A Trichotomy for Transductive Online Learning
This provides foundational theoretical insights into online learning complexity, though it is incremental in refining existing bounds.
The paper tackles the problem of determining the minimal number of mistakes in transductive online learning, proving a trichotomy where mistakes scale as n, Θ(log(n)), or Θ(1) based on VC and Littlestone dimensions, and improves a lower bound from Ω(√log(d)) to Ω(log(d)).
We present new upper and lower bounds on the number of learner mistakes in the `transductive' online learning setting of Ben-David, Kushilevitz and Mansour (1997). This setting is similar to standard online learning, except that the adversary fixes a sequence of instances $x_1,\dots,x_n$ to be labeled at the start of the game, and this sequence is known to the learner. Qualitatively, we prove a trichotomy, stating that the minimal number of mistakes made by the learner as $n$ grows can take only one of precisely three possible values: $n$, $Θ\left(\log (n)\right)$, or $Θ(1)$. Furthermore, this behavior is determined by a combination of the VC dimension and the Littlestone dimension. Quantitatively, we show a variety of bounds relating the number of mistakes to well-known combinatorial dimensions. In particular, we improve the known lower bound on the constant in the $Θ(1)$ case from $Ω\left(\sqrt{\log(d)}\right)$ to $Ω(\log(d))$ where $d$ is the Littlestone dimension. Finally, we extend our results to cover multiclass classification and the agnostic setting.