Online Ordinal Problems: Optimality of Comparison-based Algorithms and their Cardinal Complexity

We consider ordinal online problems, i.e., tasks that only require pairwise comparisons between elements of the input. A classic example is the secretary problem and the game of googol, as well as its multiple combinatorial extensions such as $(J, K)$-secretary, 2-sided game of googol, ordinal-competitive matroid secretary. A natural approach to these tasks is to use ordinal online algorithms that at each step only consider relative ranking among the arrived elements, without looking at the numerical values of the input. We formally study the question of how cardinal algorithms (that can use numerical values of the input) can improve upon ordinal algorithms. We give first a universal construction of the input distribution for any ordinal online problem, such that the advantage of any cardinal algorithm over the ordinal algorithms is at most $1+\varepsilon$ for arbitrary small $\varepsilon\gt 0$. This implies that lower bounds from [Buchbinder, Jain, Singh, MOR 2014], [Nuti and Vondrák, SODA 2023] hold not only against any ordinal algorithm, but also against any online algorithm. Another immediate corollary is that cardinal algorithms are no better than ordinal algorithms in the matroid secretary problem with ordinal-competitive objective of [Soto, Turkieltaub, Verdugo, MOR 2021]. However, the value range of the input elements in our construction is huge: $N=$ $O\left(\frac{n^{3} \cdot n ! \cdot n !}{\varepsilon}\right) \uparrow \uparrow(n-1)$ (tower of exponents) for an input sequence of length n. As a second result, we identify a class of natural ordinal problems and find cardinal algorithm with a matching advantage of $1+\Omega\left(\frac{1}{\log (c) N}\right)$, where $\log^{(c)} N=\log \log \ldots \log N$ with c iterative logs and c is an arbitrary constant $c \leq n-2$. This suggests that for relatively small input numerical values N the cardinal algorithms may be significantly better than the ordinal algorithms on the ordinal tasks, which are typically assumed to be almost indistinguishable prior to our work. This observation leads to a natural complexity measure (we dub it cardinal complexity) for any given ordinal online task: the minimum size $N(\varepsilon)$ of different numerical values in the input such the advantage of cardinal over ordinal algorithms is at most $1+\varepsilon$ for any given $\varepsilon\gt 0$. As a third result, we show that the game of googol has much lower cardinal complexity of $N=O\left(\left(\frac{n}{\varepsilon}\right)^{n}\right)$.