Kolmogorov Complexity and Metacomplexity

Problems
Aniruddh Pramod Supervisor: Dr. Satyadev Nandakumar
Mathematics and Scientific Computing Computer Science and Engineering
IIT Kanpur IIT Kanpur
[email protected] [email protected]

Abstract—In this report, we will discuss some key papers from Cryptomania to Heuristica are known, however the other
in computational complexity to understand the main proof direction of these implications are all open questions. The
techniques that have been used to make progress in the subject establishment of each of these will exclude a world from the
of metacomplexity. Starting from Natural Proofs, which work by
identifying a notion of variation that is high in SAT but low for five possibilities.
polynomial circuits, we then move on to Black Box Generators,
which have been the most important tool of the past decade.
We also discuss a non-black-box worst-to-average-case reduction
proposed by Hirahara, which shows promise in overcoming
the barriers that the previous techniques were running into.
We pay special attention to Kolmogorov Complexity techniques
that have been applied to these problems. By studying these
results, we would like to illuminate the complex interplay between
computational complexity, randomness, and circuit lower bounds.

Index Terms—natural proofs, circuit complexity, learning,

minimum circuit size problem, Kolmogorov complexity, non-
black-box reduction

I. I NTRODUCTION
The main discussion in this report is centered around 4 Fig. 1. Illustration of Impagliazzo’s Five Worlds
influential results in the field of metacomplexity. The first is
a natural proof by Sipser [1] which establishes that P arity Organization: Section II reviews background and proof
cannot be computed in AC 0 circuits. We then move on to techniques that are used in the report. In section III, we look
black box generator proofs. We cover Carmosino’s work [4] at natural proofs. In section IV we talk about proofs via black
that establishes a learning algorithm from a natural proof and box generators and finally in section V we look at a non-black-
Hirahara’s proof [3] on the hardness of partial function variants box reduction. In addition to section II, every section will start
of the Minimum Circuit Size Problem - MCSP. Finally we with an introduction to the key concepts and techniques used
cover [5], Hirahara’s novel non-black-box worst to average in that section.
case reduction for MINKT. II. BACKGROUND
Impagliazzo [6] studied the interplay between pseudoran- We define key notation and terminology in this section and
domness and metacomplexity and established that there are shed light on some essential techinques we will be using in
five possible worlds that could correspond to ours. The first the main proofs.
of these is Algorithmica, where NP is easy in the worst case
and P = NP. The second is Heuristica, where NP is hard A. Notation
in the worst case but easy on average case, ie. P ̸= NP and For an integer n, we define [n] = {1, . . . , n}. For a finite
DistNP ⊆ AvgP. The third is Pessiland, where even though set D we will use x ∼ D to indicate that x is randomly
NP is hard on average, there are no one-way functions. The sampled from the distribution D or x ∈R X to indicate that
fourth is Minicrypt, where one-way functions exist but public x is uniformly randomly picked from the set X.
key cryptography does not, and finally we have Cryptomania We also use ≲ and ≳ to represent an approximate inequality,
where we can have public key cryptography. Fig. 1 provides in particular, a ≲ b is the same as a ≤ (1 + o(1)) · b, where
a picture to better illustrate this. o(1) approaches 0 for suitable choice of constraints.
Proving the inexistence of these worlds, in particular Pes- For a function f , we will denote its truth table tt(f ) =
siland and Heuristica is where much of the attention of f (z1 ) . . . f (z2l ), with zi ∈ {0, 1}l being lexicographically
compelxity theorists is devoted to. One way implications ordered l-bit strings.
 For a language L ⊆ {0, 1}∗ , the characteristic function of such that U A (d) outputs tt(g) of length 2l in t steps and
the language is L : {0, 1}∗ → {0, 1} such that L(x) = 1 iff dist(f, g) ≤ 0.5 − δ.
x ∈ L for every string x. D EFINITION 2.6: (KT Complexity). For any oracle A ⊆
Promise Problems: A pair (LY , LN ) of languages such that {0, 1}∗ , the KT Complexity of x relative to A is defined as
LY ∩ LN = ∅. These are called the YES and NO instances KT A (x) := min{|d| + t : U A (d) = x in t steps}.
respectively. A language A such that LY ⊆ A ⊆ {0, 1}∗ \ LN Problems formulations on Kolmogorov Complexity: We are
is called a solution to the promise problem. A promise problem interested in MKTP and MINKT, which are formulated thusly,
where the ‘promise’ is the entire language is called a decision D EFINITION 2.7: For an oracle A, we define MKTPA :=
problem. {(x, 1s ) : KT A (x) ≤ s} .
Gap Problems: A Gap problem effectively turns an approx- D EFINITION 2.8: For any oracle A, define MINKTA :=
imation task into a promise problem. We set two thresholds {(x, 1t , 1s ) : KtA (x) ≤ s}.
for the approximation. The YES instances are the ones where It is easy to check that MINKT ∈ NP since we can guess a
the quantity is above the higher threshold, the NO instances certificate of length atmost s and check if U (d) outputs x in
are the ones where the quantity is below the lower threshold. t steps.
The task is to decide, given an input, if it is a YES or NO
D EFINITION 2.9: (Certificate) A string d is called a certifi-
instance.
cate for KtA (x) ≺ s if U A (d) outputs x within t steps and
Circuits: For a boolean circuit C, we represent its size by |d| ≤ s.
|C|. We assume size is measured by gates, although the exact
We also define the promise and search versions of MINKT,
method of measuring of gates is not really important and
equivalence can be shown at the cost of small changes to the D EFINITION 2.10: (Promise-MINKT) Let σ, τ : N × N be
bounds. functions such that σ(n, s) ≥ s and τ (n, t) ≥ t for n, s, t ∈ N.
Then Gapσ,τ MINKT is the promise problem defined as -
AC Circuits: We focus on proofs involving a class of
circuits called AC circuits. These are circuits which can • YES instances: (x, 1t , 1s ) such that Kt (x) ≤ s
use a polynomial number of unlimited fan-in AND and OR • NO instances: (x, 1t , 1s ) such that Kt′ (x) > σ(|x|, s) for
gates. AC 0 is the class of AC circuits with constant depth. t′ := τ (|x|, t)
Additionally, we’ll use AC 0 [p] for a prime p when we want D EFINITION 2.11: (Search-MINKT) For functions σ, τ de-
to include MOD p gates. fined in 2.10, the search version of Gapσ,τ MINKT is defined
Additionally in the context of a circuit, an d-circuit is a as -
circuit of depth d. Similarly a i-parity circuit is a circuit that
• Inputs: A binary string x and integer t in unary.
computes the parity function on i variables.
• Output: A certificate for Kt′ (x) ≺ σ(|x|, Kt (x)) for any
B. Pseudorandom Generators t′ ≥ τ (|x|, t).

D EFINITION 2.1: ((l, n, d)-Design): A family of n-sets

D. Other Important Problems
{Ii }m
i=1 ⊆ {1, . . . , l} such that |Ij ∩ Ik | ≤ d.
D EFINITION 2.2: (Nisan-Widgerson generator) Given a D EFINITION 2.12: (Boolean Satisfiability Problem - SAT)
function f and a (l, n, d)-design {Ii }m i=1 , the NW Generator Given a boolean expression, built from variables, the operators
GfN W (z) = f (zI1 )f (zI2 ) . . . f (zIm ). AND, OR and NOT, and parantheses, check whether there is
D EFINITION 2.3: (Distinguisher) A circuit D is an ϵ- an assignment that satisfies it.
distinguisher for a PRG if we have E|P r[D(P RG(x)) = 3-SAT in particular is known to be NP-Complete and many
1]−P r[D(w) = 1]| ≥ ϵ, where x and w are randomly sampled initial proofs used this as the standard example for a hard
from appropriate distributions. problem. However, as [12] suggests, MCSP has become a
more attractive starting point for most complexity theory
C. Kolmogorov Complexity proofs now.
We fix an efficient universal Turing Machine U that takes D EFINITION 2.13: (Minimum Circuit Size Problem -
n
as input the description of a Turing Machine M and simulates MCSP). Given a truth table tt(f ) ∈ {0, 1}2 and a size
it on input x. Now, we define: parameter s, is there a circuit on n-variables of size atmost s
D EFINITION 2.4: (Time-bounded Kolmogorov complexity). which agrees with tt(f ).
For any oracle A ⊆ {0, 1}∗ and integer t ∈ N, the Kolmogorov D EFINITION 2.14: (Decision Version of Occam’s Learning
complexity of x within time t relative to A is defined as - MINLT). Given a set of pairs {(xi , bi )}m i=1 , where xi ∈
KtA (x) := min{|d| : U A (d) = x in t steps}. {0, 1}n and bi ∈ {0, 1} a size constraint s and a time constraint
D EFINITION 2.5: (Approximation version of Time-bounded t, is there a t-time program M of size s such that M (xi ) = bi
Kolmogorov complexity). For functions f, g : {0, 1}l → for every i ∈ [n]
{0, 1}, define dist(f, g) := P rx∈R {0,1}l [f (x) ̸= g(x)]. For The relationship between number of pairs n and the size
a function f : {0, 1}l → {0, 1}, an integer t and an oracle of the the program, s determines the kind of problem we are
A
A, define Kt,δ (f ) as the minimum length of a string d solving. Broadly these refer to a task called PAC Learning
 III. NATURAL P ROOFS compute m-parity functions and have 2-circuits of size atmost
A. Key Concepts bc · 2bc .
Finally, these parity d-circuits can be converted to parity
First let us define the notion of a natural property - (d − 1)-circuits by rewriting the a-size 2-circuits using dis-
D EFINITION 3.1: (γ-avoids). A promise problem γ-avoids tributive law as AC circuits of size a·2a . Also since d ≥ 3, we
a circuit class G if ∀n, z, Gn (z) ∈ LN and P r[w ∈ LY ] ≥ can merge adjacent levels of OR gates and apply DeMorgan’s
γ(n). Additionally, we say that G is a HSG secure against a laws to swap AND and OR gates to make adjacent levels
complexity class C if there is no promise-problem that avoids similar. Thus at the cost of a constant factor we can create a
G. series of parity (d − 1)-circuits. Which is impossible due to
D EFINITION 3.2: (Γ-Natural Property). A promise problem the minimality of d. ■
(LN , LY ) is called a Γ-natural property useful against circuits This result will have consequences for us later when we
of size s() on l inputs with largeness γ if it γ-avoids all such try to create learning algorithms from natural proofs, since we
circuits and membership of the property is computable in Γ. will use MOD p gates and for p > 2, MOD p will not be
A natural proof involves searching for a natural property able to compute parity either, and we will need some extra
useful against a class you care about, followed by an inductive transformations to show the result.
argument to show that polynomials cannot compute functions
with a high value of this natural property. A classic NP C. Barrier
problem like SAT is then used to demonstrate that hard Razbarov and Rudich showed that no natural proof could
problems have a high value of the property and thus the proof prove P ̸= NP. This proof relies on the conjecture that
follows. pseudorandom functions exist, which is widely believed to be
true. This has greatly discouraged the pursuit of a natural proof
B. Application
for the problem since even if we could do it, we would end
We examine [1], which proves that PARITY cannot be up disproving the existence of pseudorandom functions and
computed in AC 0 circuits. Formally stating, crush the backbone of modern cryptography.
T HEOREM 3.1: Parity cannot be computed by constant-
depth, polynomial size circuits, ie. AC 0 circuits. □ IV. B LACKBOX G ENERATOR P ROOFS
Proof Sketch: We use a result by Lupanov [8] which states Taking the lesson from pseudorandomness, we look at a
that parity 2-circuits are exponentially large. Let d ≥ 3 be the proof technique that is closed related to PRGs. In a blackbox
minimal depth that admits polysize parity d-circuits. We will generator proof we often rely on the construction of a blackbox
contradict the minimality of d. generator from a hard function and then using reconstruction
Consider a sequence of parity d-circuits {Cn }n=1 where algorithms to show a reduction. Recent work also typically
Cn computes n-parity and has size ≤ nk . We claim that there focuses on proving the hardness of variants of metacomplexity
is a restriction rho that induces Cnρ such that, problems. The first work we look at will be a reduction from
√ a Natural Proof to a Learning Algorithm [4] and the other will
(i) Cnρ computes an m-parity for some m ≥ n/2
(ii) The size |Cnρ | ≤ nk is polynomial in m be the proof of hardness of a partial function variant of MCSP
(iii) All the 1-circuits of Cnρ are bounded in size, indepen- [3].
dently of n
A. Notions
ρ : X n → {0, 1, ∗} is chosen from a distribution that inde-
We need some hardness amplification tools. Typically these
pendently assigns, for each i, the probabilities
√ - P r[ρ(xi ) =
are the Direct Product (DP) Construction and the XOR Con-
∗] = √1n and P r[ρ(xi ) ̸= ∗] = 1−1/ 2
n
. Using Chebyshev’s
struction.
inequality we can prove that the failure probability is less than
D EFINITION 4.1: (K-wise Direct Product). For a boolean
1.
function f on n-bit strings and a parameter k ∈ N, the k-wise
Next use the sequence of induced circuits {Dn = Cnρ }, and
Direct Product of f is f k (x1 , . . . , xk ) = (f (x1 ), . . . , f (xk )).
show that there is another restriction which induces Dnρ , such
This is a hardness amplification for f in the sense that
that
√ we can use an approximation circuit for f k to create an
(i) Dnρ computes an m-parity for some m ≥ n/2 approximation circuit for f efficiently. The algorithm for the
(ii) All the 2-circuits of Dnρ are bounded in size, indepen- same is discussed later.
dently of n We use MODp gates as a replacement for XOR, since
If we fix a constant bc , ρ fails if ρ assigns too few ∗’s AC 0 [p] circuits cannot compute parity.
(unlikely) or if some 2-circuit of Dnρ depends on ≥ bc inputs. D EFINITION 4.2: (Black Box (ϵ, L)-Generator within Λ).
By induction we can show that for every c there is a bc such Given an error parameter ϵ : N → [0, 1] and a stretch
that the probability of failure is O(n−k ). function L : N → N, a map GEN from a n-var function
But now all 2-circuits depend on bc inputs can be converted to a family of l-var functions GEN (f ) = {gz }z∈{0,1}m with
into one which has size atmost bc · 2bc . Use this fact to convert l = logL(n) satisfying the smallness, non-uniform efficiency
the restrictions of Dn to a sequence of circuits En , which and reconstruction properties -
 •Smallness: m ≤ poly(n, 1/ϵ) C ONSTRUCTION S TEP: Taking input n-bit string x,
•Non-Uniform Λ-Efficiency: gz ∈ Λf [poly(m)] 1) Assign z|Sj of Gf to x to get m-bit z
• Reconstruction: Using a distinguisher for gz we can 2) For 1 ≤ j < i, fix j th input of D to wj = f (z|Sj ) via
create a lossy approximation circuit for f in polynomial table lookup
time using a randomized algorithm with oracle access to 3) If D(w1 , . . . , wL ) = 1, output wi else output 1 − wi
f Repeat this algorithm poly(L) times and estimate with
D EFINITION 4.3: (Black Box (ϵ, δ)-Amplification within membership queries the agreement between the constructed
Λ). A map from f → AM P (f ) such that - circuit and f to get a good circuit.
′
• Short input: Input length of AM P (f ), n ≤ A LGORITHM 4.2: (Direct Product Reconstruction). There is
poly(n, 1/ϵ, log 1/δ) a constant c such that given a circuit C ′ that approximates f k ,
f ′
• Non-uniform Efficiency: AM P (f ) ∈ Λ [poly(n )] the direct product of f , to 1 − δ, we can make a circuit C that
f
• Uniform P-Efficiency: AM P (f ) ∈ P approximates f to ϵ where, δ > e−ϵk/c .
• Reconstruction: Given an approximation circuit for the P REPROCESSING S TEP:
AM P (f ) we can create an approximation circuit for f 1) Pick a set B0 of k n-bit strings.
in polynomial time, given oracle access to f 2) Pick A ⊂ B0 of size k/2.
Where we denote by Λf the class of oracle circuits in Λ 3) Evaluate C ′ on k-tuple b⃗0 , a random permutation of
which have f -oracle gates and Λ[s] denotes circuits in Λ with strings in B0 .
maximum size s. Additionally one should note that the exis- 4) Note answers a⃗0 for strings in A.
tence of blackbox amplification in a circuit class is non-trivial. C ONSTRUCTION S TEP: On input x, if x ∈ A then return the
Indeed, showing the existence of blackbox amplification in corresponding answer from ⃗a. Else for m = O((log1/ϵ)/δ)
AC 0 [p] is one of the main theorems we need to prove. times,
D EFINITION 4.4: (Goldreich-Levin (GL) Construction). For 1) Sample a random k-set B such that A ∪ {x} ⊂ B.
a function g : {0, 1}m → {0, 1}k , we define: g GL : {0, 1}m ×
2) Evaluate C ′ on k-tuple ⃗b, a random permutation of
F k → F as -
strings in B.
k
X 3) if C ′ (⃗b) for A are consistent with ⃗a then output C ′ (⃗b)x
g GL (x1 , . . . , xm , r1 , . . . , rk ) = ri · g(x1 , . . . xm )i and stop.
i=1
. If there is no output after m iterations, output a random bit.
D EFINITION 4.5: (von Neumann trick function). For an The circuit we construct here is randomized, however we
integer parameter t > 0, define E vN : (F 2 )t → {0, 1}. For can specify the randomness in the preprocessing stage to make
pairs (a1 , b1 ), . . . , (at , bt ) ∈ F × F , set it deterministic.
A LGORITHM 4.3: (Goldreich-Levin Reconstruction). Given
E vN ((a1 , b1 ), . . . , (at , bt ))
 an arbitrary h ∈ F k and B : F k → F such that P r[B(r) =
1, if ai = bi ∀i ∈ [t]
 ⟨h, r⟩] ≥ 1/p + γ. Given oracle access to B and a parameter
= 1, if ai > bi for the first unequal pair γ, there is an algorithm that runs in poly(k, 1/γ) time and
 outputs a list of size O(1/γ 2 ) with probability > 1/2 that h
0, if ai < bi for the first unequal pair

is on the list.
B. Key Techniques Proceed in k rounds, maintaining after round i, a list Hi of
The key technique that we will rely on is the concept of a length-i tuples in F i
reconstruction algorithm. 1) Extend each tuple by one element in all |F | ways
Firstly, given a Nisan-Widgerson generator and a distin- 2) For each extended ⃗c ∈ F i , include ⃗c in Hi ⇐⇒ it
guisher against it with advantage δ, we can reconstruct the satisfies the TEST subroutine.
function f . TEST SUBROUTINE
A LGORITHM 4.1: (NW Generator Reconstruction). Given • Randomly pick m = poly(k/γ) tuples, s1 , . . . sm ∈
a distinguisher D for a NW Generator Gf : {0, 1}m → F k−i .
{0, 1}L based on a n-var boolean function f using a design • For each s ⃗i and σ ∈ F , estimate P r[B(⃗r, ⃗s) = ⟨⃗c, ⃗r⟩+σ]
S1 , . . . , SL ⊆ [m]. The goal is to make a circuit C that 1
• If atleast one estimate is ≫ p , accept, else reject
computes f on 1/2+Ω(1/L) fraction of inputs with probabilty
1/poly(L). C. Learning Algorithms from Natural Proofs
P REPROCESSING S TEP: Carmosino [4] uses a Blackbox generator technique to prove
1) Pick a random i ∈ [L] the following main theorem
2) For i ≤ j ≤ L, fix j th input to D to a rand bit wj T HEOREM 4.1: For every prime p ≥ 2 there is a randomized
3) For j ∈ [m]/Si , fix j th input to Gf to a rand bit zj algorithm that takes membership queries for an arbitrary n-
4) For 1 ≤ j < i, enumerate all possible x ∈ {0, 1}n variate Boolean function f ∈ AC 0 [p], runs in quasipolytime,
1
consistent with z|Sj , query f (x) and build table T of ie. npoly log n and finds a circuit computing f on all but poly
pairs (x, f (x)) fraction of input. □
 This learning algorithm can also be converted into a com- D. NP Hardness of MINLT
pression algorithm. We focus on the proof for the learning We now look at a proof that leans more into the pseudo-
algorithm. randomness aspect and uses a secret sharing scheme. In [3],
T HEOREM 4.2: Blackbox amplification is present in Hirahara proves the NP Hardness of MINLT and partial MCSP
AC 0 [p]. □ by reducing it to Minimum Monotone Satisfying Assignment
Proof : We first look at the proof for the case where p = 2. (MMSA). We focus on the first of these, ie. the hardness of
Here, we can compute XOR, whereas we won’t be able to learning programs.
do so when p > 3. Given a function f ∈ AC 0 [2] of size s, T HEOREM 4.3: Under randomized polynomial-time one-
set g to be the k-wise direct product of f , ie. g = f k where query reductions, it is NP-hard to distinguish the following
k = ⌈3c · 1/ϵ · ln 1/γ⌉ + 1. Set AM P (f ) = g GL , the GL cases, given a size parameter s ∈ N and a distribution E with
construction on g, over F = GF (2). This is our blackbox supp(E) ⊆ {0, 1}n × {0, 1} as input for some n ∈ N.
amplification candidate.
YES : There is a polynomial time program M of size s such
The first 3 properties are easy to verify. We just need to
that P r[M (x) = b] = 1, moreover, M computes a linear
prove the reconstruction algorithm. So, given a circuit that
function over GF (2)
approximates g GL , we need to build a circuit to approximate g.
NO : For any program M of size s · nϵ , P r[M (x) = b] ≤
Reconstruction of f from g is done by the DP Reconstruction 1 −n1−δ
algorithm 4.2. Let AGL represent the GL Reconstruction 2 +2
Algorithm from 4.3. Consider the following algorithm - where, (x, b) is sampled from E, δ > 0 is an arbitrary constant
For a given input x ∈ {0, 1}nk , define a circuit Bx (r) := and ϵ = 1/(log log n)O(1) . □
′
C (x, r), for r ∈ {0, 1}k . Run AGL on Bx with parameter γ/2 If we were to draw m samples from E, then this problem is
to get a list L of k-bit strings. Output a uniformly random k-bit reduced to MINLT.
string from L. Ko [11] gave evidence that the NP-hardness of MINLT is
The correctness of this algorithm is easy to verify and fol- subject to a relativization barrier. Most complexity-theoretic
lows from the correctness of the GL Reconstruction algorithm proofs tend to be relativizing in nature, and in this work,
itself. Hirahara overcomes this relativization barrier with a blackbox
For the case where p > 2, we need to put in some additional proof.
work since we can’t use a XOR construction, because parity is Proof of Theorem 4.3: We start by defining the MMSA
not computable by small AC 0 [p] circuits. The main struggle problem.
is the fact that g GL is no longer boolean valued and hence D EFINITION 4.6: (Minimum Monotone Satisfying Assign-
can’t be plugged into an NW generator. To convert it back ment) Given a monotone formula φ on n-variables and a
into a boolean value, we need to use the von Neumann trick constraint P θ ∈ N. Find a satisfying assignment α ∈ {0, 1}n
function on g GL . such that αi ≤ θ.
It is known that approximating MMSA to a factor of
hvN ((a1 , b1 ), . . . , (at , bt )) n1/(log log n)
O(1)
is NP-hard as discussed in [10]. So if we can
= E vN ((g GL (a1 , b1 )), . . . , (g GL (at , bt ))) show the reduction then we are done.
Use a secret sharing scheme (Share(φ, −), Rec(φ, −)) for
Proof of the reconstruction property is in [4]. ■ a monotone formula φ. We say that a subset T ⊆ [n] is
Proof of Theorem 4.1: We are given a P -natural property authorized if the characteristic function of T satisfies φ. We
against Λ[u]. We need to show that there is a randomized al- can now share a secret b among n parties which will form our
gorithm which, given oracle access to f from Λ[sf ], produces distribution E.
−1
a ϵ-approximation circuit for f in poly(n, 1/ϵ, 2u (.) ) time. To sample (x, b) from E we share b as Share(φ, b) =
We start by designing the (ϵ, L)-blackbox generator (s1 , . . . , sn ) and hide each share si in the input x in a way
GEN (f ) with log L(n) > u−1 (poly(n, . . . )). By the non- such that a large program will be able to read many hidden
uniform efficiency of black box generation we have gz ∈ shares whereas a small one will not. This will be accomplished
Λf [poly(n, 1/ϵ)]. with a blackbox generator, in particular we will use a slight
If we further replace the f -oracle with a Λ-circuit for f , variation of the k-wise DP generator that we refer to as the
then we have gz ∈ Λ[sg ]. DPk+ .
Now we need to restrict sg < u(log L(n). Let D be the D EFINITION 4.7: (Extended Direct Product Generator). The
circuit from the natural property restricted to truth tables of map DPk+ : {0, 1}λ × {0, 1}λ×k → {0, 1}λk+k which takes
size L(n). a string f ∈ {0, 1}λ and a λ × k matrix z, both over GF (2),
• By usefulness, P rz [¬D(gz ) = 1] = 1 and returns DPk+ (f, z) = (z, f · z).
• By largeness, P ry [¬D(y) = 1] ≤ 1 − 0.2 As with the k-wise direct product, we have a reconstruction
• So ¬D is a distinguisher for GEN (f ) property -
And now that we have a distinguisher, we can use it to apply D EFINITION 4.8: (DPk+ Reconstruction). If there is a ϵ-
the reconstruction property to get an approximation circuit for distinguisher D which can distinguish between the output
f in polynomial time. ■ distribution of DPk+ (f, −) and the uniform distribution, ie.
|P rz∈R U [D(DPk+ (f, z)) = 1] − P rw∈R U [D(w) = 1]| ≥ ϵ. fi . We can show that |B| is small,
Then we have K D (f ) ≤ k + O(log(λk/ϵ)).
|B| · λ − |M |
This construction is quite simple (in fact it is linear for a
≲ K(fB ) − |M | (by the assumption)
fixed f ), and still has all the properties we would like. If D
does not know f , in the sense that K D (f ) ≫ k then DPk+ is ≲ K(fB |M ) (by definition of K(x|y))
X
secure against it. ≲ K(fi |M ) (describe fB description of allfi ’s)
Now consider an instance (φ, θ) of MMSA and pick λ-bit i

strings f1 , . . . fn uniformly at random. We can associate each ≤ |B| · θ (by definition of B)

fi with the i-th variable of φ. Now define E = E(f1 , . . . , fn )
in the following way -
If we take large enough λ we can have |B| ≲ |M |/λ as
To sample (x, b), first pick b randomly and distribute it required. Now we must prove that M cannot distinguish Xi
among n parties using the scheme for φ. Let there be n shares from Ui ∀ i ∈ [n]\B, which can be done by a hybrid argument.
si and let |si | = k. Now we want to hide these in x such that We demonstrate this argument for n = 2, B = {2}. Our goal
only programs that know fi can read the i-th share. To that is then to show
end, set ξi = (zi , (fi · zi ) ⊕ si ) = DPk (fi , zi ) ⊕ (0λk , si ) for
i ∈ [n]. Then x := (ξ1 , . . . , ξn ) is the required sample. P rz1 ,z2 [M (DPk (f1 , z1 ), DPk (f2 , z2 )) = 1]
This completes our description of the distribution, now we ≈ P rU1 ,z2 [M (U1 , DPk (f2 , z2 )) = 1]
just need to prove its correctness.
Suppose not, then we can show that K(f1 |M ) ≤ theta, which
C OMPLETENESS
P : Let’s say that there is an α that satisfies will contradict the fact that M does not know f1 . We do this
φ, with αi ≤ θ. Let T be set of indices such that αi = 1, by using DP reconstruction.
so T is an authorized subset. If we apply it naively, we get K D (f1 ) ≲ k where D outputs
Now we write the program M . Hardwire the values of {fi : M (w, DPk (f2 , z2 )) on input w, z2 . Now D can be simulated
i ∈ T }. Now on input x, let (zi , ηi ) := ξi , si := (fi , zi ) ⊕ ηi with M and f2 , so we have,
and reconstruct the secret b using Rec(φ, −) for {si : i ∈ T }. K(f1 |M, f2 ) ≲ K D (f1 ) ≲ k
Output b. Since T is authorized, b is reconstructed correctly.
Also, |M | is bounded by the hardwired strings, so |M | ≤ To remove f2 , we need to use the concept of Trevisan-
P Vadhan advice from [9]. Observe that D is a randomized
i∈T |fi | + O(1) = |T | · λ ≤ θλ.
algorithm taking an a-bit advice string. Specifically here,
S OUNDNESS: First look at the success condition for the
D takes w and random bits z2 as inputs and outputs
reduction. We need that K(fT ) ≳ |T | · λ for every T . But
M (w, DP (f2 , z2 )) = M (w, (z2 , f2 ·z2 )). But we can compute
this happens with high probability for a random choice of
this with k-bit advice string f2 · z2 which is smaller than
f1 , . . . fn .
f2 . Using this fact, we obtain K(f1 |M ) ≲ k + a, where
If there is no authorized set T of size θ then no program M a ≤ (n − 1)k. So we can show,
of size θλ/2 can output b on input x with high probability over
θ
(x, b) ∼ E. This is formulated and proved as the Algorithmic K(f1 |M ) ≲ nk ≤
2
Information Extraction Theorem below.
But this implies that M knows f1 which is the contradiction
Hence, we have demonstrated the correctness of the reduc- we were trying to show. ■
tion and we are done. ■ We can see that this proof technique is algorithmic in-
T HEOREM 4.4: (Algorithmic Information Extraction). Let formation theoretic, and could be applied to other reduction
fi ’s be λ-strings such that K(fT ) ≳ |T | · λ for every T ⊆ [n] problems in the future. Also, it does not exploit the efficiency
where λ is sufficiently large. Let M be a program. Then there of the program for the NO case, and hence is blackbox in
is a set B ⊆ [n] such that |B| ≲ |M |/λ and nature.
E. Barriers
P r[M (XB , U[n]\B ) = 1] ≈ P r[M (XB , X[n]\B ) = 1] Bogdanov [7] showed that if L reduces to a DistNP problem
via blackbox nonadaptive randomized polytime reduction, then
L ∈ NP/poly ∩ coNP/poly. This means that we cannot use this
where Xi is a random variable identical to DPk (fi , zi ) for a technique for problems outside NP ∩ coNP.
random choice of zi and Ui is uniform distribution. □ This however isn’t as serious a roadblock as it might appear.
The previous proof is blackbox but non-relativizing, and in the
Proof of Theorem 4.4: Firstly, we define that M knows fi next section we will study a proof that is non-blackbox but
if K(fi |M ) ≤ θ := 2nk, since if M knows fi it should be relativizing. Combining the two proof techniques could pave
easy to describe fi with it. the way for a non-blackbox non-relativizing proof that can
Now let B be the set of indices i ∈ [n] such that M knows overcome both barriers simulataneously.
 V. T OWARDS N ON -B LACKBOX R EDUCTION • Output Tb (f (zS1 ) . . . f (zSi−1 .c.wi+1 . . . wm ) ⊕ c ⊕ 1.
Hirahara [5] overcomes the barrier mentioned previously • If c = f (zSi ) = f (x) then we get Hi , so Tb will likely
and gives a non-blackbox worst-to-average-case reduction for output 1, so we output c to predict f . So this circuit
a problem believed to be outside NP ∩ coNP. computes f with probability 0.5 + δ/m.
T HEOREM 5.1: In the following, we have 1 =⇒ 2 =⇒ We can show that P Tb has aPsmall description. Hardwiring
3 =⇒ 4, and additionally if P romise−ZP P = P romise− the required values of f uses j<i 2|Si ∩Sj | bits. With oracle
P then we also have 4 =⇒ 2. access to T we can describe the ( 12 + m δ
)-fraction of the truth
1 DistN P ⊆ AvgP table of f by specifying the parametersP involved.
2 (M IN KT [r], Du ) ∈ AvgP The length of the description will be j<i 2|Si ∩Sj | + (m −
3 There is a zero-error randomized polytime algorithm i) + (d − l) + O(log(md)) ≤ exp l2 /d · m + d + O(log(md)).
solving search version of Gapσ,τ M IN KT This procedure needs time poly(m, d) + poly(2l ). So we have
4 Gapσ,τ M IN KT ∈ P romise − ZP P □ KtT (f ) ≤ exp l2 /d · m + d + O(log(md)).
We focus on the proof of 2 =⇒ 3, which is where the For efficiency, observe that we only need to c ∈R {0, 1},
non-blackbox technique is used. iıR [m], w[m]\[i] ∈R {0, 1}m−i and z[d]\Si ∈R {0, 1}d−l
The key idea here is to note that Kolmogorov complexity randomly. Using a Markov Style argument one can show that
can be used to distinguish true randomness from pseudoran- we get P rx [P Tb (x) = f (x)] ≥ 21 + 2m δ
with probability atleast
dom strings. If we assume that (MINKT[r], Du ) is easy, then δ/2m. Repeat the random choice O(m/δ) times and try both
we can take the set of strings accepted by it for a fixed time values of b to get the required certificate with high probability.
t as a set of true random strings and a test against any PRG. This however is a certificate for approximate Kolmogorov
This oracle of random strings will help us create an algorithm complexity. To convert this into a certificate for KtT (x) we
to give a certificate for GapMINKT. can use the List-decodable error-correcting code to show the
Firstly, however we need to change our understanding of following theorem -
NW generators to a weaker notion. T HEOREM 5.3: KtA′ (x) ≤ Kt,ϵ A
(Encn,ϵ (x)) + O(log(n/ϵ))
for any string x, any oracle A and any t′ ≥ t + poly (n, 1/ϵ).
A. Some Definitions A
Given a certificate for Kt,ϵ (Encn,ϵ (x)) ≺ s, we can find a cer-
A
D EFINITION 5.1: ((l, ρ)-design). A family tificate for Kt′ (x) ≺ s + O(log(n/ϵ)) in time t + poly(n, 1/ϵ)
S = (S1 , . . .P , Sm ) of subsets of [d] is an (l, ρ)-design with oracle access to A. □
i−1
if |Si | = l and j=1 2|Si ∩Sj | + m − i ≤ ρm for every i ∈ [m] Finally we need to remove the oracle T from our expression
The advantage of this weaker notion is that it is more but this is simple. We have a program that on input x outputs a
efficient to construct such a family. short program d that describes x under the oracle T . But since
T HEOREM 5.2: For any (m, l, d) ∈ N such
that d/l ∈ N, T is accepted by a polytime algorithm, namely the errorless
there is a (l, exp(l2 /d)-design Sm,l,d ⊆ [d] that can be heuristic algorithm for (M IN KT [r], Du ) which is given to
l
constructed by a detereministic algorithm in time poly(m, d). us, we can describe x using d and a source code for the
D EFINITION 5.2: (List-decodable error-correcting code). algorithm accepting T .
For every n, m, L ∈ N and n And hence starting with a errorless heuristic algorithm

ϵ > 0, a function Enc : {0, 1} →
m 1
{0, 1} is a L, 2 − ϵ -list decodable error correcting code for (MINKT[r], Du ). We have created a scheme to obtain a
if ∃ Dec : {0, 1}m → ({0, 1}n )L such that ∀ x, r with certificate for GapMINKT. ■
dist(Enc(x), r) ≤ 12 − ϵ we have x ∈ Dec(r).
This will let us convert a certificate for approximate com- C. Evidence for complexity
plexity into one for precise complexity. While it is not yet proven that GapMINKT is outside
N P ∩ coN P , it is believed to be so. There are two main
B. Main Theorem factors contributing to this belief. The property that there is
Proof of Theorem 5.1: Start with an oracle T that accepts a no efficiently computable HSG that is γ-secure against coNP
dense subset of random strings. Given tt(f ) and this oracle, for any admissible γ and the conjecture that there is a hitting
we need to give a certificate for KtT (f ). We are assured that set generator secure against NP.
T is a statistical test against N W f with advantage δ. To make T HEOREM 5.4: Let s(n) = (log n)ω(1) for n ∈ N. Let
things a little less verbose, set Tb (w) = T (w) ⊕ b to ignore σ, τ be functions such that σ(n, s(n) + O(log(n))) ≤ n − 2
absolute values in the statistical test. for any n ∈ N. If Gapσ,τ MKTP ∈ coNP/poly then there is a
We will now use a hybridization argument. Look at the NP/poly-natural property useful against P/poly with largeness
distribution Hi = f (zS1 ) . . . f (zSi )wi+1 . . . wm . Since, H0 = 1
2. □
w and Hm = N W f , we have E[Tb (Hm ) − Tb (H0 )] ≥ δ. For The second shows that Gapσ,τ MINKT is Random-3SAT
δ
any i ∈R [m], we will have Ei [Tb (Hi ) − Tb (Hi−1 )] ≥ m . We hard.
can use this advantage to predict the next bit of the PRG. T HEOREM 5.5: Let σ, τ be functions such that for any con-
Consider the following circuit, P Tb to predict f - stant c0 > 0 and for all large N ∈ N, σ(N, N −c0 N/ log N ) ≤
l
• On input x ∈ {0, 1} , set zSi := x and construct z. N −c1 N/ log N . Then Gapσ,τ MINKT is Random 3SAT-Hard.
 The proof follows by looking at the distributional NP [6] R. Impagliazzo, ”A personal view of average-case complexity,”
version of Random 3-SAT as an HSG. Proceedings of Structure in Complexity Theory. Tenth Annual
IEEE Conference, Minneapolis, MN, USA, 1995, pp. 134-
Finally, we must restate that while this proof is non-black 147, doi: 10.1109/SCT.1995.514853. keywords: Computer
box, it is a relativizing proof, and there is a known relativiza- science;Drives;Distributed computing;Cryptography;Complexity
tion barrier established by [11]. Further progress would likely theory;Bibliographies.
[7] A. Bogdanov and L. Trevisan, “On Worst-Case to Average-Case Reduc-
combine this proof technique with non-relativizing techniques tions for NP Problems,” SIAM Journal on Computing, vol. 36, no. 4,
to establish non-blackbox and non-relativizing results. pp. 1119–1159, 2006, doi: 10.1137/S0097539705446974.
[8] O. Lupanov, Implementing the algebra of logic functions in terms of
VI. C ONCLUSION constant-depth formulas in the basis +,*, -. English translation inSov.
Phys.-Dokl., 6, 2, 1961; orig. inDokla. Akad. Nauk SSSR, 136, 5.
We explored various techniques related to metacomplexity [9] Trevisan, L., Vadhan, S. Pseudorandomness and Average-Case Com-
plexity Via Uniform Reductions. comput. complex. 16, 331–364 (2007).
problems, beginning with natural proofs, which are intuitive https://doi.org/10.1007/s00037-007-0233-x.
and straightforward but ultimately end up being overly de- [10] Irit Dinur and Shmuel Safra. 2004. On the hardness of approximat-
manding. The existence of pseudorandomness pretty much ing label-cover. Inf. Process. Lett. 89, 5 (16 March 2004), 247–254.
https://doi.org/10.1016/j.ipl.2003.11.007.
ensures that natural proofs cannot be used to reach our goal. [11] Ker-I Ko. 1991. On the complexity of learning minimum time-bounded
We then examined black-box proofs, which use minimal Turing machines. SIAM J. Comput. 20, 5 (Oct. 1991), 962–986.
information about the underlying structure and ignore the https://doi.org/10.1137/0220059
[12] Rahul Santhanam. Why MCSP Is a More Important Problem Than
efficiency of the underlying algorithm. However, this approach SAT (Invited Talk). In 42nd IARCS Annual Conference on Foundations
often proves too conservative. We talked briefly about the of Software Technology and Theoretical Computer Science (FSTTCS
relativization and black-box barriers in the pursuit of our goal, 2022). Leibniz International Proceedings in Informatics (LIPIcs), Vol-
ume 250, p. 2:1, Schloss Dagstuhl – Leibniz-Zentrum für Informatik
and some of the papers we looked at were able to overcome (2022).
these barriers individually. Hirahara expects that these building
blocks are sufficient to overcome the barriers together, and that
a combination of these proof techniques will soon be able to
land a breakthrough.
Kolmogorov Complexity appears well-suited for tackling
metacomplexity challenges and newer papers often use more
and more concepts from algorithmic information theory to
achieve their results. Much like MCSP did with SAT [12], it’s
possible that questions surrounding MINKT or MKTP may
eventually take center stage in metacomplexity research.

ACKNOWLEDGMENT
I would like to express my sincere gratitude to my project
supervisor, Dr. Satyadev Nandakumar, for his guidance and
mentorship throughout this project. It has been a privilege to
explore this challenging field, which is graced by extremely
elegant proofs, and I am grateful for this opportunity to be
able to explore the same by going through some of the most
important papers in this field through the ages.

R EFERENCES
[1] Furst, M., Saxe, J.B. & Sipser, M. Parity, circuits, and the polynomial-
time hierarchy. Math. Systems Theory 17, 13–27 (1984)
[2] Alexander A. Razborov and Steven Rudich. 1994. Natural proofs. In
Proceedings of the twenty-sixth annual ACM symposium on Theory of
Computing (STOC ’94). Association for Computing Machinery, New
York, NY, USA, 204–213. https://doi.org/10.1145/195058.195134
[3] S. Hirahara, ”NP-Hardness of Learning Programs and Partial MCSP,”
2022 IEEE 63rd Annual Symposium on Foundations of Com-
puter Science (FOCS), Denver, CO, USA, 2022, pp. 968-979, doi:
10.1109/FOCS54457.2022.00095.
[4] Marco L. Carmosino, Russell Impagliazzo, Valentine Kabanets, and
Antonina Kolokolova. Learning Algorithms from Natural Proofs. In
31st Conference on Computational Complexity (CCC 2016). Leibniz
International Proceedings in Informatics (LIPIcs), Volume 50, pp. 10:1-
10:24, Schloss Dagstuhl – Leibniz-Zentrum für Informatik (2016)
[5] S. Hirahara, ”Non-Black-Box Worst-Case to Average-Case Reductions
within NP,” 2018 IEEE 59th Annual Symposium on Foundations of
Computer Science (FOCS), Paris, France, 2018, pp. 247-258, doi:
10.1109/FOCS.2018.00032.