Theory of Computing
-------------------
Title : Special Issue: CCC 2019: Guest Editor's Foreword
Authors : Yuval Filmus
Volume : 16
Number : 19
Pages : 1-5
URL : http://www.theoryofcomputing.org/articles/v016a019
Abstract
--------
This collection comprises the expanded and fully refereed versions of
selected papers presented at the 34th Computational Complexity
Conference (CCC 2019), held July 18--20, 2019 in New Brunswick, NJ,
USA. These papers were selected by the Program Committee from among
the $32$ papers that appeared in the conference proceedings.
Preliminary versions of the papers were presented at the conference,
and the extended abstracts appeared in the proceedings of the
conference, published by Dagstuhl Publishing, LIPIcs. The CCC Program
Committee selected 32 out of 99 submissions for presentation at
the conference; of these, the six described below were invited to this
Special Issue. These six papers were refereed in accordance with the
rigorous standards of Theory of Computing.
* "UG-hardness to NP-hardness by losing half" by Amey Bhangale and
Subhash Khot.
Many hardness of approximation results in the literature are
conditioned on Khot's celebrated Unique Games Conjecture, which
states that it is NP-hard to distinguish instances of Unique Label
Cover which are $(1-\epsilon)$-satisfiable from those which are only
$\epsilon$-satisfiable. Recently, in a breakthrough result, Khot,
Minzer and Safra (building on previous work together with Dinur and
Kindler) proved the 2-to-1 conjecture, which states that it is
NP-hard to distinguish instances of Unique Label Cover which are
$(1/2-\epsilon)$-satisfiable from those which are $\epsilon$-satisfiable.
It might seem that many hardness of approximation results which are
conditional on the Unique Games Conjecture can be adapted to use the
2-to-1 theorem instead, losing a half in the resulting approximation
ratio. However, it turns out that such results need a stronger promise
in the $(1/2-\epsilon)$-satisfiable case: instead of just requiring a
$1/2-\epsilon$ fraction of the constraints to be satisfiable, they
require the existence of a $1/2-\epsilon$ fraction of the vertices
(on one of the sides), all of whose adjacent constraints can be
simultaneously satisfied. The main result of the paper is this
stronger form of the 2-to-1 theorem. As applications, the authors
prove the following hardness of approximation results:
1. Independent set is $O(d/\log^2 d)$-hard to approximate in
$d$-regular graphs. This is tight up to constants due to a matching
algorithm of Bansal, Gupta, and Guruganesh, and improves the previous
known hardness of $O(d/\log^4 d)$, due to Chan.
2. Maximum acyclic subgraph is $2/3$-hard to approximate, improving on
the previous known hardness of $14/15$, due to Austrin, Manokaran, and
Wenner. The Unique Games Conjecture itself implies the optimal hardness
of $1/2$, as shown by Guruswami, Manokaran, and Raghavendra, matching
a trivial algorithm.
3. For any predicate $P$, if $P$-CSP has a $(c,s)$-SDP integrality gap,
then $P$-CSP is $(c/2,s)$-hard to approximate. Raghavendra famously
showed that the Unique Games Conjecture implies the optimal
$(c,s)$-hardness.
* "Sherali--Adams strikes back" by Ryan O'Donnell and Tselil Schramm.
How does linear programming compare with semidefinite programming?
A case in point is the fundamental problem of Maximum Cut. Goemans
and Williamson famously gave an $0.878$-approximation algorithm
based on semidefinite programming, which is tight assuming the
Unique Games Conjecture. In contrast, linear programming relaxations
are much weaker: Charikar, Makarychev, and Makarychev showed that the
integrality gap of _any_ linear programming relaxation of size
$2^{n^\delta}$ has an integrality gap of $2-\epsilon$ on a random
$\Delta$-regular graph, for constant $\Delta$.
Should we, then, give up on linear programming? The main result of
this paper shows that $n^{O(c/\log \Delta)}$ rounds of Sherali--Adams
(a linear programming hierarchy) can certify, with high probability,
that the maximum cut in a random $\Delta$-regular graph is at most
$1/2 + 2^{-c}$. This refutes the common assumption that certifying
such bounds requires linear programs of truly exponential size
(that is, $2^{\Omega(n)}$). The result is actually stronger -- it
suffices that all nontrivial eigenvalues of the graph be small enough,
a condition that a random $\Delta$-regular graph satisfies with high
probability.
Using similar techniques, the authors prove analogous results for
random CSPs and for Independent Set and Vertex Cover on regular graphs.
* "Limits on the universal method for matrix multiplication" by Josh Alman.
This paper received the Best Student Paper award.
"Barriers for fast matrix multiplication from irreversibility" by
Matthias Christandl, Peter Vrana, and Jeroen Zuiddam.
We discuss these two papers together, since their main results are
essentially identical.
The area of fast matrix multiplication has recently seen a resurgence,
with work of Stothers, Vassilevska-Willams, and Le Gall progressively
improving the decades-old algorithm of Coppersmith and Winograd, from
$\omega < 2.376$ to $\omega < 2.373$. Can this approach prove that
$\omega = 2$? Ambainis, Filmus and Le Gall ruled this out, proving a
barrier of $2.3725$ for this particular approach.
The algorithm of Coppersmith and Winograd and its recent improvements
all rely on an analysis of the so-called Coppersmith--Winograd tensor
using Strassen's laser method. The result of Ambainis et al. shows
that applying the laser method to this tensor cannot give a bound
better than $\omega < 2.3725$. Alman and Vassilevska-Williams
subsequently showed that even if the laser method is replaced by
monomial degeneration (a vast generalization of the laser method),
then the Coppersmith--Winograd tensor cannot be used to prove that
$\omega = 2$.
In the present two papers, similar results are proved for an even more
general method, degeneration. In this method, one takes a large tensor
power of the starting tensor, and then degenerates it into a matrix
multiplication tensor. The papers consider three different classes of
tensors:
1. Coppersmith--Winograd tensors: these are the tensors used by
Coppersmith and Winograd, as well as in all recent work. They are
parametrized by an integer $q$.
2. Simple (or small) Coppersmith--Winograd tensors: these were used
as a "warm-up" in the work of Coppersmith and Winograd, and are also
parametrized by an integer $q$.
3. Cyclic group tensors, also known as reduced polynomial multiplication
tensors: these arise from the group-theoretic approach to fast matrix
multiplication, due to Cohn and Umans. Proving $\omega = 2$ is ruled
out using all of these tensors, except for the simple
Coppersmith--Winograd tensor in the case $q = 2$ (this exception
corresponds to an observation of Coppersmith and Winograd).
Alman proves his results using a parameter known as _asymptotic rank_,
whereas Christandl et al. bound _asymptotic subrank_, which is a priori
tighter. However, Strassen showed that for the tensors considered in
the papers, the two parameters coincide, and moreover are given by
a third parameter, the _minimum quantum functional_, which is easy
to compute, and is used in both papers.
* "Fourier and circulant matrices are not rigid" by Zeev Dvir and Allen Liu.
Matrix rigidity was invented by Valiant as an avenue toward proving
superlinear size lower bounds on circuits of logarithmic depth for
explicit functions. He showed that if a matrix $A$ is
$(n/\log\log n, n^{1+\epsilon})$-rigid (Valiant-rigid for short),
meaning that reducing its rank below $O(n/\log\log n)$ requires
changing $\Omega(n^{1+\epsilon})$ entries, then $Ax$ cannot be computed
using a circuit of size $O(n)$ and depth $O(\log n)$. Valiant showed
that a random matrix is $(r,(n-r)^2)$-rigid, and in particular
Valiant-rigid, but the best explicit constructions are only
$(r, (n^2/r) \log (n/r))$-rigid.
It has long been conjectured that common matrices such as the
Walsh--Hadamard matrix are Valiant-rigid. Recently, Alman and Williams
surprisingly refuted this conjecture, showing that the Walsh--Hadamard
matrix is not Valiant-rigid. The current paper shows that several broad
classes of matrices are also not Valiant-rigid:
1. Discrete Fourier Transform (DFT) matrices over any finite abelian
group $G$. When $G=\Z_2^n$, this is the Walsh--Hadamard
matrix. When $G$ is a cyclic group, we get the classical DFT matrix.
2. Circulant matrices. These are matrices whose rows consist of all
possible rotations of the first row.
3. Toeplitz matrices. These are matrices in which all diagonals are
constant.
4. $G$-circulant matrices for any finite abelian group $G$. These are
$|G|\times |G|$ matrices in which the $(x,y)$ entry $(x,y\in G)$ depends
only on $x-y$. When $G$ is a cyclic group, we get the classical circulant
matrices.
Analogous results are also obtained over finite fields $\F_q$
under the assumption that $\gcd(|G|,q)=1$.
* "Hardness magnification near state-of-the-art lower bounds" by
Igor Carboni Oliveira, Jan Pich, and Rahul Santhanam.
Hardness magnification is the phenomenon in which a weak lower bound
on a specific problem implies a much stronger lower, often beyond the
reach of current techniques. The paper proves several such results:
1. If distinguishing Boolean functions whose circuit complexity is at
least $2^{\beta n}$ from those whose circuit complexity is at most
$2^{\beta n}/cn$ is hard for circuits of size $N^{1+\epsilon}$
(where $N = 2^n$ is the input size), then NP doesn't have polynomial
size circuits.
2. If distinguishing Boolean functions whose Kt complexity (a kind of
Kolmogorov complexity taking into account the running time) is at least
$2^{\beta n} + cn$ from those whose Kt complexity is at most $2^{\beta n}$
is hard for de Morgan formulas of size $N^{2+\epsilon}$ or for
$U_2$-formulas (also allowing XOR) of size $N^{3+\epsilon}$, then EXP
doesn't have polynomial size formulas. Many other such results concerning
Kt complexity are proved in the paper. The results on Kt complexity follow
from techniques in a previous paper of the authors, while the result on
circuit complexity is the main technical contribution.
These hardness magnification results are complemented by several lower
bounds of a similar form but involving different parameters:
1. Neither de Morgan formulas of size $N^{2-\epsilon}$ nor $U_2$-formulas
of size $N^{3-\epsilon}$ can distinguish functions whose Kt complexity
is at least $2^{n-1}$ from those whose Kt complexity is at most
$2^{(1-\delta)n}$.
2. $U_2$-formulas of size $N^{2-\epsilon}$ cannot distinguish functions
whose Kt complexity is at least $2^{(\epsilon/2)n-2}$ from those whose
Kt complexity is at most $Cn^2$.
3. Such formulas also cannot distinguish functions whose circuit complexity
is at least $2^{(\epsilon/2-o(1))n}$ from those whose circuit complexity
is at most $n^{1+\delta}$.
---------------------------
I would like to thank the authors for their contributions, the CCC
program committee for their initial reviews, Amir Shpilka for his
fantastic work as chair of the program committee, Laszlo Babai for his
advice on matters related to _Theory of Computing,_ and the anonymous
referees for their hard work. It was a pleasure to edit this
Special Issue for Theory of Computing.
December 31, 2020
Yuval Filmus
Guest Editor
---------------------------
CCC 2019 Program Committee
* Andrej Bogdanov (Chinese University of Hong Kong)
* Irit Dinur (Weizmann Institute of Science)
* Yuval Filmus (Technion -- Israel Institute of Technology)
* Pavel Hrubeš (Czech Academy of Sciences)
* Valentine Kabanets (Simon Fraser University)
* Gillat Kol (Princeton University)
* Troy Lee (University of Technology Sydney)
* Raghu Meka (University of California at Los Angeles)
* Ramprasad Saptharishi (Tata Institute of Fundamental Research)
* _Amir Shpilka_ (Tel Aviv University) (Chair)
* Madhu Sudan (Harvard University)
---------------------------