Some highlights of ICALP25

The 52nd EATCS International Colloquium on Automata, Languages, and Programming (ICALP) took place in the campus of Aarhus University. I really liked the conference center in which ICALP25 took place and enjoyed visiting the charming city of Aarhus again.

Below are my comments on some of the talk I have attended.

Dana Ron - On Tolerant Testing and Distance Approximation

Tolerant testing is a generalization of property testing. It refer to the study of sublinear complexity algorithms for solving an approximate decision problem in which one has to distinguish between instances that are close to the predetermined property and instances that are far from the property. The actual definition specifies a way of accessing parts of the instance as well as a distance measure and two parameters that specify what is far and what is close. These parameters may be arbitrary in some cases, and related in other cases (e.g., being far may mean being at distance that is greater than twice the distance that is considered close).

Tolerant testing (in the arbitrary parameters version) is closely related to distance approximation, but may be weaker when only specific parameters pairs of distance parameters are allowed.

Approximating Dasgupta Cost in Sublinear Time from a Few Random Seeds (by M. Kapralov, A. Kumar, S. Lattanzi, A. Mousavifar, W. Wrzos-Kaminska)

The Dasgupta Cost of a graph is the sum $\sum_{v\in V}\rk(v)\cdot \deg(v)$, where $\rk(v)$ is the ranking of $v$ when ordering the vertices according to their degree (such that $\rk(v) \leq \rk(w)$ iff $\deg(v) \geq \deg (w)$. Can this quantity be approximated using sublinear number of queries? The paper provides a positive answer in the special case that the graph is expanding.

Recall that the sum of vertex degrees as well as the sum of powers of degrees can be approximated on sublinear time (see work by Eden, Ron, and Seshadhri). A general formulation that covers all cases is given by the sum $\sum_{v\in V} f(v)$, where $f$ is an arbitrary ``nice'' function. Note that one may consider the problem when allowing degree queries only (as done in the current work), but incidence query may be useful especially in case the function $f$ is not arbitrary (as in the cases discussed above).

A 0.51-Approximation of Maximum Matching in Sublinear n sup 1.5 Time (by S. Mahabadi, M. Roghani, J. Tarnawski)

What I found appeaking in this work is the extensive and clever use of local computation. Specifically, a problem about graphs is solved by local computation of auxiliary graphs, and application of other sublinear-time algorithms to the resulting graphs.

Relative-error testing of conjunctions and decision lists (by X. Chen, W. Pires, T. Pitassi, R. Servedio)

The relative-error model was introduced in order to study the complexity of testing sparse Boolean functions. This model deviates from the standard property testing model by defining a function $f:[N]\to\{0,1\}$ as $\eps$-close to a set $F$ if there exists $f'\in F$ that differs from $f$ on at most $\eps\cdot|f^{-1}(1)|$ (rather than $\esp\cdot N$) entries. Testers in this model are also given access to a sampling devcice that returns uniformly distributed element of $f^{-1}(1)$. (Unfortuntaely, the term `relative error' only captures the first aspect but not the second one.)

One interesting result proved in this paper is that, under some reasonable assumptions, the query complexity of standard testing is upper-bounded by the query complexity of testing in the relative-error model. One may assume, wlog, that $F$ is not empty, so the main condition is that the query complexity of the given tester behaves nicely (essentually, increases at least linearly in $1/\eps$).

As a corollary to the foregoing combined with their tester for Monomials in the relative-error model, they obtain a one-sided tester in the standard model with linaer in $1/\eps$ complexity. In fact, they can also obtain a (constant-query) proximity-oblivious tester for Monomials with polynomial rejection probability (cf. Section 1.4 in HERE).

Sepehr Assadi - Presburger Award lecture

The focus was recent lower bounds on the space complexity of multi-pass graph streaming problems, where the focus is on a small number of passes (e.g., a contant, polylog, or evern just sublinear). A key technique that was highlighted is obtaining a result for $p+1$ passes by considering a related problem and a lower bound that refers to $p$-pass algorithms. This reminded by of a talk by Jukka Suomela at the 3rd WOLA, which refered to the (distributed computing) LOCAL model (where locality lower bounds were obtained through round elimination).

Mary Wootters - List-Recovery and (Non)-Applications

List-recovery can be seen as a generalization of list decoding. Recall that in the latter case, for parameters $\delta$ and $L$, one is interested in the list of size at most $L$ that consists of all codewords that are at relative distance at most $\delta$ from a given word. In that case, the focus is on positive $\delat$ (since the case of $\delta=0$ is trivial).

In contrast, in the generalization also the case of $\rho=0$ is interesting, and an additional parameter $\ell$ is added (where list decoding corresponds to the case $\ell=1$). In that case, we consider codes $C\subseteq\Sigma^n$, and for each sequence of sets $S_1,...,S_n\subseteq\Sigma$ such that $|S_i| \leq \ell$ for every $i$, we seek a list (of size at most $L$) of all codewords $(c_1,...,c_n)\in C$ such that $c_i\in S_i$ for at least $(1-\delta)\cdot n$ of the $i$'s.

As in the case of list decoding (qand unique decoding), there is the combinatorial problem of existence of such codes, and the algorithmic question of constructing them and performing enconding and decoding efficiently.

Reference

See the program and the proceedings.


Back to list of Oded's choices.