# DS-2012-05: Query-Efficient Computation in Property Testing and Learning Theory

DS-2012-05: Soriano, David García (2012) Query-Efficient Computation in Property Testing and Learning Theory. Doctoral thesis, University of Amsterdam.

## Abstract

How can we solve computational tasks when we simply don't have enough time to
process all the data? For example, given a sequence of numbers, can we
determine if they start repeating by looking at just a few of them? Or, given
query access to some boolean function $f$, how many queries to $f$ do we need
to test whether it is monotone? How about deciding if $f$ is essentially''
the same as another, known, function $g$?

One of the central goals of theoretical computer science is to understand the
limits of computation in a variety of models, and to characterize the
resources required to solve certain tasks. The kinds of problems outlined
above are particularly amenable to study in the property testing model,
and it is under this prism that we investigate them. In property testing,
algorithms are required to distinguish those objects which satisfy the desired
property from those which are far from it. We seek query-efficient randomized
testers to solve them (upper bounds), along with rigorous proofs
explaining why no significantly better solutions are possible, even in
principle (lower bounds). The results presented here draw from
techniques in probability theory, graph theory, extremal combinatorics, the
theory of permutation groups, coding theory, the analysis of boolean
functions, and number theory.

The starting point is our work on the problem of testing function
isomorphism in Chapter 2. Two boolean functions on $n$-bit inputs are
isomorphic if they are the same up to permutations of the $n$ input
variables. The main case of interest is when $g$ is known and $f$ needs to
be queried. It is known that the task can be accomplished with
$\widetilde{O}(n)$ queries, but this is exponentially larger than the best
lower bounds from prior work. Here we bridge the gap by contributing an
almost-optimal adaptive lower bound of $\Omega(n)$ queries for worst-case
functions~$f$. Several other variants of the problem are also discussed.

In addition, we show in Chapter 3 that when $f$ is a $k$-junta (meaning that
it is determined by just $k$ of the $n$ input variables), the complexity
of testing isomorphism to $f$ is reduced to $\widetilde{O}(k)$. (In contrast,
another result of ours is that if we impose the seemingly weak restriction
that the tester have one-sided error, the complexity of testing
isomorphism to $k$-juntas becomes roughly $\log\binom{n}{k}$, which is
much larger for small $k$.) In the process we construct objects of independent
interest called sample extractors. These are efficient algorithms that
allow us to draw samples from the truth table of the core'' function on $k$
variables that determines a given $k$-junta function.

Next we give a partial characterization of the set of functions against which
it is possible to test isomorphism with constant query complexity. We show in
Chapter 4 that, for any function $f$ with polynomially many different
permutations, isomorphism to $f$ is testable with constantly many queries.
This theorem extends previous results from junta testing and in fact covers
all functions known to date for which isomorphism testing is easy. On a
related note, observing the connection between testing function isomorphism
and testing hypergraph isomorphism, we turn our attention to testing
isomorphism for uniform hypergraphs. We give a characterization of the
class of hypergraphs of constant rank for which isomorphism can be efficiently
tested, generalizing a result of Fischer (STOC'04)
regarding graph isomorphism.

In Chapter 5 we establish links with the field of group testing,
and observe that ideas from isomorphism testing can be applied to the study
of a naturally-arising relaxation of group testing problems; both the lower
bound methods and the algorithms are of use. We determine the exact query
complexity of the relaxed group testing problem for non-adaptive algorithms,
up to constant factors. The question of obtaining explicit lower bounds for
these problems is also addressed; it turns out that parity functions (XORs of
subsets of the input variables) exemplify the worst-case lower bounds for
testing function isomorphism, both for one-sided and for two-sided testers.

We move on to discuss other property testing problems in Chapter 6. It turns
out that our sample extractors can be used to enhance the best algorithms
known for many other testing problems, specifically those defined by the
property of having concise representations, such as testing if $f$ can
be computed by small circuits, or decision trees of small size. We also
provide new lower bounds for some of them, resolving several open questions
posed by Diakonikolas et al.~(FOCS'07).

In Chapter 7 we examine parity functions under a different light:
computational learning theory. Testing and learning are two closely
related areas. Instead of testing whether the function $f$ is a parity on a
small number of variables, in testing we assume that it is. Then, based on
some example values of $f$, the learner tries to predict $f(x)$ for other
inputs $x$ with good accuracy. We work in the mistake-bound model of
learning, which is stronger than the (more usual) $\PAC$ model. It is a
standard fact that parities can be learned with mistake bound (or sample
complexity) $O(k \log n)$, but no polynomial-time implementation of such a
learning algorithm is known. We design a simple, deterministic,
polynomial-time algorithm for learning $k$-parities with mistake
bound $O(n^{1-\frac{1}{k}})$. This is the first polynomial-time
algorithm to learn $\omega(1)$-parities with mistake bound $o(n)$,
and using standard conversion techniques it implies an improvement
over the results of Klivans and Servedio (COLT'04) for learning $k$-parities
in the $\PAC$ model.

Apart from this, we also consider one of the fundamental problems in property
testing: monotonicity of functions (not necessarily boolean). In Chapter 8 we
consider functions defined on the $n$-dimensional hypercube, and give an
$\Omega(n)$ lower bound for one-sided, non-adaptive testers of monotonicity.
As there is still a sizable gap between this and the best upper bounds known,
we look at a natural approch to obtaining upper bounds via the study of
routing properties of the hypercube. It had been previously observed that
if any set of source-sink pairs on the directed hypercube (where all
sources and sinks are distinct) could be connected with edge-disjoint
paths, then monotonicity of functions on the $n$-dimensional hypercube
would be testable with $O(n)$ queries. Determining whether this property
holds was posed as an open problem by Lehman and Ron (J.~Comb.~Theory,
Ser.~A, 2001), but the question had remained elusive for nearly a decade.
By analyzing the combinatorial properties of the hypercube, we show that the
answer is negative, and that this approach must fail short of matching the
current lower bounds, or even approaching them.

In the final chapter we address the cycle detection problem. Perhaps
counter-intuitively, we prove that if the smallest periodic part of the
sequence contains distinct elements, the period length $r$ can be determined
by making a number of probes to the sequence that is sublogarithmic in $r$;
moreover, this is not too far from optimal. We also study variants of the
jump'' between positions that are not too far away from each other. The