Back to blog

# CCC 2018

Few weeks ago, I attended the 33rd Computational Complexity Conference (CCC 2018) at UCSD. It was my first time going to CCC and I really enjoyed it a lot. As some of my friends told me before, since CCC is quite small and focused (in complexity), sometimes one can find it more useful and helpful. Personally I even like CCC 2018 more than FOCS 2017. The following are some papers that I found interesting to me during the conference. I decided to carefully study these papers and write some thoughts about them after the conference, however, it took much more time than I expected to finish this task.

## The Power of Natural Properties as Oracles

Author(s): Russell Impagliazzo, Valentine Kabanets, and Ilya Volkovich.

The goal of this paper is two-fold. Firstly they want to investigate how powerful a natural property is following the learning with natural proof work by (Carmosino et al., 2016). On the other hand, they provide envidence that the minimum circuit size problem, i.e., $\MCSP$, is as hard as $\NP$.

Here, we list some of their results in the following.

• $\ZPEXP^{\MCSP}\not\subseteq\Ppoly$ improving the previous $\ZPEXP^{\NP}\not\subseteq\Ppoly$.
• $\ZPP^{\MCSP}/1\not\subseteq\SIZE[n^k]$ for all $k\in\N$.
• Suppose indistinguishable ofuscation exists, then $\MCSP\in\ZPP$ if and only if $\NP=\ZPP$.
• Replace the $\NP$ oracle in many complexity relations with $\MCSP$ oracle.

#### Collapse theorem

Let $\Gamma\in\{\PaP,\ShP,\PSPACE,\EXP,\NEXP,\EXP^{\NP}\}$. If $\Gamma\subseteq\Ppoly$, then $\Gamma\subseteq\ZPP^{\MCSP}$.

To interpret this theorem, one have to compare with the previous results. Prior to this work, $\Gamma$ would only collapses to $\MA$ (see this note) which lies in $\ZPP^{\NP}$. The above theorem in some sense indicates the power of $\MCSP$ being an oracle is as powerful as $\SAT$.

#### Circuit lower bounds

$\ZPP^{\MCSP}/1\not\subseteq\SIZE[n^k]$ for all $k\in\N$.

Note that this is an analogy to the $\MA/1\not\subseteq\SIZE[n^k]$ due to Santhanam (Santhanam, 2009) (see this note). As we mentioned previously, $\MA\subseteq\ZPP^{\NP}$ and thus $\ZPP^{\MCSP}$ can be thought of as a $\MCSP$ analogy for $\MA$.

#### Key technique

The following is a kel lemma extracted from (Carmosino et al., 2016) using natural proof (see this note) to learn boolean function.

Let $\mathcal{R}$ be a strongly useful natural proof and let $L$ be a downward self-reducible and self-correctable language. Then, there exists a randomized algorithm with oracle queries to $\mathcal{R}$ such that on input $x$, output $L(x)$ correctly with probability at least $1-1/\poly(n)$ in time $\poly(\card{x},\text{size}(L))$.

An immediate corollary would be a conditional collapse of $\PSPACE$ to $\BPP^{\MCSP}$ since there is a downward self-reducible and self-correctable language in $\PSPACE$ and $\MCSP$ is a strongly useful natural proof.

To move from $\BPP^{\MCSP}$ to $\ZPP^{\MCSP}$, we need one more lemma as follows.

Let $\mathcal{R}$ be a strongly useful natural property. Then, $\BPP\subseteq\ZPP^{\mathcal{R}}$. Moreover, if $\mathcal{R}\in\Ppoly$, then $\BPP^{\mathcal{R}}=\ZPP^{\mathcal{R}}$.

#### Some thoughts

To some extent, the techniques used in this paper are not that surprising after the CIKK paper, however, the high-level message is informative. I feel that these results all highly rely on the self-reducibility of a $\PSPACE$-complete language. Is that possible to go beyond this? Say having similar results for lower complexity classes that do not have a self-reducible language?

## Pseudorandom Generators from Polarizing Random Walks

Author(s): Eshan Chattopadhyay, Pooya Hatami, Kaave Hosseini, and Shachar Lovett.

This work gives a new framework of constructing pseudo random generator (PRG) by using fractional PRG. They construct PRG for functions with low sensitivity with better seed length as well as PRGs for other family of well-studied functions. The key constrution is a fractional PRG for function with exponentially decaying $\ell_1$ tail. In the following, we will first define some notations and present some of their results.

Here, we think of PRG as a random variable $X$ distributes in ${-1,1}^n$ defined as $X=G(U_r)$ where $G:{-1,1}^r\rightarrow{-1,1}^n$ is the generator function, $U_r$ is uniformly distributed over ${-1,1}^r$, and $r$ is the seed length. We say $X$ $\epsilon$-fools a function $f:{-1,1}^n\rightarrow{-1,1}$ if $$\left|\E[f(X)]-\E[f(U_n)]\right|\leq\epsilon.$$ The goal is to minimize the seed length $r$. Note that there are other ways to define PRG and they are basically equivalent.

#### Fractional PRG

A fractional PRG prosed in this paper is a random variable $X$ distributes in $[-1,1]^n$ instead of ${-1,1}^n$ while having the same $\epsilon$-fool property. However, to avoid triviality ($X$ always being $0^n$), we say $X$ is $p$-noticeable for some $p>0$ if $\E[X^2]\geq p^2$.

The paper mentions that a $p$-noticeable fractional PRG that fools function with exponentially decaying $\ell_1$ Fourier tail can be constructed with bounded-wise independent random variables. Concretely, they have the following lemma.

For any $a,b>0$ define $\mathcal{F}_{a,b}^n=\{f\ |\ \sum_{S\subseteq[n]:|S|=k}\card{\hat{f}(S)}\leq a\cdot b^k\}$ be the family of functions with exponentially decaying $\ell_1$ Fourier tail. There exists a $p$-noticeable fractional PRG $X$ that $\epsilon$-fools $\mathcal{F}_{a,b}^n$ with seed length $O(\log\log n+\log(a/\epsilon))$ and $p=\frac{1}{4b^2}$.

Note that many well-studied family of boolean functions have exponentially decaying $\ell_1$ Fourier tail. For instance, $\AC{0}$ circuit, functions with bounded sensitivity, functions computed by certain branching programs, etc.

#### From fractional PRG to PRG

Now that we have fractional PRG for a bunch of boolean function families, the goal is to construct PRG for them. This paper shows how to use fractional PRG to construct PRG for the same family of functions with little blow-up in seed length.

The key idea is using a sequence of indpendent fractional PRG to do a random walk in $[-1,1]^n$ and then rounding to boolean value after few steps. The analysis is based on polarization in one-dimensional martingale. Here, we skip the proof and only state the theorem as follows. Note that we need the function family $\mathcal{F}$ to have some self-trstrictable property in the sense that after fixing some input, the function still lies in the same family.

Let $X_1,X_2,\dots,X_t$ be independent $p$-noticeable fractional PRG that $\epsilon$-fools self-restrictable $\mathcal{F}_{a,b}^n$. When $t=O(\log(n/\epsilon)/p)$, we have $X=M(X_1,X_2,\dots,X_t)$ a PRG that $(t+1)\epsilon$-fools $\mathcal{F}_{a,b}^n$ where $M(\cdot,\dots,\cdot)$ is some random walk function.

#### Some thoughts

The idea of fractional PRG is quite interesting and it is surprising to me that one can really use it to contruct PRG with seed lengths comparable to other highly non-trivial PRG for various computational models. Also, the recent exiciting oracle separation for $\BQP$ and $\PH$ used the analysis in this paper. I still haven’t fully understood the intuition why their tools overcome the previous barrier and it would be interesting to see if this can be applied in other places.

Another intersting research direction here could be generalizing the idea of fractional PRG as the authors suggested in the end the talk/paper. Basically, the currect reduction from fractional PRG to PRG is via a simple summation and rounding. Could a more complicated composition of fractional PRG be more efficient?

## NP-hardness of Minimum Circuit Size Problem for OR-AND-MOD Circuits

Author(s): Shuichi Hirahara, Igor C. Oliveira, and Rahul Santhanam.

The minimum circuit size problem (MCSP) is a well-known problem in $\NP$ in the situation that people do not know if MCSP lies in $\P$ or is $\NP$-complete. The input of MCSP is the truth table of a boolean function $f$ and an integer $s$ such that the goal is to decide whether there exists a circuit of size at most $s$ computing $f$. Currently there’s no evidence against the $\NP$-hardness of MCSP. The only obstacle in showing MCSP is $\NP$-hard is due to Kabanets and Cai showing that a natural reduction from SAT to MCSP would imply a super-polynomial lower bound against $\mathbf{E}$, the family of problems solvable in linear exponential time.

There are not so many results either even one considers MCSP for restricted circuit family. In 1979, Masek showed DNF-MCSP is $\NP$-hard and this is the state-of-the-art. This paper gives the first improvement for the hardness of MCSP for depth-3 circuit.

Theorem The MCSP for $\text{Or}\circ\text{And}\circ\text{Mod}_m$ is $\NP$-hard.

The main idea is viewing the $\text{Or}\circ\text{And}\circ\text{Mod}_m$ as an indicator function for an union of affine spaces. Concretely, let’s start with $m=2$ where $\text{Mod}_2$ is the parity gate. One can see that the preimage of 1 of each parity gate gives an affine space. As a result, one can think of a $\text{Or}\circ\text{And}\circ\text{Mod}_m$ circuit for boolean function $f$ as using affine spaces to cover the 1-preimage of $f$. Quantitatively, the number of affine spaces needed is a lower bound for the number of $\text{And}$ gates needed.

## A story of AM and UNIQUE-SAT

Author(s): Ilya Volkovich.

This paper was presented by Ilya in the rump session. He showed that $\AM=\SBP$ if Unique-Sat ($\USAT$) is $\NP$-hard. Recall that $\MA\subseteq\SBP\subseteq\AM$ and whether $\MA$ euqals to $\AM$ is one of the important question in complexity theory. This paper can be thought of as an indication of $\MA=\AM$. A natural open question after this work would be either (i) improving the assumption, i.e., basing on more believing condition than the $\NP$-hardness of $\USAT$, or (ii) improving the conclusion, i.e., $\MA=\AM$ instead of $\AM=\SBP$.

#### $\MA$, $\SBP$, and $\AM$

One convenient way to view $\MA$ is to think of it as $\NP$ with a randomized verifier. Formally, a language $L$ is in $\MA$ if there exists a polynomial-time computable predicate $R(x,w,r)$ such that \begin{align*} x\in L\Rightarrow\ \exists w,\ \Pr_r[R(x,w,r)=1]\geq\frac{2}{3},\\ x\not\in L\Rightarrow\ \forall w,\ \Pr_r[R(x,w,r)=1]\leq\frac{1}{3}. \end{align*}

As for $\AM$, the verifier (Arthur) will toss the randomness before the prover (Merlin) sends the witness. It immediately follows from the definitions that $\MA\subseteq\AM$. Intuitively, one can think of $\AM$ as a class of languages that have a randomized reduction to $\NP$. Formally, we say $L\in\AM$ if there exists a polynomial-time computable predicate $R(x,w,r)$ such that \begin{align*} x\in L\Rightarrow\ \Pr_r[\exists w,\ R(x,w,r)=1]\geq\frac{2}{3},\\ x\not\in L\Rightarrow\ \Pr_r[\exists w,\ R(x,w,r)=1]\leq\frac{1}{3}. \end{align*}

$\SBP$ is not as well known as $\MA$ and $\AM$ since most complexity courses would not introduce it. However, it is a fairly natural existence between these two very important complexity classes. Let us directly state the definition of $\SBP$. We say $L\in\SBP$ if there exists a polynomial-time computable predicate $R(x,r)$ such that there exists $\epsilon>0$ and $k\in\N$, \begin{align*} x\in L\Rightarrow\ \Pr_r[R(x,r)=1]\geq(1+\epsilon)\cdot\frac{1}{2^{n^k}},\\ x\not\in L\Rightarrow\ \Pr_r[R(x,r)=1]\leq(1-\epsilon)\cdot\frac{1}{2^{n^k}}. \end{align*} It is a simple exercise to check $\MA\subseteq\SBP\subseteq\AM$ by viewing the witness $w$ in $\MA$ and $\AM$ as a part of the randomness $r$.

#### Intuition for the conditional collapse theorem

We are going to briefly explain the intuition for the following main theorem.

If $\USAT$ is $\NP$-hard, then $\AM=\SBP$.

The idea is actualy quite straightforward. Recall the definition of $\AM$. Suppose for a language $L\in\AM$, each yes instance $x\in L$ has a unique witness $w$. Then it immediately follows from the definition that $L\in\SBP$ by simply let the randomness to be $r’=(w,r)$. Namely, to collapse $\AM$ into $\SBP$, it suffices to transform every language $L\in\AM$ to a language that has unique witness.

With this sufficient condition in mind, it is then no difficult to see that under the assumption that $\USAT$ is $\NP$-hard, one can collapse $\AM$ into $\SBP$.

#### Some thoughts

This is quite a cute paper that gives a detailed connection between $\MA,\SBP,\AM$. Though the conditional collapse theorem is not too deep after knowing the definitions of these classes, the paper points out a pave to prove $\AM\subseteq\MA$. A the paper suggested in the end, some natural open questions would then be either improving the assumption or improving the consequence. I personally would be more interested in finding more suitable assumption to collapse $\AM$.

## References

1. Carmosino, M. L., Impagliazzo, R., Kabanets, V., & Kolokolova, A. (2016). Learning algorithms from natural proofs. LIPIcs-Leibniz International Proceedings in Informatics, 50.