to be NP-complete, see [Coo71]. Its versatility extends beyond the-
oretical interest, with practical applications in areas like artificial intelligence, software verification,
and optimization (see [MS08, GGW06, Knu15], and references therein). In recent years, SAT has
also attracted significant attention in the fields of discrete probability and statistical physics. This
interdisciplinary interest arises because SAT exhibits behaviors, such as phase transitions, making
it a compelling subject for studying threshold behavior in combinatorial structures.
A SAT instance is a Boolean function that evaluates multiple Boolean variables and returns
a single Boolean value. The function is typically expressed in conjunctive normal form (CNF),
meaning it is a conjunction (and) of disjunctions (or) of literals. Each literal represents a variable
or its negation. A formula in which every clause contains exactly k literals is called a k-CNF
formula. The following is an example of a 2-CNF formula with four variables and five clauses:

φ(x) = (x1 ∨ x2 ) ∧ (¬x2 ∨ x3 ) ∧ (¬x3 ∨ x4 ) ∧ (¬x1 ∨ ¬x4 ) ∧ (x2 ∨ x4 ), (x ∈ {true, false}4 ).

The only assignment that makes the above formula evaluate to true is (false, true, true, true).
The objective of the satisfiability problem is to determine whether such an assignment exists; if
so, we write φ ∈ SAT. In the context of computational complexity theory, the 2-SAT problem is
NL-complete, meaning it can be solved non-deterministically with logarithmic storage space and
is one of the most difficult problems within this class (see Thm. 16.3 in [Pap03]). Consequently,
a deterministic algorithm that solves 2-SAT using only logarithmic space would imply L = N L,
which is a standing conjecture. For k ≥ 3, the k-SAT problem is NP-complete, situating it at the
core of the famous P vs. N P conjecture.
In practical applications, SAT instances are, in most cases, easily solvable, which appears to
contradict the problem’s computational hardness. This observation inspired the development of
the random k-SAT model, designed to generate typical SAT instances, see [Gol79, CKT+ 91, KS94,
GW94]. In this model, the number of input variables n, clauses m, and the clause size k are fixed.
Clauses are then sampled independently and uniformly from the 2 k clauses with non-overlapping
k n

variables. This model is called the random k-SAT model, and the distribution is denoted Fk (n, m).
This model becomes particularly interesting when n and m grow large simultaneously. Specifically,
by setting m = ⌊αn⌋, where α > 0 represents the clause density, the random k-SAT problem is
believed to undergo a phase transition: the asymptotic probability of satisfiability shifts from one
to zero as α surpasses a critical threshold αk , that is for k ≥ 2,

1,
(
α < αk ,
lim P Fk (n, ⌊αn⌋) ∈ SAT = (1.1)


n→∞ 0, α > αk .

A random k-SAT problem that is satisfiable w.h.p. is referred to as under-constrained, while it is

called over-constrained when it is unsatisfiable w.h.p. Furthermore, when a phase transition exists,
problems at this critical value are referred to as being critical.
As previously discussed, SAT problems are computationally challenging. Notably, it is near
the expected phase transition of the random k-SAT model that the hardest instances are thought
to arise, see [SML96]. Figure 1 displays how a spike in computational hardness appears when
the clause density approaches the expected phase transition. This highlights why understanding
the behavior of random k-SAT in this critical region is of substantial theoretical and practical

2
Figure 1: Computational hardness of random 3-SAT as a function of clause density. The y-axis
displays the median resolution time of 10,000 instances solved using the DPLL algorithm. Credit:
Fig. 8.2 in [BCM02]

importance. More broadly, the study of random structures near critical transitions is a significant
and complex area of research. The prominence of this field is underscored by the fact that three
Fields Medals have been awarded since 2006 for groundbreaking work on critical phenomena, with
recipients including H. Duminil-Copin, S. Smirnov, and W. Werner.
The phase transition phenomenon was in 1992 established for k = 2 in the articles [Goe96,
CR92, dLV01], where the authors independently established that α2 = 1. Recently, the sharp
satisfiability conjecture (1.1) has been affirmatively verified for all k ≥ k0 , with k0 being a large
and unknown constant, see [DSS15]. The remaining cases of k constitute an open problem. In
1999, the result on random 2-SAT was further refined in [BBC+ 01] as the rate of convergence
was determined. Additionally, it was shown that the asymptotic probability of satisfiability of a
random critical 2-SAT problem is bounded away from both zero and one, though whether this
probability converges remains an open question. Recent contributions to the random 2-SAT model
have focused on the under-constrained regime, where both the expected number of solutions and a
central limit theorem for this quantity (see [ACOHK+ 21, CCOM+ 24]) has been established. Thus,
while the phase transition of random 2-SAT was proven many years ago, ongoing research continues
to uncover new insights into the model, and several open questions remain unresolved.
A recent study [BOOS25] examined variable interactions by analyzing the degrees of freedom
in under-constrained random k-SAT problems. This concept refers to the number of variables that
can be fixed without impacting the formula’s satisfiability. For under-constrained random 2-SAT
problems, where α < 1, the degrees of freedom equal n1/2 , while in random 3-SAT problems well
below the phase transition (α < 3.145), the degrees of freedom equal n2/3 .
In this paper, we compute the degrees of freedom in critical 2-SAT problems. Our result shows
that in this critical setting, the degrees of freedom decrease with a polynomial factor, scaling only
as n1/3 . This finding underscores the emergence of complex structures near the phase transition,
where variable interdependencies become significantly more pronounced. Thus, our results highlight
this marked shift in variable correlation as random SAT problems approach criticality.

3
1.2 Main result
Consider a random 2-CNF formula Φ sampled at the phase-transition point of the random 2-SAT
problem, where the asymptotic probability of satisfiability shifts from one to zero. We aim to
determine how many input variables are free—that is, they can be assigned any value without
effecting the asymptotic probability that the formula is satisfiable.
Let L ⊆ ±[n] := {−n, . . . , −1, 1, . . . , n} be a set with |L| = f (n) elements, chosen such that if
ℓ ∈ L, then −ℓ ∈ / L (we say that L is consistent). This set L dictates the variables being fixed,
having xv = true when v ∈ L and xv = false when −v ∈ L. Formally, let B = {true, false}.
For x ∈ Bn , we define xL ∈ Bn as the vector with (xL )v = true when v ∈ L, (xL )v = false when
−v ∈ L, and (xL )v = xv for all other entries. We then consider

ΦL (x) = Φ(xL ). (1.2)

Note that ΦL denotes the mapping Φ with f variables fixed to values specified by L. Our goal is to
identify the threshold value of f that separates instances where ΦL remains solvable with positive
probability from those where ΦL becomes unsatisfiable. To formalize this notion, we introduce
the following definition, where we recall that Fk (n, m) denotes a random k-CNF formula with n
variables and m clauses.

Definition 1. The random k-SAT problem with clause density α > 0 is said to have f⋆ (n) degrees
of freedom weakly if, for Φ ∼ Fk (n, ⌊αn⌋), every consistent subset L ⊆ ±[n] with |L| = f (n), and
for all ε > 0, the following holds:

(1) Whenever f = O(f⋆ n−ε ), then

lim inf P(ΦL ∈ SAT) = lim inf P(Φ ∈ SAT) > 0.

n→∞ n→∞

(2) Whenever f = Ω(f⋆ nε ), then

lim P(ΦL ∈ SAT) = 0.
n→∞

Condition (1) states that fixing strictly fewer than f⋆ variables does not decrease the lower
bound on the probability of satisfiability. On the other hand, condition (2) implies that when
fixing strictly more than f⋆ variables, the problem becomes unsatisfied. This concept is a weaker
form of the degrees of freedom notion introduced in [BOOS25]; specifically, having f⋆ degrees of
freedom implies having f⋆ degrees of freedom weakly. Note that f⋆ is unique up to sub-polynomial
factors, meaning that if both f⋆ and g⋆ are weak degrees of freedom, then for any ε > 0, we have
f⋆ n−ε ≤ g⋆ ≤ f⋆ nε for sufficiently large n. Our main result is the following:

Theorem 2. The random critical 2-SAT problem has n1/3 degrees of freedom weakly.

We recall that critical refers to the situation with α = 1. Figure 2 shows simulations indicating
that, as n increases, the curve representing the satisfiability of the random critical 2-SAT problem
as a function of the number of fixed variables becomes increasingly steep. Moreover, this steepening
behavior points to a cutoff occurring at n1/3 .

1.3 Related work

In this section, we compare our results to related work, providing new insights and situating our
findings within a broader context.

4
Figure 2: Satisfiability of random critical 2-SAT as a function of the number of fixed variables.
The different curves represent a varying number of input variables. Each data point is comprised
of 2, 000 simulations. The vertical dotted line indicates q = 1/3.

Remark 3. Theorem 2 allows us to compare the critical random 2-SAT problem with general
random k-SAT problems:

• The paper [BOOS25] established that under-constrained random 2-SAT problems have n1/2
degrees of freedom, which reveals a pronounced difference with the behavior observed at the
critical phase-transition point. At this threshold, a dramatic reduction in degrees of freedom
occurs, reflecting a fundamental shift in the underlying structure of the formula. This is not
surprising, as at this critical ratio, the system is on the "knife edge" between being satisfiable
versus unsatisfiable, and therefore, long-range correlations between variables are expected to
appear. To our knowledge, our result is one of the first to indicate this drastic change in
variable dependence.

• The paper [BOOS25] also examines random 3-SAT problems, establishing that when α is
significantly below the expected phase-transition threshold (α < 3.145), the degrees of freedom
are n2/3 . In comparison, our main theorem shows that the degrees of freedom in critical
random 2-SAT equal the square root of this amount, indicating a notable contrast in variable
flexibility between the two cases.

The computational hardness of the satisfiability problem implies that finding solutions to chal-
lenging SAT formulas often requires traversing a substantial portion of the search tree, that is,
assigning truth values to variables sequentially and backtracking when encountering contradic-
tions. This approach forms the core of the DPLL algorithm, introduced in 1962 as one of the
first SAT-solving algorithms, [DLL62]. Decades later, in the 1990s, CDCL (Conflict-Driven Clause
Learning) solvers transformed SAT solving, enabling the solution of instances with thousands or
even millions of variables. Despite their modern enhancements, these solvers still rely on the simple
procedure of assigning truth values and backtracking (see p. 62 in [Knu15]). The concept of degrees
of freedom quantifies how deep one can navigate in the search tree before a contradiction arises
when solving a random SAT problem. Moreover, the drastic change in degrees of freedom when
comparing under-constrained problems with critical problems highlights why computational com-

5
plexity intensifies near the satisfiability threshold. This also aligns with the observations in Figure
1, which displayed the computation time of the DPLL algorithm when approaching criticality.
Let again Φ ∼ F2 (n, n), L ⊆ ±[n] be consistent with |L| = f (n), and remember that fixing
variables corresponds to shrinking the input space. Thus, it is clear that {ΦL ∈ SAT} ⊆ {Φ ∈ SAT}.
This along with our main theorem implies that whenever f = O(n1/3−ε ) for an ε > 0 we have

lim inf P Φ ∈ SAT ≤ lim inf P ΦL ∈ SAT ≤ lim sup P ΦL ∈ SAT ≤ lim sup P Φ ∈ SAT . (1.3)





n→∞ n→∞ n→∞ n→∞

Thus, if P(Φ ∈ SAT) has a limit as n → ∞, then P(ΦL ∈ SAT) also has a limit, and these two limits
coincide. In [BBC+ 01] it is shown that for all δ > 0 sufficiently small, there exists a cδ > 0 such
that if Φα ∼ F2 (n, ⌊αn⌋) with α ∈ [1 − cδ n−1/3 , 1 + cδ n−1/3 ], then

δ ≤ P(Φα ∈ SAT) ≤ 1 − δ. (1.4)

Moreover, this interval is the best possible in the sense that if a sufficiently large constant replaces
cδ , the statement becomes false. Combining (1.3) and (1.4) we get that for δ > 0 small enough

δ ≤ lim inf P ΦL ∈ SAT ≤ lim sup P ΦL ∈ SAT ≤ 1 − δ,

n→∞ n→∞

so the limiting probability is bounded away from zero and one, and this interval is not larger than
the corresponding interval for satisfiability when no variables are fixed. We observe that the length
of the scaling window in [BBC+ 01] is on the order of n−1/3 , which is the reciprocal of the degrees
of freedom for the critical 2-SAT problem. However, the proof presented in [BBC+ 01] differs from
that of the current paper, and there is no direct coupling between the two results.
The main idea of the proof in [BBC+ 01] is to consider an order parameter for the phase transition
of random 2-SAT. This is a concept often used in statistical physics and it refers to a function that
vanishes on one side of a transition and becomes non-zero on the other side. The order parameter
that they consider is the average size of the spine, where the spine of a CNF-formula φ is defined to
be the set of literals ℓ for which there is a satisfiable sub-formula ψ of φ with ψ ∧ℓ not satisfiable. By
carefully controlling this quantity in a random CNF-formula as clauses are added one by one their
result follows. Note that the size of the spine equals the number of variables that are free to be given
any truth value without making a satisfiable SAT problem unsatisfiable. The spine only describes
how each variable on its own affects the satisfiability of a CNF-formula. In contrast, we need
to understand how all the fixed variables simultaneously impact the satisfiability of the formula.
Multiple other papers, e.g. [CF86, ACIM01, AKKK01] also consider the procedure of fixing one
single variable at a time, and in [Ach00] they consider fixing two variables at a time. This is
different from the approach in the present paper where many variables are fixed simultaneously
and hereby long implication chains emerge that intervene with each other and affect satisfiability.
As previously mentioned, the paper [BOOS25] was the first to introduce and compute degrees
of freedom in certain random under-constrained k-SAT problems. Their proof is based on the idea
that fixing variables in a CNF-formula creates clauses of size one, also called unit-clauses. The
presence of these unit-clauses, in turn, corresponds to further variable fixing. Thus, variables are
fixed repeatedly in rounds, and the probability of encountering a contradiction in each round is
calculated. The sequence describing the number of fixed variables throughout the rounds is then
studied. This procedure is closely related to the unit-propagation algorithm, a well-studied tech-
nique used as a subroutine in most modern SAT solvers. We also base our proof on an appropriate
adaptation of the unit propagation algorithm. In the under-constrained regime of random 2-SAT,
it is possible to control the number of unit-clauses produced in each round r, and this number de-

6
creases exponentially at a rate of α, i.e. as αr . However, at the phase transition, we have α = 1, and
thus the expected number of unit-clauses produced in each round remains approximately constant
(αr = 1). As a result, controlling unit propagation becomes more challenging because the entire
process must be analyzed as a whole, unlike in the under-constrained regime, where the rounds
could be considered independently. This again suggests the presence of long-range correlations
between variables when α = 1. When f = Ω(n1/3+ε ) for some ε > 0 the key idea is to show that
the number of unit-clauses produced in each round remains high for a certain number of rounds
w.h.p. This implies that a contradiction is likely to occur before the process terminates. On the
other hand, when f = O(n1/3−ε ) for some ε > 0 we show that the sequence dies out w.h.p. before
encountering a contradiction.
The results of [BOOS25] extend further as they also determine the limiting satisfiability of the
random SAT problem when Θ(f⋆ ) variables are fixed, where f⋆ represents the degrees of freedom
of the random formula. In this setting, they show that the limiting probability remains bounded
away from zero and one, and they provide the exact limiting value. By adjusting a parameter, this
limiting value smoothly interpolates between the two edge cases. An open question is whether a
similar result holds for the random critical 2-SAT problem. Specifically, it remains unknown what
happens when Θ(n1/3 ) variables are fixed in such formulas, and whether the limiting probability
will also interpolate between the edge cases.

2 Preliminaries
2.1 Notation and conventions
For any set A ⊆ Z we define −A = {−a : a ∈ A}, ±A = A ∪ (−A) and we denote by |A| the
number of elements in A. For elements xi , i ∈ A belonging to some space we let (xa )a∈A denote
the vector (xa1 , . . . , xa|A| ), where {a1 , . . . , a|A| } = A and a1 < a2 < · · · < a|A| . Furthermore, for
any n, m ∈ N with m < n we let [n] = {1, . . . , n}, [m, n] = {m, m + 1, . . . , n}, and [0] = ∅. The two
sets B = {true, false} and K = {0, 1, 2, ⋆} are also considered repeatedly. For an x ∈ R we let
x+ = max{0, x}.
When considering random elements a probability space (Ω, F, P) will always be given. When-
ever new random elements are introduced, unless specified otherwise, they are independent of all
previously existing randomness. We define 00 = 0. As we will ultimately let n approach infinity,
certain inequalities will hold only for sufficiently large n. In such cases, the required size of n for
the inequality to hold may depend on q, but it will always be independent of the round r. As has
been the case thus far, n is often omitted from the notation, even though most elements depend
on this parameter.

2.2 The random SAT-problem

Let n, m ∈ N0 and k ∈ N, where n ≥ k when m > 0. The random k-SAT distribution was defined
in section 1.2, but we will infer some additional notation needed for our proof. Firstly, we will
specify the non-random case. When m > 0 we let a k-clause over n variables be a vector from the
set
D = (ℓ1 , . . . , ℓk ) ∈ (±[n])k : |ℓ1 | < · · · < |ℓk | .


7
The entries of such a vector are called the literals of the clause. Consider m such clauses (ℓj,i )i∈[k] ,
j ∈ [m]. From these clauses we define a k-SAT formula φ with n variables and m clauses by letting
m
φ=
^

ℓj,1 ∨ · · · ∨ ℓj,k .
j=1

We let the order of the clauses matter such that two formulas φ and φ′ with literals ((ℓj,i )i∈[k] )j∈[m]
and ((ℓ′j,i )i∈[k] )j∈[m] , respectively, are equal if and only if ℓj,i = ℓ′j,i for all j ∈ [m] and i ∈ [k]. This
implies a one-to-one correspondence between a formula and its (ordered set of) literals. Now, we
define a mapping related to a SAT-formula. For ℓ ∈ ±[n] we associate a mapping by letting

if sgn(ℓ) = 1,
(
x|ℓ| ,
ℓ : B → B,
n
where ℓ : x = (x1 , . . . , xn ) 7→ (2.1)
¬ x|ℓ| , if sgn(ℓ) = −1.

Letting ∧ denote the logical and and ∨ denote the logical or, we associate φ with the function
mapping Bn to B that is given by
m
^  m
φ(x) = (x) = ℓj,1 (x) ∨ · · · ∨ ℓj,k (x) ,
^
x ∈ Bn .



ℓj,1 ∨ · · · ∨ ℓj,k
j=1 j=1

We now define a distribution over the set of k-SAT formulas with n variables and m clauses and
we denote this distribution by Fk (n, m). Consider random vectors (Lj,i )i∈[k] , j ∈ [m], that are
uniformly distributed on D. We say that these are random clauses. Furthermore, let
m
Φ=
^

Lj,1 ∨ · · · ∨ Lj,k ,
j=1

then Φ has distribution Fk (n, m) and we say that Φ is a random k-SAT formula with n variables
and m clauses. For x ∈ Bn we let Φ(x) denote the point-wise evaluation of Φ in x.

2.3 Fixing variables and the unit-propagation algorithm

Let n, m ∈ N0 and k ∈ N with n ≥ k when m > 0. Let L ⊆ ±[n] be consistent. For an x ∈ Bn
we let the vector xL be as defined in subsection 1.2 and for functions g : Bn → B we define
gL (x) = g(xL ). Consider a 2-SAT formula φ with n variables and m clauses where its literals are
denoted ((ℓj,i )i∈[2] )j∈[m] . Consider the formula with fixed variables
m m
φL = = (ℓj,1 )L ∨ (ℓj,2 )L .
^
^

ℓj,1 ∨ ℓj,2 L
j=1 j=1

The set [m] is now split into four non-overlapping subsets:

C0 = j ∈ [m] : ℓj,1 ∈ −L and ℓj,2 ∈ −L ,



C1 = j ∈ [m] : ℓj,ij1 ∈
/ ±L and ℓj,ij2 ∈ −L, {ij1 , ij2 } = {1, 2} ,

(2.2)
C2 = j ∈ [m] : ℓj,1 ∈
/ ±L and ℓj,2 ∈

/ ±L ,
C⋆ = j ∈ [m] : ℓj,1 ∈ L or ℓj,2 ∈ L .


Using the definition of ij1 from above we ease notation and let ℓj,ij1 = ℓj for j ∈ C1 . Note that

8
 • When j ∈ C0 then (ℓj,1 ∨ ℓj,1 )L (x) = false for all x ∈ Bn .

• When j ∈ C1 then (ℓj,1 ∨ ℓj,2 )L (x) = ℓj (x) for all x ∈ Bn .

• When j ∈ C2 then (ℓj,1 ∨ ℓj,2 )L (x) = (ℓj,1 ∨ ℓj,2 )(x) for all x ∈ Bn .

• When j ∈ C⋆ then (ℓj,1 ∨ ℓj,2 )(x) = true for all x ∈ Bn .

Define
φ1 = (ℓj ), and φ2 = (ℓj,1 ∨ ℓj,2 ). (2.3)
^ ^

j∈C1 j∈C2

Note that the above literals will belong to the set (±[n]\ ± L). The above implies that φ ∈ SAT
if and only if C0 = ∅ and (φ1 ∧ φ2 ) ∈ SAT. We will now further determine when (φ1 ∧ φ2 ) ∈ SAT.
Define
L(φ1 ) = {ℓj ∈ C1 : −ℓj ∈
/ C1 } (2.4)
We let this be the set associated with the 1-SAT formula φ1 , and we note that it is a consistent
set. Moreover, for x ∈ Bn

when sgn(ℓj ) = +1,

(
true,
φ1 (x) = true ⇐⇒ x|ℓj | = ∀j ∈ C1 . (2.5)
false, when sgn(ℓj ) = −1.

This along with the definition of xL(φ1 ) implies that when φ1 ∈ SAT then for all x ∈ Bn we have
that φ1 (x) = true if and only if x = xL(φ1 ) . Therefore

(φ1 ∧ φ2 ) ∈ SAT ⇐⇒ φ1 ∈ SAT and (φ1 ∧ φ2 )L(φ1 ) ∈ SAT ⇐⇒ φ1 ∈ SAT and (φ2 )L(φ1 ) ∈ SAT.

Thus
φ ∈ SAT ⇐⇒ C0 = ∅, φ1 ∈ SAT, and (φ2 )L(φ1 ) ∈ SAT. (2.6)
This decomposition of the event {φ ∈ SAT} becomes a key tool in the proof. Moreover, note that
the same procedure, as just described, can now be applied to the formula (φ2 )L(φ1 ) . Hence, the
procedure of fixing variables continues recursively in rounds and this is the idea behind the unit-
propagation algorithm. One of the main ingredients in the proof of our main theorem concerns
controlling this process.

2.4 Sketch of proof

Consider a random 2-CNF formula Φ ∼ F2 (n, n) with literals (Lj,i )i∈[2],j∈[n] , and a consistent set
L ⊆ ±[n] with |L| = f . We now apply the unit-propagation procedure to Φ, thereby decomposing
the probability of interest into a collection of simpler terms.
(1)
Initial round: Let Ck , for k ∈ K, be the random sets defined from Φ and L as described in
(1) (1) (1) (1)
(2.2), and define Mk = |Ck | for k ∈ K. Additionally, let Φ1 and Φ2 be the random formulas
constructed from Φ and L, corresponding to the definitions in (2.3). Finally, let L(1) denote the set
(1)
associated with Φ1 , as defined in (2.4). From the decomposition in (2.6), we get that

P(ΦL ∈ SAT) = P(M0

(1) (1) (1)
= 0, Φ1 ∈ SAT, (Φ2 )L(1) ∈ SAT). (2.7)

The independence of the clauses of Φ implies that the three events in (2.7) only are dependent
(1)
through the random vector (Mk )k∈[K] . Moreover, the i.i.d. structure of the clauses in Φ implies
that this vector has a multinomial distribution, where the entries concentrate around their mean

9
and hence become asymptotically independent. This implies that also the events in (2.7) are
asymptotically independent, allowing for the desired decomposition:
(1) (1)
= 0 P Φ1 ∈ SAT P (Φ2 )L(1) ∈ SAT .
(1)
(2.8)




P(ΦL ∈ SAT) ∼ P(M0

Subsequent rounds: The procedure from the initial round is now repeated recursively, replacing
(1) (2) (2) (2)
Φ and L with Φ2 and L(1) , respectively. Hereby, new random elements (Mk )k∈K , Φ1 , Φ2 ,
(2)
and L(2) are constructed. The procedure is then repeated iteratively on Φ2 and L(2) , and so on.
Continuing a total of R times (R being some suitable integer), we construct the random elements
(r) (r) (r)
(Mk )k∈K , Φ1 , Φ2 , and L(r) for each r ∈ [R]. Using these constructed elements, the probability
calculation in (2.8) can be extended iteratively, leading to
R
(R) (r) (r)
P(ΦL ∈ SAT) ∼ P (Φ2 )L(R) ∈ SAT = 0 P Φ1 ∈ SAT . (2.9)

Y


P M0
r=1

The probabilistic decomposition in (2.9) plays a central role in the overall proof. To evaluate the
terms of (2.9), we need to know the distributions of the defined elements. As the elements are defined
recursively, the distributions can be found as conditional distributions, and when conditioning on
the past, we get that
(r−1) 2
M1
 
(r) (r−1) (r) (r) (r)
M0 |M1 ≈ Binomial ,n , and Φ1 |M1 ≈ F1 (n, M1 ), (r ∈ [R]).
2n
(r)
Thus, it becomes crucial to control the size of the sequence (M1 )r∈[R] , and the remaining part of
the proof concerns this.
Firstly, we establish that limn→∞ P(ΦL ∈ SAT) = 0 when f = nq with q = 1/3 + ε. Here we will
prove the existence of constants c, C > 0, such that
(r)
lim P M1 ∈ [cnq , Cnq ] ∀r ∈ [R] = 1, (2.10)


n→∞

which will imply that the product in (2.9) approaches zero and thus, this implies our main result.
When proving (2.10) a simple union bound will not do, and thus we will need to exploit the Markov
(r)
structure of the sequence (M1 )r∈[R] .
Next, we will establish that lim inf n→∞ P(ΦL ∈ SAT) ≥ lim inf n→∞ P(Φ ∈ SAT), when f = nq
(r)
with q = 1/3 − ε. In this setup, the sequence (M1 )r∈[R] is a super-martingale, and thus optional
(r)
sampling gives that M1 ≤ log n · nq for all r ∈ [R]. This further implies that the product in
(r)
(2.9) approaches one as n → ∞. Next, we will establish that the sequence (M1 )r∈[R] is close in
(R)
distribution to a critical Galton-Watson tree, and from this we can establish that M1 = 0 w.h.p.,
(R)
which implies that L(R) = ∅ w.h.p. This further gives that (Φ2 )L(R) is close in distribution to
Φ, and thus the first term of (2.9) is asymptotically equivalent to P(Φ ∈ SAT). Thus, this finally
proves our main theorem.

3 Main decomposition of probability

In this section, we present a mathematically rigorous version of the decomposition in 2.4. This
decomposition will break the proof of our main result into smaller lemmas, which will be proven

10
later. In subsection 3.1, we introduce the technical lemmas that primarily provide distributional
results for the sequences of elements that will be defined in sections 3.2 and 3.3. The two sequences
defined in these sections both serve as approximations to the unit propagation procedure. Section
3.2 addresses the case q > 1/3, where the corresponding sequence is used to establish an upper
bound on the probability, which approaches zero. In Section 3.3, the other sequence provides a
lower bound that is used for the proof in the case q < 1/3.

3.1 Technical lemmas

The first lemma of this section states that for Φ ∼ F2 (n, m) and a consistent set of literals L ⊆ ±[n],
with |L| = f we can construct a coupled SAT-formula Φ′ which has the same distribution as Φ but
where fixing the literals of L in Φ corresponds to fixing the literals of the set [n]\[n − f ]. When
considering the different rounds of the unit-propagation algorithm later on, the repeated use of this
lemma will allow us to control which variables are fixed.

Lemma 4. There exists a function G such that if Φ ∼ F2 (n, m) and L ⊆ ±[n] is a consistent set
D
of literals with |L| = f , then Φ′ := G(Φ, L) = Φ and

ΦL ∈ SAT = Φ′[n]\[n−l] ∈ SAT .

 

The below is an easy consequence of the above lemma.

Fact 5. Let Φ ∼ F2 (n, m) and let L ⊆ ±[n] be a consistent random set of literals independent of
D
Φ. Then G(Φ, L) = Φ and G(Φ, L) is independent of L.

The proof of Lemma 4 relies on the uniformity of the clauses that imply that literals can be
swapped without changing the distribution of the formula.
Next, we want to decompose a 2-CNF formula with fixed variables into its 1- and 2-CNF
sub-formulas. Let φ be a (non-random) 2-CNF formula with n variables and m clauses and let
L = [n]\[n − f ] for some f ∈ N. Define the sets

A0 = A0 (n, f ) := −L × −L, A1 = A1 (n, f ) := ±[n − l] × −L,

A2 = A2 (n, f ) := ±[n − f ] × ±[n − f ], A⋆ = A⋆ (n, f ) := ±[n] × L.

Let (ℓj,1 , ℓj,2 ), j ∈ [m], be the literals of φ and define Ck = {j ∈ [m] : (ℓj,1 , ℓj,2 ) ∈ Ak }, k ∈ K. Note
that this definition corresponds to the definition in (2.2). A clause that belongs to A0 is said to be
an unsatisfied clause, and a clause in A⋆ is said to be satisfied. Define

G1 (φ, f ) := G2 (φ, f ) = (ℓj,1 ∨ ℓj,2 ).

^ ^
ℓj,1 ,
j∈C1 j∈C2

In (2.6) we saw that

φL ∈ SAT ⇐⇒ C0 = ∅, G1 (φ, f ) ∈ SAT, G2 (φ, f )L(G1 (φ,f )) ∈ SAT,

where L(G1 (φ, f )) is defined in (2.4). In the setup with L = [n]\[n − f ] we further note that when
(ℓj,1 , ℓj,2 ) ∈ A1 , then ℓj,1 ∈ ±[n − f ] and when (ℓj,1 , ℓj,2 ) ∈ A2 then (ℓj,1 , ℓj,2 ) ∈ (±[n − f ])2 . Hence
both G1 (φ, f ) and G2 (φ, f ) can be viewed as boolean functions that map {±1}n−f into {±1}.
The above setup will now be applied to a random 2-CNF formula. The next lemma describes the
simultaneous distribution of the elements defined in this setup.

11
Lemma 6. Let Φ ∼ F2 (n, m) and L = [n]\[n − f ]. If Mk is the random variable denoting the
number of clauses in Φk := Gk (Φ, f ) for k ∈ {1, 2}, and M0 and M⋆ are the number of unsatisfied-
and satisfied clauses, respectively, then

(Mk )k∈K = (M0 , M1 , M2 , M⋆ ) ∼ Multinomial m, p(n, f ) ,

where p = (pk )k∈K and

f (f − 1) (n − f )f
p0 (n, f ) = , p1 (n, f ) = ,
4n(n − 1) n(n − 1)
(n − f )(n − f − 1) (n − f4 − 34 )f
p2 (n, f ) = , p⋆ (n, f ) = .
n(n − 1) n(n − 1)

Furthermore
Φk |(Mk )k∈K ∼ Fk (n − f, Mk ), (k ∈ {1, 2}),
and Φ1 and Φ2 are conditionally independent given (Mk )k∈K .

This lemma is again a direct consequence of the uniformity and the independence of the clauses
of a random 2-CNF formula.
The last lemma of this section gives a lower bound on the probability that a 1-CNF formula is
satisfiable.

Lemma 7. Let n, m ∈ N with n ≥ m and let Φ ∼ F1 (n, m). Then

m m
P(Φ ∈ SAT) ≥ 1 −


n .

This lemma can be proven in the same way as they prove Lemma 8 in [BOOS25]. Thus, we will
not repeat the argument here.

3.2 Decomposition of probability when many variables are fixed

Let Φ ∼ F2 (n, n), and L be a consistent set of literals with |L| = f = f (n), where f (n) = Ω(n1/3+ε )
for a small ε > 0. We will prove that limn→∞ P(ΦL ∈ SAT) = 0. In this section, the aim is to closely
regulate the unit-propagation procedure and hereby establish an upper bound on the probability
of interest. Later, it is established that this upper bound approaches zero as n → ∞.

Controlling the unit-propagation procedure

The assumption on f implies that f (n) ≥ nq , where q = 1/3 + ε for some small ε > 0. Let L′ ⊆ L
with |L′ | = ⌊nq ⌋. As {ΦL ∈ SAT} ⊆ {ΦL′ ∈ SAT} it is sufficient to establish that limn→∞ P(ΦL′ ∈
SAT) = 0. Thus, we will WLOG assume that f (n) = ⌊nq ⌋ for some 1/3 < q < 1/2.
Next, we define a sequence of random elements that resembles a controlled version of the unit-
propagation procedure. First, we define the initial elements of the procedure. Let G be the function
defined in Lemma 4. Then define
(0)
Ψ2 := G(Φ, L), S (−1) = 0, S̄ (−1) = 0, S (0) := f,
(0)
M1 = f + 1, L(0) := [n]\[n − f ], M(0) := {∅, Ω}.

12
 (0) (0)
Note that S (0) = S (−1) + (M1 − 1)+ , L(0) = [n − S (−1) ]\[n − S (0) ], and Lemma 4 states that Ψ2
is constructed such that
(0) (0)
Ψ2 ∼ F2 (n, n) = F2 n − S (−1) , (n − S̄ (−1) )+ and ΦL ∈ SAT = (Ψ2 )L(0) ∈ SAT . (3.1)

 

Furthermore, M(0) is the trivial σ-algebra and thus it provides no information. Now, additional
elements are constructed recursively. Let R := ⌊n1−2q log n⌋ denote the number of rounds. Then
for each r ∈ [R] we define the following recursively.
(r) (r−1)
Let G1 and G2 be the functions from Lemma 6 and define Φk := Gk (Ψ2 , |L(r−1) |) for
(r) (r) (r)
k ∈ {1, 2}. Also, let Mk denote the number of clauses in Φk for k ∈ {1, 2} and let M0 and
(r) (r−1)
M⋆ denote the number of unsatisfied- and satisfied clauses of (Ψ2 )L(r−1) , respectively. Define
(r)
the σ-algebra M := σ M
(r) (r−1) ∪ σ(Mk , k ∈ K) . The elements are constructed such that

(r−1) (r) (r) (r)

(Ψ2 )L(r−1) ∈ SAT = (Φ2 )L(Φ(r) ) ∈ SAT, Φ1 ∈ SAT, M0 =0 , (3.2)
 
1

see (2.6), and Lemma 6 states that



(r) (r−1)
(Mk )k∈K |M(r−1) ∼ Binomial (n − S̄ (r−2) )+ , p n − S (r−2) , (M1 − 1)+ , (3.3)
(r) (r)
Φk |M(r) ∼ F2 (n − S (r−1) , Mk ), k ∈ {1, 2}, (3.4)

(r) (r) (r) (r) (r)

and Φ2 and Φ1 are independent when conditioning on M(r) . Now, define Ψ̄2 := G(Φ2 , L(Φ1 )),
(r)
where G is the function from Lemma 4 and the set corresponding to Φ1 is defined in (2.4). As
(r) (r)
Φ2 and Φ1 are independent given M(r) , Fact 5 states that
(r) (r) (r) (r)
Ψ̄2 |M(r) ∼ F2 (n − S (r−1) , Mk ), and Ψ̄2 ⊥⊥ Φ1 |M(r) .
(r) (r)
Moreover, if M̄1 = |L(Φ1 )|, then
(r) (r) (r)
(Φ2 )L(Φ(r) ) ∈ SAT = (Ψ̄2 )L̄(r) ∈ SAT , where L̄(r) = [n−S (r−1) ]\[n−S (r−1) − M̄1 ]. (3.5)
 
1

(r)
Now, we either add clauses to Ψ̄2 or remove clauses. Define S̄ (r−1) = ⌊log n⌋ · S (r−1) and let
(r) (r) (r) (r) (r)
(Lj,1 , Lj,2 ) for j ∈ [M2 ] be the random literals of Ψ̄2 . If M2 < (n − S̄ (r−1) )+ define additional
(r) (r) (r)
random literals (Lj,1 , Lj,2 ) for j ∈ {M2 , . . . , (n − S̄ (r−1) )+ } where conditional on M(r) they are
i.i.d. and uniformly distributed on D(r) := {(ℓ1 , ℓ2 ) ∈ (±[n − S (r−1) ])2 : |ℓ1 | < |ℓ2 |}. Define
(r) (r) (r)
(r)
Ψ2 = then Ψ2 |M(r) ∼ F2 n − S (r−1) , (n − S̄ (r−1) )+ .
^

Lj,1 ∨ Lj,2 ,
j∈[(n−S̄ (r−1) )+ ]

Lastly, let
(r)
S (r) = S (r−1) + (M1 − 1)+ , and L(r) = [n − S (r−1) ]\[n − S (r) ].
(r)
Then we are in the same setting again and we can repeat the procedure on Ψ2 and L(r) under the
conditional distribution given M(r) , where we note that L(r) is deterministic given M(r) .
Note that it is mainly the sequence {S (r) }r∈[R] that controls the size of the different elements
(r)
constructed above and this sequence is defined from the sequence {M1 }r∈[R] . Thus, a big part

13
of the proof in the over-constrained setting is controlling the size of this sequence, which describes
the number of unit-clauses constructed in each round. We show that that this number remains on
the order of nq (remember that f (n) = ⌊nq ⌋) throughout the R rounds as the below lemma states.

Lemma 8. There exist constants c0 > 0 and C0 > 0 such that the two events
(r) (r)
Bl = M1 ≥ c0 nq , r ∈ [R] B u = M1 ≤ C0 nq , r ∈ [R]
 
and

satisfy
lim P Bl = lim P Bu = 1.



n→∞ n→∞

(r)
As S (r) ≤ ⌊nq ⌋ + the above Lemma also implies that:
PR
r=1 M1

Fact 9. There exists a constant C1 > 0 such that for r ∈ [R] (and n large enough) we have

S ≤ C1 n1−q log n ⊆ Bu .
 (r)

Lemma 8 is technical to prove. It is easy to find constants c0 > 0 and C0 > 0 such that for
(r)
each r ∈ [R] we have that c0 nq < M1 < C0 nq w.h.p. This does however not imply that the entire
(r)
sequence {M1 }r∈[R] is uniformly bounded w.h.p. A union bound is not tight enough to establish
the uniform boundedness so the dependence structure of the sequence needs to be exploited. We
(R)
establish that when M1 is bounded w.h.p. the previous elements will be bounded w.h.p. as well.

Decomposing the probability

The random elements defined in the controlled unit-propagation procedure above will now be related
to the probability that ΦL is satisfiable. Our aim is to show that the probability tends to zero and
thus we want to construct an upper bound on the probability. Let Bu and Bl be the events from
Lemma 8. Using equation (3.1) we first note that

(0)
P ΦL ∈ SAT, Bu , Bl = P (Ψ2 )L(0) ∈ SAT, Bu , Bl .

Next, using (3.2) and (3.5) on the term at the right gives that
 
(0)
P (Ψ2 )L(0) ∈ SAT, Bu , Bl
 
(1) (1) (1)
=P (Φ2 )L(Φ(1) ) ∈ SAT, Φ1 ∈ SAT, M0 = 0, Bu , Bl
1

(3.6)
 
(1) (1) (1)
=P (Ψ̄2 )L̄(1) ∈ SAT, Φ1 ∈ SAT, M0 = 0, Bu , Bl
(1)
(1) (1) (1) (1)
≤P Ψ̄2 = 0, Bu , Bl , M1 ≤ M̄1 + 1, M2 ≥ (n − S̄ (−1) )+


L̄(1)
∈ SAT, M0
(1) (1) (1) (1)
+P M̄1 + 1, Φ1 ∈ SAT, Bu + P M2 < (n − S̄ (−1) )+ , Bu , Bl .



M1 >

The first term in the last expression above will now be further decomposed. Note that when
(1) (1) (1) (1)
M1 ≤ M̄1 + 1 then L(1) ⊆ L̄(1) and when M2 ≥ (n − S̄ (−1) )+ then Ψ2 is a sub-formula of
(1)
Ψ̄2 . Thus
(1)
(1) (1) (1) (1)
P Ψ̄2 = 0, Bu , Bl , M1 ≤ M̄1 + 1, M2 ≥ (n − S̄ (−1) )+


L̄(1)
∈ SAT, M0
(1) (1)
(3.7)
≤P (Ψ2 )L(1) ∈ SAT, M0 = 0, Bu , Bl .

14
Now, recursively repeating (3.6) and (3.7) R times in total we eventually arrive at the decomposition
(r)
P (ΦL ∈ SAT, Bu , Bl ≤P M0 = 0, r ∈ [R], Bu , Bl

R
(r) (r) (1)
+ > M̄1 + 1, Φ1 ∈ SAT, Bu
X

P M1
r=1
(3.8)
R
(r)
+ < (n − S̄ (r−1) )+ , Bu , Bl .
X

P M2
r=1

The below lemma establishes the limits of the above upper bound.

Lemma 10. It holds that

(r)
(1) limn→∞ P M0 = 0, r ∈ [R], Bu , Bl = 0,

(r) (r) (1)

(2) limn→∞ ≥ M̄1 + 2, Φ1 ∈ SAT, Bu = 0,
PR

r=1 P M1
(r)
(3) limn→∞ < (n − S̄ (r−1) )+ , Bu , Bl = 0.
PR

r=1 P M2

When proving the above Lemma the events Bu and Bl make it possible to control the sizes of
the different random elements. Further, Lemma 8 makes it possible to evaluate one event at a time
by conditioning on previous information and hereby knowing exact distributions.
The above decomposition and lemmas make it straight forward to prove that ΦL is indeed
asymptotically unsatisfiable.

Proof of Definition 1 (2). Lemma 8 implies that it is sufficient to establish that the right-hand side
of (3.8) tends to zero as n → ∞. But this is a direct consequence of Lemma 10.

3.3 Decomposition of probability when few variables are fixed

We will now also control the unit-propagation procedure when the problem is asymptotically sat-
isfiable, where we instead need a lower bound. Let Φ ∼ F2 (n, n) and L ⊆ ±[n] be consistent with
|L| = f (n), where f (n) ≤ n1/3−ε for a small ε > 0. In section 3.2, we saw that the number of unit-
clauses remained of order nq throughout the R rounds. In this section, we instead want to show
that the unit-propagation procedure terminates and thus that the number of unit-clauses reaches
zero within the number of rounds we consider. It turns out that the number of unit-clauses gener-
ated by this algorithm will be a super-martingale (on a set of probability one) when considering the
sequence from round r = 2 and onward. This is helpful as we will make use of optional sampling.
Therefore, we will start by stating another lemma for which the entire sequence of one-clauses is a
super-martingale and then we will connect this lemma to our main theorem.
(−1) (0)
Lemma 11. Let 0 < q < 1/3 and let M1 and M1 be random variables taking values in [n]
(0)
satisfying that E[M1 ] ≤ C0 n for some C0 > 0 and also that
q

(−1) (0)
lim P M1 ≤ nq log n = lim P M1 ≤ nq log n = 1.



n→∞ n→∞

(−1) (−1) (0) (−1) (0)

Define L′ = [n − M1 ]\[n − M1 − M1 ] and M(0) = σ(M1 , M1 ) and let Φ′ be a random
function with
(−1) (−1) (0)

Φ′ |M(0) ∼ F2 n − M1 , n − M1 − M1 .

15
If Φ ∼ F2 (n, n) then
lim inf P Φ′L′ ∈ SAT ≥ lim inf P Φ ∈ SAT .



n→∞ n→∞

Controlling the unit-propagation procedure

The notation used when naming the elements of the unit-propagation procedure in subsection 3.2
is now reused in this section. As there are small differences in the definitions in the two cases it is
important to pay attention to which definitions apply to which lemmas.
(−2) (−1)
Let Φ′ , L′ , M1 and M1 be the elements of Lemma 11. Again we start by defining some
initial elements of our unit-propagation procedure:
(0) (−1) (0) (−2) (−1)

Ψ2 := Φ′ , S (−1) = M1 , S (0) = S (−1) + M1 , L(0) := L′ , M(0) = σ M1 , M1 .

(0)
Note that L(0) = [n − S (−1) ]\[n − S (0) ] and Ψ2 |M(0) ∼ F2 (n − S (−1) , n − S (0) ). Now the rest
of the elements are generated recursively. Let R = ⌊n1−2q log−3 n⌋ denote the number of rounds.
Then for each r ∈ [R] we define the following recursively. Let G1 and G2 be the functions defined in
(r) (r−1) (r)
Lemma 4 and let Φk = Gk (Ψ2 , L(r−1) ) for k ∈ {1, 2}. Also, let Mk be the number of clauses
(r) (r) (r)
in Φk for k ∈ {1, 2} and let M0 and M⋆ denote the number of unsatisfied- and satisfied clauses
(r−1) (r)
of (Ψk )L(r−1) , respectively. We further define M(r) = σ M(r−1) ∪ σ(Mk , k ∈ K) . We have

that  (r−1)  (r) (r) (r)

(Ψ2 )L(r−1) ∈ SAT = (Φ2 )Φ(r) ∈ SAT, Φ1 ∈ SAT, M0 = 0 , (3.9)
1

see (2.6), and Lemma 6 states that

(r) (r−1)
(Mk )k∈K |M(r−1) ∼ Binomial n − S (r−1) , p(n − S (r−2) , M1 ) , (3.10)

(r) (r) (r) (r)

Φk |M(r) ∼ F2 (n − S (r−1) , Mk ), k ∈ {1, 2}, Φ1 ⊥⊥ Φ2 |M(r) . (3.11)

(r) (r) (r) (r)

Now, define Ψ̄2 := G(Φ2 , Φ1 ), where G is the function from Lemma 4 and Φ1 is seen as set,
(r) (r)
see (2.4). As Φ2 and Φ1 are independent given M(r) , Fact 5 states that
(r) (r) (r) (r)
Ψ̄2 |M(r) ∼ F2 (n − S (r−1) , Mk ), and Ψ2 ⊥⊥ Φ1 |M(r) . (3.12)
(r) (r)
Moreover, if M̄1 denotes the number of distinct variables appearing in Φ1 , then
(r) (r) (r)
(Φ2 )Φ(r) ∈ SAT = (Ψ̄2 )L̄(r) ∈ SAT , where L̄(r) = [n − S (r−1) ]\[n − S (r−1) − M̄1 ]. (3.13)
 
1

(r) (r)
Now, we add additional clauses to Ψ̄2 or remove clauses. Let S (r) = S (r−1) + M1 . Recall that
(r) (r−1) (r) (r) (r)
(r) (r) (r)

= M2 − M0 + M 1 + M⋆ = n − S (r−1) − M0 + M1 + M⋆ ≤ n − S (r) .


M2
(r) (r) (r) (r)
Let (Lj,1 , Lj,2 ), j ∈ [M2 ] be the random random literals of Ψ̄2 and define additional random
(r) (r) (r)
literals (Lj,1 , Lj,2 ) for j ∈ {M2 + 1, . . . , n − that when conditioning on M(r) are i.i.d. and
S (r) }
uniformly distributed on D(r) := (ℓ1 , ℓ2 ) ∈ (±[n − S (r−1) ])2 : |l1 | < |l2 | . Define


(r) (r) (r)
(r)

Ψ2 = then Ψ2 |M(r) ∼ F2 n − S (r−1) , n − S (r) .
^

Lj,1 ∨ Lj,2 ,
j∈[n−S (r) ]

16
 (r)
Lastly, define L(r) = [n − S (r−1) ]\[n − S (r) ]. Now, the same procedure can be applied to Ψ2 and
L(r) in the conditional distribution given M(r) , where we note that L(r) is deterministic given M(r) .

Decomposing the probability

We will use the elements defined previously to create a lower bound on the probability of Φ′L′ being
satisfiable. The definitions in the initial round imply that
 ′  (0)
ΦL′ ∈ SAT = (Ψ2 )L(0) ∈ SAT .

Now, using (3.9) we get

(0) (1) (1) (1)
(Ψ2 )L(0) ∈ SAT = (Φ2 )Φ(1) ∈ SAT, Φ1 ∈ SAT, M0 =0 , (3.14)
 
1

and equation (3.13) further implies

(1) (1)
(Φ2 )L(1) ∈ SAT = (Ψ̄2 )L̄(1) ∈ SAT . (3.15)
 

(1) (1) (1) (1)

As M̄1 ≤ M1 we have that L(1) ⊆ L̄(1) and also Ψ2 is constructed such that Ψ̄2 is its
sub-formula. Thus, we get the inclusions
(1) (1)
(Ψ̄2 )L̄(1) ∈ SAT ⊇ (Ψ2 )L(1) ∈ SAT . (3.16)
 

Combining all of the above set inclusions imply that

(1) (1) (1)
P (Φ′ )L′ ∈ SAT ≥ P (Ψ2 )L(1) ∈ SAT, Φ1 ∈ SAT, M0 =0 .

(1)
Now, we are back at considering the event {(Ψ2 )L(1) ∈ SAT} and thus (3.14), (3.15) and (3.16)
can be repeated for r = 2, . . . , R. Hereby, we eventually get the lower bound
(R) (r) (1)
P Φ′L′ ∈ SAT ≥ P (Ψ2 )L(R) ∈ SAT, Φ1 ∈ SAT, M0 ∈ SAT, r ∈ [R] . (3.17)

Our next lemma gives that the above lower-bound tends to one as n → ∞.
Lemma 12. We have that
(r)
(1) limn→∞ P M1 ≤ nq log n, r ∈ [R] = 1,

(r) (r)
(2) limn→∞ P Φ1 ∈ SAT, r ∈ [R] M1 ≤ nq log n, r ∈ {−1, . . . , R} = 1,

(r) (r)
(3) limn→∞ P M0 = 0, r ∈ [R] M1 ≤ nq log n, r ∈ {−1, . . . , R} = 1,

(R)
(4) limn→∞ P M1 = 0 = 1.

(r)
That the sequence (M1 )r∈[R] is bounded from above follows using optional sampling where
we exploit that the sequence turns out to be a super-martingale. Lemma 12 (2) and (3) are then
consequences of Lemma 6. Lastly (4) is proven by a Poisson approximation and also using theory
of Galton-Watson trees. Lemma 11 is now an easy consequence of Lemma 12.
(−1) (0)
Proof of Lemma 11. The definitions of M1 and M1 along with Lemma 12 (1) imply that the
event that we condition on in (2) and (3) of Lemma 12 happens w.h.p. Therefore, Lemma 12 (2)

17
implies that
(r)
lim P(Φ1 ∈ SAT, r ∈ [R]) = 1,
n→∞

and Lemma 12 (3) implies that

(r)
lim P(M0 = 0, r ∈ [R]) = 1.
n→∞

(R)
Also, Lemma 12 (4) implies that L(R) = ∅ w.h.p. and when this is the case also Ψ2 |M(R) ∼
(r)
F2 (n − S (R−1) , n − S (R−1) ). Moreover, that M1 ≤ nq log n for all r ∈ [R] w.h.p. implies that
S (R−1) ≤ n1−q w.h.p. These observations along with Fatou’s Lemma give
(R)
lim inf P (Ψ2 )L(R) ∈ SAT


n→∞
(R)
= lim inf E P Ψ2 ∈ SAT M(R) S (R−1) ≤ n1−q , L(R) = ∅


n→∞
h i
(R)
≥E lim inf P Ψ2 ∈ SAT M(R) S (R−1) ≤ n1−q , L(R) = ∅


n→∞
= lim inf P Φ ∈ SAT .


n→∞

Combining these limits with the decomposition in (3.17) gives the result.

4 Proofs
In this section we provide proofs of the lemmas stated previously. Sections 4.1 and 4.2 are devoted
to the case q > 1/3 and sections 4.3 and 4.4 concern the case q < 1/3. Lastly, in section 4.5 the
technical lemmas of section 3.1 are proven.

4.1 Proof of Lemma 8

In this section, we again consider the elements defined in the unit-propagation procedure of section
(r) (r)
3.2. We establish that the two events Bu = {M1 ≤ C0 nq } and Bl = {M1 ≥ c0 nq } happen w.h.p.
A problem we encounter is that we cannot control the size of the sequence {S (r) }r∈[R] . Thus, we
will need to define a new sequence of random elements that approximates our previously defined
(0)
elements but for which we do not have this problem. Let T (−1) = 0, N1 = ⌊nq ⌋ and define
recursively for each r ∈ [0, R]
(r)
T (r) = min T (r−1) + (N1 − 1)+ , ⌈n1−q log2 n⌉ , T̄ (r) = ⌊log n⌋ · T (r) ,


+
+

(r+1) (1) (r) (r−1)
N1 |N1 , . . . , N1 ∼ Binomial n − T̄ (r−2) , p1 n − S (r−1) , M1 −1 .

Now, the sequence {T (r) }r∈[R] is upper-bounded by ⌈n1−q log2 n⌉ but at the same time it turns out
that it has the same distribution as {S (r) }r∈[R] w.h.p. Let c0 > 0 and C0 > 0 be two constants
(which will be further specified later) and define the events
 (r)  (r)
Dl = N1 ≥ c0 nq ∀r ∈ [R] , Du = N1 ≤ C0 nq ∀r ∈ [R]
AS = S (R) < n1−q log2 n AT = T (R) < n1−q log2 n .
 

18
Equation (3.3) implies that
(r) D  (r)
, 1AS =



M 1 1AS r∈[R]
N1 1AT r∈[R]
, 1AT .

(r) (r) (r)

Moreover, on AS it holds that {M1 1AS }r∈[R] = {M1 }r∈[R] and on AT we have {N1 1AT }r∈[R] =
(r)
{N1 }r∈[R] . Thus, if Bl and Bu are the events of Lemma 8, then

P(Blc ) ≤ P(Blc ∩ AS ) + P(AcS ) = P(Dlc ∩ AT ) + P(AcS ) ≤ P(Dlc ) + P(AcS ),

and similarly P(Buc ) ≤ P(Duc ) + P(AcS ). Thus, in order to establish Lemma 8 it is sufficient to
establish that limn→∞ P(AS ) = 1 and that limn→∞ P(Dl ) = limn→∞ P(Du ) = 1. Thus, proving
Lemma 8 reduces to proving the below two lemmas

Lemma 13. We have limn→∞ P(AS ) = 1.

Lemma 14. We have limn→∞ P(Dl ) = limn→∞ P(Du ) = 1.

We start by proving the first of the above two lemmas.

Proof of Lemma 13. We will establish this using Markov’s inequality. Note that
R R
(r) (r)
E[S (R) ] = ⌊nq ⌋ + E[(M1 − 1)+ ] ≤ E[M1 ],
X X

r=1 r=0

and equation (3.3) and the definition of p1 in Lemma 6 implies

(r−1) (r−1)
n − S (r−2) − (M1 − 1)+ (M1 − 1)+


(r) (r−2)
+
E[M1 ] = n − S̄
n − S (r−2) n − S (r−2) − 1

(r−1)
(n − ⌊log n⌋S (r−2) )+ n − S (r−2) − (M1 − 1)+
 
(r−1)  (r−1) 
≤ E M1 · · ≤ E M1 .
n − S (r−2) − 1 n − S (r−2)

In the above, we used that S (r−2) ≥ nq − 1 so n − g(n)S (r−2) ≤ n − S (r−2) − 1. Repeating the above
argument we eventually get that
(r) (0)
E[M1 ] ≤ E[M1 ] ≤ nq .

Thus, Markov’s inequality implies that

E S (R) (R + 1)nq (n1−2q log n + 1)nq

 
P(AcS ) =P S (R) 1−q
log n ≤ 1−q
2
→ 0 as n → ∞,


≥n ≤ ≤
n log2 n n1−q log2 n n1−q log2 n

where we use that for q < 1/2 we have 1 − q > q.

To prove the next Lemma we need the below technical lemma.

Lemma 15. Let r, s ∈ [0, R] with s < r. Then

(r) (1) (s) (s)
(1) E[N1 |N1 , . . . , N1 ] ≤ N1 ,
(r) (1) (s) (s) (s)
(2) E[(N1 )2 |N1 , . . . , N1 ] ≤ RN1 + (N1 )2 ,

19
 (s)
(3) Assume N1 ≤ C0 nq for some C0 > 0. Then there exists C1 > 0 (dependent on C0 but
(r) (1) (s) (s)
independent of r and s) such that E[N1 |N1 , . . . , N1 ] ≥ N1 − C1 n1−2q log4 n
(s)
(4) Assume c0 nq ≤ N1 ≤ C0 nq for some c0 , C0 > 0. Then there exists C2 > 0 (dependent on c0
 (r) (1) (s) 
and C0 but independent of r and s) such that V N1 N1 , . . . , N1 ≤ C2 n1−q log4 n.

Proof. The inequalities will be established one at a time.

(1) Direct calculations give that
 (r) (1) (s) 
E N1 N1 , . . . , N1
  (r) (1) (r−1)  (1) (s) 
=E E N1 N1 , . . . , N 1 N1 , . . . , N1
(r−1) (r−1)
n − T (r−2) − (N1 − 1)+ (N1 − 1)+ (1)


(s)
=E (n − T̄ (r−2)
) N1 , . . . , N1
n − T (r−2) n − T (r−2) − 1

(r−1)
(r−1) (n − ⌊log n⌋T (r−2) )+ n − T (r−2) − (N1 − 1)+


(1) (s)
≤E N1 · N1 , . . . , N 1
n − T (r−2) − 1 n − T (r−2)
 (r−1) (1) (s)  (s)
≤E N1 N1 , . . . , N1 ≤ · · · ≤ N1 ,

where we in the first inequality use that T (r−2) ≥ nq − 1.

(2) For the second moment, we use that when X ∼ Binomial(n, p), then

2
E[X 2 ] = np(1 − p) + n2 p2 ≤ E[X] + E[X] . (4.1)

This along with the calculations and result of (1) imply

(r) (1) (s)  (r) (1) (r−1)  (1) (s) 
E (N1 )2 N1 , . . . , N1 =E E (N1 )2 N1 , . . . , N1
  
N1 , . . . , N1
 (r−1) (1) (s)  (r−1) 2 (1) (s) 
+ E (N1 ) N1 , . . . , N1

≤E N1 N1 , . . . , N1
(s) (s)
≤ · · · ≤ (r − s)N1 + (N1 )2
(s) (s)
≤RN1 + (N1 )2 .

(3) Next, we want to find a lower bound on the mean. Here we use that T (r−2) ≤ n1−q log2 n + 1

20
and we also make use of (1) and (2).
 (r) (1) (s) 
E N1 N1 , . . . , N1
(r−1) (r−1)
n − T (r−2) − (M1 − 1)+ (N1 − 1)+ (1)


(s)
=E (n − T̄ (r−2)
) N1 , . . . , N1
(n − T (r−2) )(n − T (r−2) − 1)
(r−1)
(r−1)

n − T (r−2) + N1 N1 −1
 
(1) (s)
≥E n − n 1−q
log n − log n
3
· N1 , . . . , N 1
n − T (r−2) n
(r−1)
n − n1−q log3 n − log n N1
   
(r−1) (1) (s)
≥ · E (N1 − 1) 1 − N1 , . . . , N1
n n−n 1−q log n − 1
2 (4.2)
(r−1) (1) (s)  
n − n1−q log3 n − log n  (r−1) (1) E (N1 )2 N1 , . . . , N1
 
(s) 
≥ E N1 N1 , . . . , N1 − 1 −
n n − n1−q log2 n − 1
(s) (s)
n − n1−q log3 n − log n  (r−1) (1) (s)  RN1 + (N1 )2
≥ · E N1 N1 , . . . , N1 − 1 −
n n − n1−q log2 n − 1
!r−s (s) (s)
n − n1−q log3 n − log n RN1 + (N1 )2
 
(s)
≥··· ≥ N1 − (r − s) 1 + .
n n − n1−q log2 n − 1

We will now bound the above two terms one at a time. For the first term, we will need the below
inequality which is true for x ≥ 2 and y > 0:
h
iy
2 x
h

iy h

iy
1− = exp x log 1 − 2
≥ exp x − 4
= exp − 4y ≥ 1 − 4y.


x x x

This and that r − s ≤ R ≤ n1−2q log n now implies

!r−s nq log−3 n !n1−3q log4 n
n − n1−q log3 n − log n 2

≥ 1− q ≥ 1 − 4n1−3q log4 n. (4.3)
n n log−3 n

(s)
For the other term, we use the assumption that N1 ≤ C0 nq . Then as 2q < 1 we get
(s) (s)
RN1 + (N1 )2 C0 n1−q log n + C02 n2q
≤ →0 as n → ∞.
n − n1−q log2 n − 1 n − n1−q log2 n − 1
Thus, for a C > 1 we have that
(s) (s)
RN1 + (N1 )2
!
(r − s) 1 + ≤ (r − s)C ≤ Cn1−2q log n. (4.4)
n − n1−q log3 n − 1

(s)
Now, combining (4.2), (4.3) and (4.4) along with the assumption that N1 ≤ C0 nq we get that
 (r) (1) (s) 
(s)
E N1 N1 , . . . , N1 ≥ 1 − 4n1−3q log4 n N1 − Cn1−2q log n
(s) (s)
≥N1 − 2n1−3q log4 nC0 nq − Cn1−2q log n ≥ N1 − C1 n1−2q log4 n.

(s)
(4) Lastly, we combine (2) and (3) along with the extra assumption that N1 ≥ c0 nq (which implies

21
 (s)
that N1 − C1 n1−2q log4 n ≥ 0) to conclude that
(r) (1) (s) (r) (1) (s)   (r) (1) (s) 
2
V(N1 |N1 , . . . , N1 ) = E (N1 )2 N1 , . . . , N1

− E N1 N1 , . . . , N1
(s) (s) (s)
2
≤ RN1 + (N1 )2 − N1 − C1 n1−2q log4 n
(s) (s) (s)
≤ RN1 + (N1 )2 − (N1 )2 + 2C0 nq C1 n1−2q log4 n
≤ C2 n1−q log4 n.

(0)
Fact 16. As N1 = ⌊nq ⌋ the above Lemma implies the existence of constants c1 > 0 and C1 > 0
such that
(R) (R) (R)
E[N1 ] ≤ nq + 1, E[N1 ] ≥ c1 nq , V[N1 ] ≤ C1 n1−q log4 n.
We are now ready to prove the last lemma of this section which will imply Lemma 8.

Proof of Lemma 14. Let c1 > 0 and C1 > 0 be the constants of Fact 16. We let c0 = c1 /2 and
C0 = 2C1 be the constants of our lemma. Note that when q > 1/3 then (1 − q)/2 < q why we can
1−q
choose a q1 ∈ 2 , q . Then using Chebyshev’s inequality and Fact 16 we get

 (R) 
(R) (R) V N1 C1 n1−q log4 n
P N1 − E[N1 ] ≥n q1

≤ ≤ →0 as n → ∞.
n2q1 n2q1
Fact 16 then implies
(R)
≤ c1 nq − nq1 → 0 as n → ∞, (4.5)


P N1
(R)
P N1 ≥ C1 n q + n q1

→0 as n → ∞. (4.6)

The above implies that the sequence is still of order nq at step R. We will use this to show that
the sequence cannot have been too small or too large in previous steps. Remember that we want
to establish that limn→∞ P(Dl ) = 1, where the complimentary event is given by
(r) (r)
Dlc = ∃r ∈ [R] s.t. N1 < c0 nq = ∃r ∈ [R] s.t. N1 < 12 c1 nq .
 

Using (4.5) we see that the above is implied if we show that

(R)
lim inf P N1 ≤ c1 nq − nq1 Dlc > 0. (4.7)


n→∞

Define
(r)  (1) (2) (r−1) (r)
Dl = N1 > 12 c1 nq , N1 > 12 c1 nq , . . . , N1 > 21 c1 nq , N1 ≤ 21 c1 nq , (r ∈ [R]).
(r)
Then the above events are disjoint and Dlc = ∪r∈[R] Dl . Using Markov’s inequality we get

(R) (r)
(R) (r) (R) (r) E[N1 |Dl ]
P(N1 ≤ c1 n − n q q1
|Dl ) = 1 − P(N1 > c1 n − n q q1
|Dl ) ≥ 1− , (r ∈ [R]). (4.8)
c1 nq − nq1
Next, using Lemma 15 (1) we see
(R) (r) (R) (1) (r) (r)  (r) (r)
E[N1 |Dl ] = E E[N1 |N1 , . . . , N1 ] Dl ≤ E[N1 |Dl ] ≤ 21 c1 nq , (r ∈ [R]).


22
This upper bound is then inserted in (4.8):

(R) (r)
1
2 c1 n
q
1
P(N1 ≤ c1 nq − nq1 |Dl ) ≥ 1 − ≥ , (r ∈ [R]).
c1 nq − nq1 4
This finally implies that
 (R)
≤ c1 nq − nq1 ∩ Dlc


(R) P N1
q q1
Dlc =


P N1 ≤ c1 n − n
P(Dlc )
(R) (r)
(r)
P(Dl )
PR
r=1 P N1 ≤ c1 nq − nq1 Dl
=
P(Dlc )
(r)
1 r=1 P(Dl ) 1
PR
≥ = ,
4 P(Dlc ) 4

which is (4.7).
Next, we will establish that limn→∞ P(Du ) = 1, where the complimentary event is given by
(r) (r)
Duc = ∃r ∈ [R] s.t. N1 > C0 nq = ∃r ∈ [R] s.t. N1 > 2C1 nq .
 

Using (4.6) we get that this is implied if we can show that

(R)
lim inf P N1 ≥ C1 nq + nq1 Duc > 0. (4.9)


n→∞

Define
 (1) (2) (r−1) (r)
Du(r) = N1 < ⌊2C1 nq ⌋, N1 < ⌊2C1 nq ⌋, . . . , N1 < ⌊2C1 nq ⌋, N1 ≥ ⌊2C1 nq ⌋ , (r ∈ [R]).

(r)
Note that the above events are disjoint and Duc ⊆ ∪r∈[R] Du . Let r ∈ [R] be fixed. The event
(r) (r)
Du does not give us an upper bound on N1 which implies that we do not have good bounds on
(R) (1) (r)
E[N1 |N1 , . . . , N1 ]. Therefore, we split the below probability into two terms. Write
(R)
< C1 nq + nq1 Du(r)


P N1
=P N1
(R)
< C1 nq + nq1 Du(r) ∩ {N1
(r)
< 2⌊2C1 nq ⌋} P N1
(r)
< 2⌊2C1 nq ⌋ Du(r) (4.10)

(R) (r) (r)

+P < C1 nq + nq1 Du(r) ∩ ≥ 2⌊2C1 nq ⌋} P ≥ 2⌊2C1 nq ⌋ Du (r)



N1 {N1 N1 .

We will now consider the above two terms separately. For the first term, we now have a bound on
(r) (r)
N1 , as we condition on the event ⌊2C1 nq ⌋ ≤ N1 < 2⌊2C1 nq ⌋. However, we can not use Markov’s
inequality as before as our inequality points in the wrong direction. Thus, we will instead use
Chebyshev’s inequality. In Lemma 15 the bounds (3) and (4) imply that there exists a constant
C > 0 (which is independent of r) such that
(R) (1) (r) (R) (1) (r)
E[N1 |N1 , . . . , N1 ] ≥ 2C1 nq − Cn1−2q log4 n and V(N1 |N1 , . . . , N1 ) ≤ Cn1−q log4 n.

Then
(R) (1) (r)
E[N1 |N1 , . . . , N1 ] − (C1 nq + nq1 ) ≥ C1 nq − Cn1−2q log4 n − nq1 > 0,

23
and we can use Chebyshev’s inequality to establish that
(R) (1) (r)

P N1 < C1 nq + nq1 N1 , . . . , N1
(R) (R) (1) (r) (1) (r)

≤P N1 − E[N1 |N1 , . . . , N1 ] > C1 nq − Cn1−2q log4 n − nq1 N1 , . . . , N1
(R) (1) (r)
V[N1 |N1 , . . . , N1 ] (4.11)
≤
(C1 nq − Cn1−2q log4 n − nq1 )2
Cn1−q log4 n 1
≤ ≤ .
(C1 nq − Cn1−2q log4 n − nq1 )2 4

Lastly, we used that the fraction is of order n1−3q log4 n and thus it approaches zero as n → ∞.
(r) (r)
For the second term of (4.10) we want to show that P(N1 ≥ 2⌊2C1 nq ⌋|Du ) is small. Note that
(r) (r−1) (r)
Du contains the event N1 < ⌊2C1 nq ⌋ which makes it unlikely that N1 ≥ 2⌊2C1 nq ⌋. When
(r−1)
N1 < ⌊2C1 nq ⌋ we have:
(r) (1) (r−1) (r)
≥ 2⌊2C1 nq ⌋ N1 , . . . , N1 ≥ ⌊2C1 nq ⌋


P N1 , N1
n
(r) (1) (r−1) (r)
= = x N1 , . . . , N1
X
≥ ⌊2C1 nq ⌋


P N1 , N1
x=2⌊2C1 nq ⌋
(r) (1) (r−1)

n
P N1 = x N1 , . . . , N1 (4.12)
=
X
(r) (1) (r−1)

x=2⌊2C1 nq ⌋ P N1 ≥ ⌊2C1 nq ⌋ N1 , . . . , N1
(r) (1) (r−1)

n
X P N1 = 2⌊2C1 nq ⌋ N1 , . . . , N1
≤ (r) (1) (r−1)

.
x=2⌊2C1 nq ⌋ P N1 = ⌊2C1 nq ⌋ N1 , . . . , N1

In the last inequality, we made the nominator larger and the denominator smaller. Here we used
(r) (1) (r−1) (r−1)
that N1 |N1 , . . . , N1 has a Binomial distribution and as N1 < ⌊2C1 nq ⌋ the mode of its
probability mass function is smaller than 2⌊2C1 nq ⌋. Now, if X ∼ Binomial(n, p) then

2y p (1 − p)
N
2y n−2y
P X = 2y y! (n − y)!



= n y = (p)y (1 − p)−y
P X=y k p (1 − p) (2y)! (n − 2y)!


n−y

y
e · y · ye
(np)y exp − y log(1 − p)


≤ 2y
2y
e· e
y  4 −y  np y
4 y y (np) exp yp = y exp(yp).
y

≤

e y
e y

n
n
In the above we have used that log(1 − p) ≥ −p and that e ne ≤ n! ≤ en ne . Furthermore,
coupling this with (4.12) we note that in our setup we have that the number of trials equals
n − ⌊log n⌋T r−2 and the probability parameter equals
(r−1)
(r−1)
N1
p1 n − T (r−2) , N1 ≤ ≤ Cnq−1 , (4.13)
n − T (r−2)
for a constant C > 0 chosen large enough. This implies that the factor in the middle becomes less

24
than one, so in (4.12) we get
(r) (1) (r−1) (r)
> 2⌊2C1 nq ⌋ N1 , . . . , N1 ≥ ⌊2C1 nq ⌋


P N1 , N1
(r) (1) (r−1)

n
X P N1 = 2⌊2C1 nq ⌋ N1 , . . . , N1
≤ (r) (1) (r−1)

x=2⌊2C1 nq ⌋ P N1 = ⌊2C1 nq ⌋ N1 , . . . , N1
 −2C1 nq
4   1
≤n2C1 n q
exp 2C1 Cn2q−1 ≤ .
e 4
q
Lastly, we used that n1+q (4/e)−2C1 n → 0 and exp(2C1 Cn2q−1 ) → 1 when n → ∞. Using this and
(4.11) in (4.10) we get
(R)
≤ C1 nq + np A(r)


P N1 u
(R) (1) (r)
(r)
≤ C1 nq + np N1 , . . . , N1 A(r) ≤ 2⌊2C1 nq ⌋}


≤E P N1 u ∩ {N1
(r) (1) (r−1) (r) 1 1 1
+ E P N1 > 2⌊2C1 nq ⌋ N1 , . . . , N1 ≥ ⌊2C1 nq ⌋ Au(r) ≤ + = .


, N1
4 4 2
This finally implies
 (R)
> C1 nq + np ∩ Au


(R) P N1
> C1 n + n Au =
q p

P N1
P(Au )
(R) (r)
(r)
> C1 nq + np Au P(Au )
PR
r=1 P N1
= (r)
P(Au )
(r)
1 r=1 P(Au ) 1
PR
≥ = ,
2 P(Au ) 2

and this implies (4.9) and finishes the proof.

4.2 Proof of Lemma 10

The three limits of this lemma are established one at a time. Remember that we use the defined
elements of Section 3.2, i.e. the unit-propagation procedure elements constructed for the regime
with q > 1/3.
We will first establish that
(r)
lim P M0 = 0, r ∈ [R], Bu , Bl = 0.


n→∞

Proof of Lemma 10 (1). Let c0 > 0 be the constant of Lemma 8 and C1 > 0 be the constant of
Fact 9. Remember that S̄ (r) = ⌊log n⌋S (r) for r ∈ [R]. As n(4⌊log n⌋)−1 > C1 n1−q log n, we note
that it is sufficient to establish that
(r) (r)
lim P M0 = 0, M1 ≥ c0 nq , r ∈ [R], S̄ (R−2) < n
= 0.


n→∞ 4

Recall that for r ∈ [R] equation (3.3) gives that

 
(r) (r−1)
M0 |M(r−1) ∼ Binomial (n − S̄ (r−2) )+ , p0 n − S (r−2) , (M1 − 1)+


,

25
and the function p0 , defined in Lemma 6, is given by
(r−1) (r−1)
(M1 − 1)+ (M1 − 1)+ − 1


(r−1)

p0 n − S (r−2)
, M1 = .
4(n − S (r−2) )(n − S (r−2) − 1)

Define i.i.d. random variables

n c0 nq − 2 2
j k  
X (1) , . . . , X (R) ∼ Binomial , .
2 2n

The above considerations imply that

(r) (r)
= 0, M1 ≥ c0 nq , r ∈ [R], S̄ (R−2) < n


P M0 4
(r) (r)
= 0, r ∈ [R], M1 ≥ c0 nq , r ∈ [R − 1], S̄ (R−2) < n


≤P M0 4
h i
(R)
=E P M0 = 0 M(r−1) 1{M (r) =0, M (r) ≥c


0 1
q
0n , r∈[R−1], S̄ (R−2) < n
4
}
h i
≤E P X (R) = 0 1{M (r) =0, M (r) ≥c


0 1 0n
q, r∈[R−1], S̄ (R−2) < n
4
}
(r) (r)
≤P X (R) = 0 P M0 = 0, M1 ≥ c0 nq , r ∈ [R − 1], S̄ (R−3) ≤ n



4 ,

where we lastly use that S̄ (r) increases in r. Now, the argument can be repeated on the last factor
of the above upper bound. Eventually, we then derive that
R
(r) (r)
= 0, M1 ≥ c0 nq ∀r ∈ [R], S̄ (R−2) < P(X (r) = 0).
Y
n


P M0 4 ≤
r=1

Let c > 0 denote a constant satisfying

c0 nq − 2
 2
⌊n/2⌋ ≥ cn, R ≥ cn1−2q log n, ≥ cn2q .
2
Then
R  c nq − 2 2 ⌊n/2⌋ R
 !
0
= 0) = 1−
Y
(r)
P(X
r=1
2n
cn cn1−2q log n
c

≤ 1−
n2(1−q)
n2(1−q) !c2 log n
c

= 1− →0 as n → ∞,
n2(1−q)

and this finishes the proof.

The next limit that will be established is the following

R
(r) (r)
lim ≥ M̄1 + 2, Bu = 0.
X

P M1
n→∞
r=1

Proof of Lemma 10 (2). Let C0 be the constant of Lemma 8 and C1 be the constant of Fact 9. As

26
n/2 > C1 n1−q log n we note that it is sufficient to establish that
R
(r) (r) (r)
lim ≥ M̄1 + 2, M1 = 0.
X
≤ C0 nq , S (r−1) < n


P M1 2
n→∞
r=1

Let r ∈ [R] be fixed and consider the conditional probability

(r) (r) (r)
≥ M̄1 + 2, M1 ≤ C0 nq , S (r−1) < n
M(r) .


P M1 2

(r)
Let V1 , . . . , VM (r) be the variables of the 1-SAT formula Φ1 . From (3.4) we get that these are i.i.d.
1
and uniformly distributed on [n − S (r−1) ] when conditioning on M(r) . Thus
(r) (r) (r)
≥ M̄1 + 2, M1 ≤ C0 nq , S (r−1) < n
M(r)


P M1 2
 [ 
=P Vj1 = Vj2 = v1 , Vj3 = Vj4 = v2
[  (r)
M 1{M (r) ≤C q S (r−1) < n }
1 0n , 2
v1 ,v2 j1 ,j2 ,j3 ,j4
distinct
P Vj1 = Vj2 = v1 , Vj3 = Vj4 = v2 M(r) 1{M (r) ≤C
X X

≤ q S (r−1) < n }
1 0n , 2
v1 ,v2 j1 ,j2 ,j3 ,j4
distinct
1
 4
(r)
≤(n − S (r−1) )2 (M1 )4 1{M (r) ≤C q S (r−1) < n }
n − S (r−1) 1 0n , 2

(C0 nq )4 C
≤ = 2(1−2q) ,
(n/2) 2 n
(r)
where we sum over v1 , v2 ∈ [n − S (r−1) ] and j1 , j2 , j3 , j4 ∈ [M1 ] and C = 4C04 . Therefore, we get
R
(r) (r) (r)
≥ M̄1 + 2, M1
X
≤ C0 nq , S (r−1) < n


P M1 2
r=1
R
(r) (r) (r)
= ≥ M̄1 + 2, M1
X
≤ C0 nq , S (r−1) < n
M(r)


E P M1 2
r=1
R
C C log n
→ 0 as n → ∞,
X
≤ ≤
r=1
n2(1−2q) n1−2q

as 1 − 2q > 0 for q < 1/2 and this was the claim.

The last limit to be established in this section is the following

R
(r)
lim < (n − S̄ (r−1) )+ , Bu , Bl = 0.
X

P M2
n→∞
r=1

Proof of Lemma 10 (3). Let C0 , c0 , and C1 be the constants of Lemma 8 and Fact 9. As n(4⌊log n⌋)−1 >
C1 n1−q log n we note that it is sufficient to establish that
R
(r) (r−1) (r−1)
lim < (n − S̄ (r−1) )+ , M1 = 0.
X
≤ C0 n q , M 1 ≥ c0 nq , S̄ (r−2) < n


P M2 4
n→∞
r=1

27
Let r ∈ [R] be fixed. We now consider the conditional distribution given M(r−1) and assume that
(r−1) (r−1)
M1 ≤ C 0 n q , M1 ≥ c0 nq and S̄ (r−2) < n/4 which also implies that S (r−2) < n/4. Using the
definition of equation (3.3), and the definition of p2 in Lemma 6, we get
(r) (r−1)
E[M2 |M(r−1) ] = n − S̄ (r−2) p2 n − S (r−2) , (M1 − 1)+

(r−1) (r−1)
n − S (r−2) − (M1 − 1) n − S (r−2) − (M1 − 1) − 1



= n − S̄ (r−2)

n − S (r−2) n − S (r−2) − 1

(r−1) (r−1)
M  − 1  M1 −1 
= n − S̄ 1 − 1 (r−2) 1 −
(r−2)

n−S n − S (r−2) − 1
 C0 n q  2
≥ n − S̄ (r−2) 1 −


n/2

2
≥ n − S̄ (r−2) 1 − 21 Cnq−1 ,

(4.14)
where C = 4C0 . Using the above we get


(r) (r) (r−1)
E[M2 |M(r−1) ] − n − S̄ (r−1) =E[M2 |M(r−1) ] − n − S̄ (r−2) − ⌊log n⌋ M1 −1

 
(r−1)
≥ n − S̄ (r−2) (1 − 12 Cnq−1 )2 − 1 + ⌊log n⌋ M1 −1 (4.15)

n
≥ − Cnq−1 + ⌊log n⌋ c0 nq − 1 ≥ Cnq ,



2
(r−1)
The conditional variance can also be bounded (again when M1 ≤ C0 nq and S̄ (r−1) < n/4). To
do so we again make use of (3.3) and the calculations in (4.14)
(r) (r−1)
 (r−1)

V(M2 |M(r−1) ) = n − S̄ (r−2) p2 n − S (r−2) , (M1 − 1)+ 1 − p2 n − S (r−2) , (M1 − 1)+

(r−1)
≤n 1 − p2 n − S (r−2) , (M1 − 1)+




2

≤n 1 − 1 − 12 Cnq−1 ≤ Cnq .
(4.16)
Using (4.15) and (4.16) along with Chebyshev’s inequality we get
 
(r) (r) (r)
< n − S̄ (r−1) M(r−1) ≤P M(r−1) > Cnq M(r−1)

 
P M2 M2 − E M2
(r)
V(M2 |M (r−1) ) 1
≤ ≤ .
(Cnq )2 Cnq

This finally implies that

(r) (r−1) (r−1)
< (n − S̄ (r−1) )+ , M1 ≤ C0 n q , M 1 ≥ c0 nq , S̄ (r−2) < n


P M2 4
h i
(r)
=E P M2 < n − S̄ (r−1) M(r−1) 1{M (r−1) ≤C


(r−1)
1 0 nq , M1 ≥c0 nq , S̄ (r−1) < n
4
}
1 (r−1) (r−1) 1
≤ C0 n q , M 1 ≥ λnq , S (r−1) < n/4 ≤


≤ q P M1 ,
Cn Cnq

28
and thus we get the limit
R
(r) (r−1) (r−1)
< (n − S̄ (r−1) )+ , M1
X
≤ C0 n q , M 1 ≥ c0 nq , S̄ (r−2) < n


P M2 4
r=1
R
1 1
· n1−3q log n → 0 as n → ∞,
X
≤ q
≤
r=1
Cn C

where we use that 1 − 3q < 0 when q > 1/3.

4.3 Proof of Lemma 12

The four limits of this lemma are established one at a time. Remember that we use the defined
elements of Section 3.3, i.e. the unit-propagation procedure elements constructed for the case
q < 1/3. To begin with we want to show that
(r)
lim P M1 ≤ nq log n, r ∈ [R] = 1.


n→∞

Proof of Lemma 12 (1). For each r ∈ [R] we use (3.10) and the definition of p1 in Lemma 6 to get
that
(r) (r−1)

E[M1 |M(r−1) ] = n − S (r−1) p1 n − S (r−2) , M1

(r−1)
2
n − S (r−2) − M1 (r−1) (r−1)
=
· M1 ≤ M1 .
−1


n−S (r−2) n−S (r−2)

(r−1) (r−1)
The last inequality is obviously true when M1 ≥ 1 and when M1 = 0 both sides of the
(r) (R)
inequality equals zero. Now, by letting M1 = M1 and M = M(r) (R) for r > R we can extend
(r)
our sequence and consider {M1 }r∈N0 which then becomes a super-martingale w.r.t. the filtration
{M(r) }r∈N0 . Define the stopping time
(r) (r)
τ = min{r ∈ N0 : M1 = 0 or M1 > nq log n}.
(r)
Let C0 be the constant of Lemma 11. As our sequence {M1 }r∈N0 is a non-negative super-
martingale we can make use of the optional sampling theorem (Thm. 28, Chapter V in [DM11]).
Hereby we get that
(0) (τ ) (τ )
C0 nq ≥ E[M1 ] ≥ E[M1 ] ≥ nq log n · P M1 > nq log n .

Rearranging the above terms implies that

(τ )
> nq log n ≤ C0 log−1 n.


P M1

As the sequence terminates when hitting zero this establishes the result.

The next limit to establish is the following

(r) (r)
lim P Φ1 ∈ SAT, r ∈ [R] M1 ≤ nq log n, r ∈ [−1, R] = 1.


n→∞

29
 (r)
Proof of Lemma 12 (2). We will use that when M1 ≤ nq log n for all r ∈ [−1, R] then S (r) ≤ n
2
for all r ∈ [0, R]. We will further use that if X, Y, Z are random variables then
D
⊥ (Y, Z) ⇒ (X ⊥
X⊥ ⊥ Y )|Z, and (X ⊥⊥ Y )|Z ⇒ X|(Y, Z) = X|Z. (4.17)

From (3.11) we got that

(r) (r)
Φ1 ⊥⊥ Φ2 | M(r) , (r ∈ [R]).

(r) (r) (r) (r)

The random function Ψ2 is constructed from Φ1 and Φ2 , but we noticed in (3.12) that Ψ2
(r)
and Φ1 are independent given M(r) . From this point and on all remaining random objects are
(r)
constructed from Ψ2 and from L(r) , which is deterministic given M(r) , and then also from random
(r)
objects that are defined independently of Φ1 |M(r) . This implies that
(r) (r+1) (R) (r+1) (R) 
Φ1 ⊥
⊥ (Mk )k∈K , . . . , (Mk )k∈K , Φ1 , . . . , Φ1 M(r) , (r ∈ [R]).


Thus, the first implication of (4.17) implies that

(r) (r+1) (R)

Φ1 ⊥
⊥ Φ1 , . . . , Φ1 M(R) , (r ∈ [R]),

and the second implication of (4.17) gives that

(r) D (r)
Φ1 |M(r) = Φ1 |M(R) , (r ∈ [R]).

From (3.11) we have that

(r) (r)

Φ1 |M(r) ∼ F1 n − S (r−1) , M1 , (r ∈ [R]),

m m (r)
and Lemma 7 states that if Φ1 ∼ F1 (n, m) then P Φ1 ∈ SAT ≥ 1 − . Thus, when M1



n ≤
nq log n it holds that
nq log n
nq log n

(r)
P Φ1 ∈ SAT M (r)

≥ 1− .
n/2

30
Combining the above we get that
(r) (r)
P Φ1 ∈ SAT, r ∈ [R] M1 ≤ nq log n, r ∈ [−1, R]


h i
(r) (r)
=E P Φ1 ∈ SAT, r ∈ [R] M(R) M1 ≤ nq log n, r ∈ [−1, R]

h i
(1) (r) (r)
=E P Φ1 ∈ SAT M(1) P Φ1 ∈ SAT, r ∈ [2, R] M(R) M1 ≤ nq log n, r ∈ [−1, R]

nq log n
nq log n

(r) (r)
≥ 1− P Φ1 ∈ SAT, r ∈ [2, R] M1 ≤ nq log n, r ∈ [−1, R]


n/2
nq log n !R
nq log n

≥··· ≥ 1−
n/2
 1
 n1−q log n


log n
2
1 − 1−q  →1 as n → ∞
 
≥
 n


log n

which was the claim.

Next up, we will establish that

(r) (r)
lim P M0 = 0, r ∈ [R] M1 ≤ nq log n, r ∈ [−1, R] = 1,


n→∞

Proof of Lemma 12 (3). In (3.10) it is stated that

 
(r) (r−1)

(Mk )k∈K M(r−1) ∼ Multinomial n − S (r−1) , p n − S (r−2) , M1 , (r ∈ [R]). (4.18)

We will use the following fact:

pi

If (X1 , . . . , Xn ) ∼ Multinomial n, (p1 , . . . , pn ) then Xi |Xj ∼ Binomial n − Xj , for i ̸= j.


1−pj

This implies that

 
(r−1)

(r) (r) (r) p0 n − S (r−2) , M1
M0 |M(r−1) , M1 ∼ Binomial n − S (r−1) − M1 , (r−1)

, (r ∈ [R]).
1 − p1 n − S (r−2) , M1
(4.19)
(r)
Now, given M1 ≤ log n for all r ∈ [−1, R] we get that
nq ≤ n/4 for all r ∈ [0, R] and using
S (r)
the definitions of p0 and p1 given in Lemma 6 we get for each r ∈ [R]:

(r−1)
(nq log n)2 nq log n 3 (R−1)

p0 n − S (r−1) , M1 ≤ , 1 − p1 n − S (R−1) , M1 ≥ . ≥1−
( 34 n)2 3
4n
4
(4.20)
Using (3.10) we also note that there exists functions g (r) , r ∈ [R] such that for r ∈ [R − 1] we have

31
that
(s) (r) (r)

P M1 = m(s) , s ∈ [r + 1, R] M(r−1) , M1 , M0
(R) (r) (r) 
=E P M1 = m(R) M(R−1) 1{M (s) =m(s) , s∈[r+1,R−1]} M(r−1) , M1 , M0


1
(−1) (r) (s) (r) (r)

=g (R)
(M1 , . . . , M1 , m(r+1) , . . . , m(R) )P M1 = m(s) , s ∈ [r + 1, R − 1] M(r−1) , M1 , M0
R
(−1) (r)
=··· =
Y
g (s) M1 , . . . , M1 , m(r+1) , . . . , m(s) .

s=r+1
(4.21)
(r+1) (R) (r)
This implies that (M1 , . . . , M1 ) is independent of M0 when conditioning on M(r−1) and
(r)
M1 . Now using (4.19) and (4.20) we get that
(r) (r)
= 0, r ∈ [R] M1 ≤ nq log n, r ∈ [−1, R]


P M0
h i
(R) (R)
(r)
=E P M0 = 0 M(R−1) , M1 1{M (r) =0, r∈[R−1]} M1 ≤ nq log n, r ∈ [−1, R]
0

(nq log n)2 n

 
(r)
≥E 1− 1{M (r) =0, r∈[R−1]} M1 ≤ nq log n, r ∈ [−1, R]
( 12 )3 n2 0
!n
(nq log n)2 (r) (r)
= 1− = 0, r ∈ [R − 1] M1 ≤ nq log n, r ∈ [−1, R] .


P M0
( 12 )3 n2

Then using (4.21) the above argument can be repeated on the last factor above
(r) (r)
= 0, r ∈ [R − 1] M1 ≤ nq log n, r ∈ [−1, R]


P M0
(R−1) (R−1) (R)
(r)
=E P M0 = 0 M(R−2) , M1 ≤ nq log n, r ∈ [−1, R]
 
, M1 1{M (r) =0, r∈[R−2]} M1
0
(R−1) (R−1)
(r)
=E P =0M (R−2)
≤ nq log n, r ∈ [−1, R]
 
M0 , M1 1{M (r) =0, r∈[R−2]} M1
0
!n
(nq log n)2 (r) (r)
≥ 1− = 0, r ∈ [R − 2] M1 ≤ nq log n, r ∈ [−1, R] .


P M0
( 12 )3 n2

Repeating the above R times in total we eventually arrive at the expression

(r) (r)
= 0, r ∈ [R] M1 ≤ nq log n, r ∈ [−1, R]


P M0
 1
 n2(1−q) log n

!Rn 
2n
(nq log n)2 23 
log

≥ 1− 1 − → 1, as n → ∞,
 
≥ 
( 12 )3 n2  n2(1−q) 
log2 n

which was the claim.

Next, we will establish that

(R)
= 0 → 1 as n → ∞,


P M1

i.e. we will now establish that our process of 1-clauses terminates in less than R rounds w.h.p.
We will show this by proving that our recursive sequence of Binomial random variables can be
approximated by a recursive sequence of Poisson random variables. Afterwards, it is proven that

32
the recursive sequence of Poisson random variables terminates.
(−1)
Lemma 17. Let (X (r) )r∈[−1,R] be a sequence of random variables where X (−1) = M1 , X (0) =
(0)
M1 and X (r) |X (r−1) ∼ Poisson X (r−1) for r ∈ [R]. Then for x(−1) , x(0) , . . . , x(R) ∈ [0, ⌊nq log n⌋]

it holds that
(r) (−1) (r)
= x(r) , r ∈ [R] M1 = x(−1) , M1 = x(0)


P M1
≥P X (r) = x(r) , r ∈ [R] X (−1) = x(−1) , X (0) = x(0) · E(n),

where E is a function satisfying that limn→∞ E(n) = 1.

Proof. We will make use of the below inequality which holds for y > x > 0 and z > 0:
y−z y ! !
x x x x x2  x2
     
1− ≥ 1− = exp y log 1 − ≥ exp y − − 2 = exp (−x) exp − .
y y y y y y

Let s(−2) = 0 and s(r) = s(r−1) + x(r) for r ∈ [−1, R]. The elements x(−1) , . . . , x(R) are chosen such
that s(r) ≤ n2 for all r ∈ [−1, R]. Also, note that the definition of p1 in Lemma 6 implies that

(n − f )f f
2
≤ p1 (n, f ) ≤ , (n ≥ f ≥ 0).
n n
Using the above inequalities along with Lemma 6 we now get
(r) (−1) (0)
= x(r) , r ∈ [R] M1 = x(−1) , M1 = x(0)


P M1
R
(r) (s)
= = x(r) M1 = x(s) , s ∈ [−1, r − 1]
Y

P M1
r=1
R
!
n − s(r−1)  x(r)  n−s(r−1) −x(r)
= p1 (n − s(r−2) , x(r−1) ) 1 − p1 (n − s(r−2) , x(r−1) )
Y
x (r)
r=1
R (r)
#x(r) " #n−s(r−2) −(x(r−1) +x(r) )
(n − s(r−1) − x(r) )x (n − s(r−1) )x(r−1)
"
x(r−1)
1−
Y
≥
r=1
x(r) ! (n − s(r−2) )2 n − s(r−2)
R
#2x(r) (r−1) (r)
n − s(r−2) − (x(r−1) + x(r) ) (x(r−1) )2 (x(r−1) )x
" !
e−x
exp −
Y
≥
r=1
n − s(r−2) n − s(r−2) x(r) !
R  2nq log n
2nq log n n2q log2 n
!
1− exp − P X (r) = x(r) |X (r−1) = x(r−1)
Y

≥
r=1
n/2 n/2
1−q
 nlog  log2 n
4 2
 n
 
1− exp − P X (r) = x(r) , r ∈ [R] X (−1) = x(−1) , X (0) = x(0) .


≥ n1−q
log n
log n

And as 1−q
 nlog  log2 n
4 2
 n
 
1− exp − → 1 as n → ∞,
n1−q
log n
log n

the result follows.

33
 Let (X (r) )r∈[−1,R] be the sequence of random variables from the above lemma. Note that

(R) (R) (r)

= 0 ≥ P M1 = 0, M1 ≤ nq log n, r ∈ [−1, R]



P M1
⌊nq log n⌋ ⌊nq log n⌋
(−1) (R−1) (R)
= = x(−1) , . . . , M1 = x(R−1) , M1 =0 ,
X X

··· P M1
x(−1) =0 x(R−1) =0

and using Lemma 17 and letting x(R) = 0 each summand can be upper bounded by
(−1) (R)
= x(−1) , . . . , M1 = x(R)


P M1
(r) (−1) (0) (−1) (0)
=P M1 = x(r) , r ∈ [R] M1 = x(−1) , M1 = x(0) P M1 = x(−1) , M1 = x(0)

≥P X (r) = x(r) , r ∈ [R] X (−1) = x(−1) , X (0) = x(0) P X (−1) = x(−1) , X (0) = x(0) E(n)

=P X (−1) = x(−1) , . . . , X (R) = x(R) E(n).

Inserting this lower bound in the sum gives that

(R)
= 0 ≥ E(n)P X (R) = 0, X (r) ≤ nq log n, r ∈ [−1, R − 1] , (4.22)



P M1

where E is the function from Lemma 17. To establish our result we thus only need the two lemmas
below
Lemma 18. We have that

P X (r) ≤ nq log n, r ∈ [−1, R − 1] → 1

as n → ∞.

Lemma 19. We have that

P X (R) = 0 → 1


as n → ∞.

Proof of Lemma 18. Let X (r) := X (R) for r > R and also define the σ-algebras F (r) = σ X (−1) , . . . , X (r)

for r ≥ −1. Then for each r ≥ −1 we have

E[X (r) |F (r−1) ] = X (r−1) ,

why (X (r) )r≥−1 is a martingale w.r.t. the filtration (F (r) )r≥−1 . As it is non-negative, we can make
use of optional sampling (Thm. 28, Chapter V of [DM11]). Consider the stopping time

τ = min{r ∈ N0 : X (r) = 0 or X (r) ≥ nq log n},

and let C0 be the constant of Lemma 11. Then

C0
C0 nq ≥ E[X (0) ] ≥ E X (τ ) ≥ nq log nP X (τ ) ≥ nq log n ⇒ P X (τ ) ≥ nq log n ≤
 


.
log n
As 0 is an absorbing state this implies that

P X (r) ≤ nq log n, r ∈ [−1, R] → 1 as n → ∞,

which was the claim.

Proof of Lemma 19. Note that the distribution of (X (r) )r∈[R] has the same law as a critical Galton-
Watson tree with offspring distribution P oisson(1) cut off at depth R, see Chapter 1 in [ANN04].

34
Thus, using standard results for such processes (see e.g. Thm. 1 in section 1.9 of [ANN04]), there
exists a constant C > 0 such that

C X (0) C E[X (0) ] Cn3q−1 log3 n C0 nq

   
P(X (R)
= 0) ≥ E 1− ≥ 1− 1−2q ≥ 1− → 1, as n → ∞,
R n log−3 n nq

where C0 is the constant of Lemma 11 and Jensen’s inequality is also used. Thus, the result is
established.

4.4 Establishing Definition 1 (1) using Lemma 11

We will now couple Lemma 11 to our main result in the regime q < 1/3. We will do this by closely
controlling the first couple of rounds in the unit-propagation algorithm. Thus, we will once again
repeat the notation used when going through this procedure. However, as this section uses none of
the defined elements from the other sections this will not be a problem.
Let Φ ∼ F2 (n, n) and let L ⊆ ±[n] be a consistent set of literals with |L| = ⌊nq ⌋. We need to
show that lim inf n→∞ P(ΦL ∈ SAT) = lim inf n→∞ P(Φ ∈ SAT). Let G be the function of Lemma 4
and define
(0) (0)
Ψ2 := G(Φ, L), N1 = ⌊nq ⌋, T (−1) := 0,
T (0) := ⌊nq ⌋, L(0) := [n]\[n − T (0) ], N (0) = {∅, Ω}.

(0)
Note that Ψ2 ∼ F2 (n, n) and also
(0)
ΦL ∈ SAT = (Ψ2 )L(0) ∈ SAT . (4.23)
 

Unlike previously we now only repeat the unit-propagation procedure twice. Thus, recursively for
r = 1, 2 define the following:
(r) (r−1) (r)
Let G1 and G2 be the functions of Lemma 6 and define Φk = Gk (Ψk , L(r−1) ). Let Nk be
(r) (r) (r)
the number of clauses in Φk for k ∈ {1, 2} and let N0 and N⋆ be the number of unsatisfied- and
(r−1) (r)
satisfied
clauses of (Ψ2 )L(r−1) , respectively. Define the σ-algebra N (r) = σ N (r−1) ∪ σ(Nk :
k ∈ K) . The elements are constructed such that
(r−1) (r) (r) (r)
(Ψ2 )L(r−1) ∈ SAT = (Φ2 )Φ(r) ∈ SAT, Φ1 ∈ SAT, N0 =0 . (4.24)
 
1

and also
 
(r) (r−1)

(Nk )k∈K |N (r−1) ∼ Multinomial n − T (r−2) , p n − T (r−2) , N1 , (4.25)
(r) (r)
Φk |N (r) ∼ Fk (n − T (r−1) , Nk ), (k ∈ {1, 2}),

(r) (r) (r)

and Φ1 and Φ2 are conditionally independent. Let N̄1 be the number of distinct variables
(r)
appearing in Φ1 and define further
(r) (r)
T (r) := T (r−1) + N1 , L̄(r) := [n − T (r−1) ]\[n − T (r−1) − N̄1 ], L(r) := [n − T (r−1) ]\[n − T (r) ].
(r) (r) (r)
Also, let Ψ2 := G(Φ2 , L(Φ1 )), where again G is defined in Lemma 4. Then using Lemma 6 we

35
see
(r) (r)

Ψ2 |N (r) ∼ F2 n − T (r−1) , N2 ,
(r) (r) (r)
(Φ2 )Φ(r) ∈ SAT = (Ψ2 )L̄(r) ∈ SAT ⊇ (Ψ2 )L(r) ∈ SAT , (4.26)
  
1

where we lastly used that L̄(r) ⊆ L(r) . Now, we are in the same setup as initially and our recursive
step has ended. Combining (4.23), (4.24), and (4.26), we now see
(2) (2) (1) (2) (1)
P ΦL ∈ SAT ≥ P (Ψ2 )L(2) ∈ SAT, Φ1 ∈ SAT, Φ1 ∈ SAT, N0 = 0, N0 =0 . (4.27)

The above equation implies that it is sufficient to lower bound the right-hand side of the above
expression to establish our main theorem in the case q < 1/3. We do this by proving the below
lemmas.

Lemma 20. We have

(1) (2)
lim P N0 = 0 = lim P N0 = 0 = 1.



n→∞ n→∞

and
(1) (2)
lim P Φ1 ∈ SAT = lim P Φ1 ∈ SAT = 1.



n→∞ n→∞

Lemma 21. We have

(2)
lim inf P (Ψ2 )L(2) ∈ SAT ≥ lim inf P Φ ∈ SAT .



n→∞ n→∞

These lemmas will imply our main theorem when q < 1/3

Proof of Definition 1 (1). As {ΦL ∈ SAT} ⊆ {Φ ∈ SAT} this implies that

lim inf P ΦL ∈ SAT ≤ lim inf P Φ ∈ SAT .

n→∞ n→∞

On the other hand, equation (4.27) along with Lemma 20 and Lemma 21 gives

lim inf P ΦL ∈ SAT

n→∞
(2) (2) (1) (2) (1)
≥ lim inf P (Ψ2 )L(2) ∈ SAT, Φ1 ∈ SAT, Φ1 ∈ SAT, N0 = 0, N0 =0


n→∞
≥ lim inf P Φ ∈ SAT .


n→∞

Combining the above implies that the two limit infimum coincide.

To prove our main theorem it thus suffices to establish Lemma 20 and 21. To do so we need
the following technical result.
(r)
Lemma 22. There exists a constant C0 > 0 such that E[Nk ] ≤ C0 nq for r ∈ {1, 2} and k ∈ {1, ⋆}.
Furthermore,
(r)
lim P Nk ≤ 12 nq log n = 1, k ∈ {1, ⋆}, r ∈ {1, 2},


n→∞

and
(1) (2)
lim P N⋆ + N⋆ ≥ nq = 1.


n→∞

36
Proof. Note that p0 (n, ⌊nq ⌋) ≤ p1 (n, ⌊nq ⌋) ≤ p⋆ (n, ⌊nq ⌋) ≤ nq−1 , see the definitions in Lemma 6,
and thus
(1)
E[Nk ] = n · pk (n, ⌊nq ⌋) ≤ nq , (k ∈ {0, 1, ⋆}). (4.28)
Using the previous observations, we get

(1) (1) nq
E p1 (n − T (0) , N1 ) ≤ E p⋆ (n − T (0) , N1 ) ≤
   
,
n − nq − 2
and thus if we let q1 ∈ ( 2q , q), then

(2)  (1) (1) n1+q

E[Nk ] = E N2 · pk (n − T (0) , N1 ) ≤ ≤ nq + nq1 , (k ∈ {1, ⋆}), (4.29)

n − nq − 2
(1)
where we used the upper bound N2 ≤ n. Note that (4.28) and (4.29) imply the first claim of the
lemma. Furthermore, these two equations along with Markov’s inequality give
(r)
(r) E[Nk ] nq + nq1
> 12 nq log n < →0 as n → ∞, (k ∈ {1, ⋆}, r ∈ {1, 2}),


P Nk ≤
2 n log n 2 n log n
1 q 1 q

and this is the second claim of the lemma. Next, using (4.25), (4.28) and a Chernoff bound we get
(1) (1)
≥ E[Nk ] + nq1 ≤ exp − 13 n2q1 −q , (k ∈ {0, 1, ⋆}),



P Nk
(1) (1)
E[Nk ] q1

≤ exp − 13 n2q1 −q , (k ∈ {0, 1, ⋆}).


P Nk ≤ −n

(1)
Using this, (4.28), and also that E[Nk ] ≥ nq − nq1 for k ∈ {1, ⋆} (this is a direct consequence of
the definition) we see that
(1)
≥ nq + nq1 ≤ exp − 31 n2q1 −q , (k ∈ {0, 1, ⋆}),



P Nk
(1)
(4.30)
1 2q1 −q
≤ n − 2n
q q1

≤ exp − (k ∈ {1, ⋆}).


P N1 3n ,

Next, we again use that for X ∼ Binomial(n, p) we have E[X 2 ] ≤ E[X] + E[X 2 ], see (4.1). Then
we get the bound:
h
2 
(2) (2) (2)
E (N⋆ )2 ≤ E E[N⋆ |N (1) ] + E[N⋆ |N (1) ]
 

h
2 i
(1) (1) (1) (1)
= E N2 · p⋆ (n − T (0) , N1 ) + N2 · p⋆ (n − T (0) , N1 )
2 
n n
 
(1) (1) 
≤ E[N ] + E (N1 )2
n−n −2q 1
n−n −2q
2 (4.31)
n n

(1) (1)
q
+ E[N1 ] + (E[N1 ])2


≤ n
n−n −2q n−n −2q
2 4
n n
 
≤2 n +q
n2q
n − nq − 2 n − nq − 2
≤ n2q + Cnq ,

for a constant C chosen large enough. We also want to lower bound the mean. Using that
(2) (2) (2) (2)
N2 = n − N0 − N1 − N⋆ ,

37
and that
(2) (1) (1)
N⋆ |N (1) ∼ Binomial N2 , p⋆ (n − T (0) , N1 ) ,

(n−l)l
where p⋆ (n, l) ≥ p1 (n, l) ≥ n2
, we get that
 (2) (1) (1)
> nq − 2nq1 and Nk < nq + nq1 for k ∈ {0, 1, ⋆}

E N⋆ N1

n − nq − (nq + nq1 ) (nq − 2nq1 )


≥ n − 3(n + n ) q q1
≥ nq − Cnq1 ,
n2
again for C chosen large enough. This and (4.30) now implies that
(2) (1) (1)
E[N⋆ ] ≥(nq − Cnq1 )P N1 > nq − 2nq1 and Nk < nq + nq1 for k ∈ {0, 1, ⋆}


 
(1) (1)
≥(nq − Cnq1 ) 1 − P N1 ≤ nq − 2nq1 − ≥ nq + nq1

X
P Nk
k∈{0,1,⋆}

≥(n − Cn ) 1 − 4 exp
q q1
− 13 n2p1 −q



,

and by redefining C we get that

 (2) 
E N⋆ ≥ nq − Cnq1 . (4.32)
Combining this with (4.31) we now get
(2)
(2) (2)
2
V N⋆ = E[(N⋆ )2 ] − E[N⋆ ] ≤ Cnq+q1 ,

where C is again redefined. Now, let q2 ∈ ( q+q

2 , q). Then
1

(2)

(2) (2) V N⋆ nq+q1
P N⋆ − E[N⋆ ] >n q2

≤ ≤C → 0 as n → ∞.
n2q2 n2q2
Therefore, using this and (4.32) we see
(2)
≥ nq − Cnq1 − nq2 → 1 as n → ∞


P N⋆

and this along with (4.30) gives

(1) (2) (1) (2)
P N⋆ + N⋆ > nq ≥ P N⋆ ≥ nq − 2nq1 , N⋆ ≥ nq − Cnq1 − nq2 → 1 as n → ∞,

which finishes the proof.

Now, we can prove our two remaining lemmas of this section.

l2
Proof of Lemma 20. Using Lemma 6 we note that p0 (n, l) ≤ 4(n−1)2
, why
n
n2q

(1)
=0 ≥ 1− → 1, as n → ∞,


P N0
4(n − 1)2

38
and
(2) (2) (1) (1)
= 0 ≥ P N0 = 0 N1 ≤ 21 nq log n P N1 ≤ 12 nq log n




P N0

2 n
log n 1 q
2n

(1)
≥ 1− ≤ 12 nq log n → 1 as n → ∞,


P N1
4(n − nq − 2)2

where we lastly used Lemma 22. Now, using this Lemma again along with Lemma 7 we get
(1) (1) (1) (1)
P Φ1 ∈ SAT ≥ P Φ1 ∈ SAT N1 ≤ 12 nq log n P N1 ≤ 12 nq log n

 1 nq log n
log n 1 q
2n

2 (1)
≥ 1− ≤ 12 nq log n → 1 as n → ∞.


P N1
n − nq − 2

A similar argument gives that

(2) (2) (r) (r)
P Φ1 ∈ SAT ≥ P Φ1 ∈ SAT N1 ≤ 12 nq log n, r ∈ {1, 2} P N1 ≤ 12 nq log n, r ∈ {1, 2}

 12 nq log n
log n
1 q
2n

(r)
≥ 1− ≤ 12 nq log n, r ∈ {1, 2} → 1 as n → ∞.


P N1
n − nq log n

Proof of Lemma 21. Remember that

(1) (1) (2)
L(2) = [n − T (1) ]\[n − T (2) ] = [n − ⌊nq ⌋ − N1 ]\[n − ⌊nq ⌋ − N1 − N1 ],
(2) (r)
and using that N2 =n− we see
P
k∈{0,1,⋆},r∈{1,2} Nk

(2) (2) (1) (1) (2) (2)

Ψ2 |N (2) ∼ F2 (n − T (1) , N2 ) = F2 (n − T (1) , n − T (2) + ⌊nq ⌋ − N0 − N⋆ − N0 − N⋆ ).
(2) (2) (2)
Let the literals of Ψ2 be given by (Lj,1 , Lj,2 ) for j ∈ [N2 ]. If n − T (2) > N2 define additional
(2)
random variables (Lj,1 , Lj,2 ) for j ∈ [N2 + 1, n − T (2) ], where conditional on N (2) the pairs of
random variables are independent and uniformly distributed in {(ℓ1 , ℓ2 ) ∈ (±[n − T (1) ])2 : |ℓ1 | <
|ℓ2 |}. Define
n−T (2)
′
Φ = (Lj,1 ∨ Lj,2 ),
^

j=1

and let also L′ := L(2) . Then

(2) (2)
(2)

P (Ψ2 )L(2) ∈ SAT ≥ P Φ′L′ ∈ SAT, n − T (2) ≥ N2 ≥ P Φ′L′ ∈ SAT + P n − T (2) ≥ N2 − 1.

Note that
(2)  (1) (2)
n − T (2) ≥ N2 ⊇ N⋆ + N⋆ ≥ nq ,


(2)

and thus this along with Lemma 22 implies that limn→∞ P n − T (2) ≥ N2 = 1. Thus
(2)
lim inf P (Ψ2 )L(2) ∈ SAT ≥ lim inf P Φ′L′ ∈ SAT . (4.33)



n→∞ n→∞

39
Define now further
(−1) (1) (0) (2) (−1) (0)

M1 := ⌊nq ⌋ + N1 , M1 := N1 , M(0) := σ M1 , M1 .

Note that as M(0) ⊆ N (2) and

(−1) (−1) (0)

Φ′ |N (2) ∼ F2 n − T (1) , n − T (2) = F2 n − M1 + M1


, n − M1 ,

we get that
(−1) (−1) (0)

Φ′ |M(0) ∼ F2 n − M1 , M1 + M1
Moreover, the definitions imply that
(−1) (−1) (0)
L′ = [n − M1 ]\[n − M1 − M1 ].
(0)
Lemma 21 also gives that E[M1 ] ≤ C0 nq for some C0 > 0 and also that
(−1) (0)
lim P M1 ≤ nq log n = lim P M1 ≤ nq log n = 1



n→∞ n→∞

Thus, our defined elements satisfy all assumptions of Lemma 11. Therefore

lim inf P Φ′L′ ∈ SAT ≥ lim inf P Φ ∈ SAT .

n→∞ n→∞

Combining this with (4.33) establishes the lemma.

4.5 Proof of technical lemmas

We begin by proving Lemma 4 and this Lemma is established by a coupling argument where literals
are swapped. The swapping will not change the distribution of the resulting formula as the clauses
are uniformly distributed.

Proof of Lemma 4. Let φ be a non-random SAT-formula with n variables and m clauses and let
(ℓj,i )j∈[m],i∈[2] denote its literals. Write L = {ℓ1 , . . . , ℓm }, where |ℓ1 | < · · · < |ℓl | and let Labs =
{|ℓj | : j ∈ [f ]}. Also let 0 < ℓf +1 < . . . < ℓn ∈ [n] be defined such that {|ℓ1 |, . . . , |ℓn |} = [n].
Define a function Γ : [n] → [n] by letting Γ(|ℓi |) = n − i for i ∈ [n]. Then Γ is a permutation
satisfying that Γ(Labs ) = L′ , where L′ = [n]\[n − f ]. Define another function θ : [n] → {±1} where
θ(|ℓi |) = sgn(ℓi ). Define a new SAT-formula φ′ with literals (ℓ′j,i )j∈[m],i∈[2] , where

{ℓ′j,1 , ℓ′j,2 = θ(|ℓj,i |) · sgn(ℓj,i ) · Γ(|ℓj,i |) : i ∈ [2] .



Then we define G(φ, L) := φ′ . Let x = (x1 , . . . , xn ) ∈ Bn and define x′ = (x′1 , . . . , x′n ) ∈ Bn by

letting x′v = xΓ−1 (v) for v ∈ [n]. Note that x 7→ x′ is a bijection. Let for j ∈ [m], i ∈ [2] be chosen
such that ℓ′j,1 = θ(|ℓj,i |) · sign(ℓj,i ) · Γ(|ℓj,i |).
Now, if ℓj,i ∈ L then (ℓj,i )L (x) = true. Also, there exists a v ∈ [f ] such that ℓj,i = ℓv and also
Γ(|ℓv |) ∈ L′ . Thus
′
(lj,1 )L′ (x′ ) = θ(|ℓv |) · sgn(ℓv ) · Γ(|ℓv |) (x′ ) = Γ(|ℓv |)L′ (x′ ) = true.
 
L′

If −ℓj,i ∈ L then (ℓj,i )L (x) = false. Again there exists v ∈ [l] such that ℓj,i = −ℓv and also

40
Γ(|ℓv |) ∈ L′ and thus
′
(lj,1 )L′ (x′ ) = θ(|ℓv |) · sgn(−ℓv ) · Γ(|ℓv |) (x′ ) = −Γ(|ℓv |)L′ (x′ ) = false.
 
L′

Lastly, if ±ℓj,i ∈
/ L then (ℓj,i )L (x) = ℓj,i (x) and also ±Γ(|ℓj,i |) ∈
/ L′ . Therefore
′
(lj,1 )L′ (x′ ) = sgn(ℓj,i ) · Γ(|ℓj,i |) (x′ ) = sgn(ℓj,i ) · x′Γ(|ℓj,i |) = sgn(ℓj,i ) · x|ℓj,i | = ℓj,i (x).
 

Repeating the argument on ℓ′j,2 implies that

(lj,1 ∨ lj,2 )L (x) = (ℓj, 1′ ∨ lj,2

′
)L′ (x′ ),

and thus φL ∈ SAT if and only if φ′L′ ∈ SAT.

Let Φ ∼ F2 (n, m) and define Φ′ = G(Φ, L). Then the above argument implies that

G(Φ, L) ∈ SAT = Φ′[n]\[n−l] ∈ SAT .

 

Let (Lj,i )j∈[m],i∈[2] be the random literals of Φ and let (L′j,i )j∈[m],i∈[2] be the random literals of Φ′ .
Note that as the clause L′j,1 , L′j,2 is constructed from Lj,1 , Lj,2 for each j ∈ [m] the clauses

of Φ′ are independent. Let j ∈ [m] and assume WLOG that Γ(|Lj,1 |) < Γ(|Lj,2 |). Then, for
(ℓ′1 , ℓ′2 ) ∈ (±[n])2 with |ℓ′1 | < |ℓ′2 | we have
 
P L′j,i = ℓ′i , i ∈ [2] = P θ(|Lj,i |) · sgn(Lj,i ) · Γ(|Lj,i |) = ℓ′i , i ∈ [2]

 
= P |Lj,i | = Γ−1 (|ℓ′i |), sgn(Lj,i ) = θ Γ−1 (|ℓ′i |) · sgn(ℓ′i ), i ∈ [2]

= P Lj,i = ℓi , i ∈ [2] ,

where ℓi = θ Γ−1 (ℓ′i ) · sgn(ℓ′i ) · Γ−1 (|ℓ′i |) for i ∈ [2] and as the clause (Lj,1 , Lj,2 ) is uniformly

distributed, the result follows.

More or less direct calculations imply the next lemma. We recall that equation (3.1) defines
the sets
A0 (n, f ) := −L × −L, A1 (n, f ) := ±[n − f ] × −L,
A2 (n, f ) := ±[n − f ] × ±[n − f ], A⋆ (n, f ) := ±[n] × L.

Proof of Lemma 6. Let Φ ∼ F2 (n, m) and let L = (Lj,i )j∈[m],i∈[2] be its literals. As the clauses are
i.i.d. and
Mk = j ∈ [m] : (Lj,1 , Lj,2 ) ∈ Ak (n, f ) , (k ∈ K),


where each clause belongs to exactly one of the sets Ak (n, f ) for k ∈ K this implies that

M := (M0 , M1 , M2 , M⋆ ) ∼ Multinomial m, p(n, f ) ,

where p = (p0 , p1 , p2 , p⋆ ) and pk (n, f ) = P (L1,1 , L1,2 ) ∈ Ak (n, l) for k ∈ K. As the clauses

are uniformly distributed on D = {(ℓ1 , ℓ2 ) ∈ (±[n])2 : |ℓ1 | < |ℓ2 |} we further get that pk (n, l) =

41
|Ak (n,f )∩D|
|D| for k ∈ K, so

f

f (f − 1)
p0 (n, f ) = 2

= ,
22 n2
4n(n − 1)
2(n − f )f (n − f )f
p1 (n, f ) = = ,
2 2
2 n

n(n − 1)
22 n−f (n − f )(n − f − 1)


p2 (n, f ) = 2 n2
= ,
2 2 n(n − 1)
f (n − f4 − 34 ))
p⋆ (n, f ) = 1 − p0 (n, f ) − p1 (n, f ) − p2 (n, f ) = ,
n(n − 1)

We will need the following result to establish the last part of the lemma. For X, Y independent
random functions, sets A, B, and elements x, y with x ∈ A and y ∈ B we have

P X = x, Y = y|X ∈ A, Y ∈ B = P X = x|X ∈ A P Y = y|Y ∈ B .

Define Ck = j ∈ [m] : (Lj,1 , Lj,2 ) ∈ Ak (n, f ) for k ∈ K and let C = (Ck )k∈K . For elements


ℓj,1 ∈ ±[n − f ] for j ∈ C1 and (ℓj,1 , ℓj,2 ) ∈ A2 (n, f ) for j ∈ C2 we use the independence of the clauses
and the above equation and get
 
P Φ1 = (ℓj,1 ) C = P Lj,1 = ℓj,1 j ∈ C1 ,
^ Y

j∈C1 j∈C1
  (4.34)
P Φ2 = (ℓj,1 ∨ ℓj,2 ) C = P (Lj,1 , Lj,2 ) = (ℓj,1 , ℓj,2 ) j ∈ C2 ,
^ Y

j∈C2 j∈C2

and  
P Φ1 = (ℓj,1 ), Φ2 = (ℓj,1 ∨ ℓj,2 ) C
^ ^

j∈C1 j∈C2
(4.35)
= P Lj,1 = ℓj,1 |j ∈ C1 P (Lj,1 , Lj,2 ) = (ℓj,1 , ℓj,2 )|j ∈ C2 .
Y
Y

j∈C1 j∈C2

Now, using that the clauses are uniformly distributed on D along with the definitions of the sets
Ak (n, f ), k ∈ {1, 2} we first get for j ∈ C1

P Lj,1 = ℓj,1 j ∈ C1 = P (Lj,1 , Lj,2 ) = (ℓj,1 , ℓj,2 ) j ∈ C1

X

ℓj,2 ∈−L
1
P (Lj,1 , Lj,2 ) = (ℓj,1 , ℓj,2 ) 22 (n
2) 1


=
=l· = n−f
,
X

ℓj,2 ∈−L
P (Lj,1 , Lj,2 ) ∈ A1 (n, f ) p1 (n, f ) 2 1

and next for j ∈ C2

1
P (Lj,1 , Lj,2 ) = (ℓj,1 , ℓj,2 ) 22 (n
2) 1


P (Lj,1 , Lj,2 ) = (ℓj,1 , ℓj,2 ) j ∈ C2 =
= =



.
P (Lj,1 , Lj,2 ) ∈ A2 (n, l) p2 (n, f ) 2 2
2 n−f

42
Inserting this in (4.34) gives
M1 M2
1 1
     
P Φ1 = (ℓj,1 ) C = and P Φ2 = (ℓj,1 ∨ ℓj,2 ) C =
^ ^
.
j∈C1 2 n−f

1 j∈C2 22 n−f

2

Thus for φ1 a 1-SAT formula with n − l variables and M1 clauses and φ2 a 2-SAT formula with
n − l variables and M2 clauses we get
M1
1

P Φ1 = φ1 M = E P Φ1 = φ1 C M =



,
2 n−l

1
M2
1

P Φ2 = φ2 M = E P Φ2 = φ2 C M =



,
22 n−l

2

and this corresponds to having Φk |M ∼ Fk (n − l, Mk ) for k ∈ {1, 2}. Repeating this argument with
equation (4.35) gives that

P Φ1 = φ1 , Φ2 = φ2 M = P Φ1 = φ1 M P Φ2 = φ2 M

which implies the conditional independence.

References
[Ach00] Dimitris Achlioptas. Setting 2 variables at a time yields a new lower bound for
random 3-SAT. In Proceedings of the thirty-second annual ACM symposium on
Theory of computing, pages 28–37, 2000.

[ACIM01] Dimitris Achlioptas, Arthur Chtcherba, Gabriel Istrate, and Cristopher Moore. The
phase transition in 1-in-k SAT and NAE 3-SAT. In Proceedings of the twelfth annual
ACM-SIAM symposium on Discrete algorithms, pages 721–722, 2001.

[ACOHK+ 21] Dimitris Achlioptas, Amin Coja-Oghlan, Max Hahn-Klimroth, Joon Lee, Noëla
Müller, Manuel Penschuck, and Guangyan Zhou. The number of satisfying assign-
ments of random 2-SAT formulas. Random structures & algorithms, 58(4):609–647,
2021.

[AKKK01] Dimitris Achlioptas, Lefteris M Kirousis, Evangelos Kranakis, and Danny Krizanc.
Rigorous results for random (2+ p)-SAT. Theoretical Computer Science, 265(1-
2):109–129, 2001.

[ANN04] Krishna B Athreya, Peter E Ney, and PE Ney. Branching processes. Courier Cor-
poration, 2004.

[BBC+ 01] Béla Bollobás, Christian Borgs, Jennifer T Chayes, Jeong Han Kim, and David B
Wilson. The scaling window of the 2-SAT transition. Random Structures & Algo-
rithms, 18(3):201–256, 2001.

[BCM02] Giulio Biroli, Simona Cocco, and Rémi Monasson. Phase transitions and com-
plexity in computer science: an overview of the statistical physics approach to the
random satisfiability problem. Physica A: Statistical Mechanics and its Applications,
306:381–394, 2002.

43
[BOOS25] Andreas Basse-O’Connor, Tobias Lindhardt Overgaard, and Mette Skjøtt. On the
regularity of random 2-SAT and 3-SAT. arXiv:2504.11979 [math.PR], 2025.

[CCOM+ 24] Arnab Chatterjee, Amin Coja-Oghlan, Noela Müller, Connor Riddlesden, Maurice
Rolvien, Pavel Zakharov, and Haodong Zhu. The number of random 2-SAT solutions
is asymptotically log-normal. arXiv preprint arXiv:2405.03302, 2024.

[CF86] Ming-Te Chao and John Franco. Probabilistic analysis of two heuristics for the
3-satisfiability problem. SIAM Journal on Computing, 15(4):1106–1118, 1986.

[CKT+ 91] Peter C Cheeseman, Bob Kanefsky, William M Taylor, et al. Where the really hard
problems are. In Ijcai, volume 91, pages 331–337, 1991.

[Coo71] Stephen A Cook. The complexity of theorem-proving procedures. In Proceedings

of the Third Annual ACM Symposium on Theory of Computing, STOC ’71, page
151–158. Association for Computing Machinery, New York, NY, USA, 1971.

[CR92] Vašek Chvátal and Bruce Reed. Mick gets some (the odds are on his
side)(satisfiability). In Proceedings of the forty-seventh annual ACM symposium
on Theory of computing, pages 620–627. IEEE Computer Society, 1992.

[DLL62] Martin Davis, George Logemann, and Donald Loveland. A machine program for
theorem-proving. Communications of the ACM, 5(7):394–397, 1962.

[dLV01] W Fernandez de La Vega. Random 2-SAT: results and problems. Theoretical com-
puter science, 265(1-2):131–146, 2001.

[DM11] Claude Dellacherie and P-A Meyer. Probabilities and potential, c: potential theory
for discrete and continuous semigroups. Elsevier, 2011.

[DSS15] Jian Ding, Allan Sly, and Nike Sun. Proof of the satisfiability conjecture for large k.
In Proceedings of the forty-seventh annual ACM symposium on Theory of computing,
pages 59–68, 2015.

[GGW06] Aarti Gupta, Malay K Ganai, and Chao Wang. SAT-based verification methods and
applications in hardware verification. In International School on Formal Methods
for the Design of Computer, Communication and Software Systems, pages 108–143.
Springer, 2006.

[Goe96] Andreas Goerdt. A threshold for unsatisfiability. Journal of Computer and System
Sciences, 53(3):469–486, 1996.

[Gol79] Allen T Goldberg. On the complexity of the satisfiability problem. New York Uni-
versity, 1979.

[GW94] Ian P Gent and Toby Walsh. The SAT phase transition. In ECAI, volume 94, pages
105–109. PITMAN, 1994.

[Knu15] Donald E Knuth. The art of computer programming, Volume 4, Fascicle 6: Satisfi-
ability. Addison-Wesley Professional, 2015.

[KS94] Scott Kirkpatrick and Bart Selman. Critical behavior in the satisfiability of random
boolean expressions. Science, 264(5163):1297–1301, 1994.

44
[MS08] Joao Marques-Silva. Practical applications of boolean satisfiability. In 2008 9th
International Workshop on Discrete Event Systems, pages 74–80. IEEE, 2008.

[Pap03] Christos H Papadimitriou. Computational complexity. In Encyclopedia of computer

science, pages 260–265. John Wiley and Sons Ltd., 2003.

[SML96] Bart Selman, David G Mitchell, and Hector J Levesque. Generating hard satisfia-
bility problems. Artificial intelligence, 81(1-2):17–29, 1996.

45