[stellalo]

Recasting quantile estimation as an optimization problem is trivial: the q-quantile minimizes the “pinball” loss (see first eqn in http://statweb.stanford.edu/~owen/courses/305a/lec18.pdf) with parameter q. What they do in the paper is to take subgradient steps with respect to the latest observation (just think about subgradients as gradients, since the loss function is everywhere differentiable except for one point)

[zaroth]

I hate it when the complexity of the lingo dramatically exceeds the complexity of the algorithm. Language shouldn’t be the barrier to understanding.

This seems to be particularly true in computer learning. We’re taking about a conditional step function here, right?

[eru]

The lingo is complex here, because it's general enough to be used for much more complicated cases.

Think of it as a 'hello world' program. The typically 'hello world' program in eg Java teaches you more about the lingo of Java than about solving the problem of putting 'hello world' on the screen.

(Of course, there are still plenty of bad reasons to describe simple things in complex lingo. But the above is one good reason.)