Eq. (5) —
ProvedForecastAggLearn.eq_5forecast-aggregationonline-learningp2o-gran-per-chapterp2o-plan-paperp2o-v1probability
Let , , , and let be a Bernoulli variable with success probability (in the paper ). With the loss of eq. (4),
So the per-period loss (4), which depends only on the realized state, is an unbiased version of the expected loss ; regret bounds for the realized losses transfer to expected regret.
Differences from the source:
- encoding: E_{ω_t}[L̄_t(h)] with E[ω_t | z̃_t] = r*(z̃_t) → q · L̄(h)|{ω=1} + (1 − q) · L̄(h)|{ω=0} for every q ∈ [0, 1]; reason: the expectation of a function of a Bernoulli(q) state; r* ranges over [0, 1]
Preamble
import Mathlib import Definitions.Def_ForecastAggLearn_barL
Formal statement
namespace ForecastAggLearn
/-- **Eq. (5): the expected per-period loss.**
Babichenko, Garber, arXiv:1802.07107v1, §4.2.1, eqs. (4)–(5), p.18: with
`L̄_t(h) = (1 − ω_t)(h · z̃_t + µ̃_t) + ln(1 + exp(−h · z̃_t − µ̃_t))` (4), and `ω_t` a Bernoulli
variable with success probability `r*`, "∀h ∈ ℝⁿ, t ∈ [T] : E_{ω_t}[L̄_t(h)] = L̃(h · z̃_t + µ̃_t)"
where `L̃(r̃) = −(r* − 1) r̃ + ln(1 + exp(−r̃))` (p.16).
**Formalization Note.** `q` is the success probability `r* = E[ω_t | z̃_t]`, and the expectation
over `ω_t` is `q L̄_t(h)|_{ω=1} + (1 − q) L̄_t(h)|_{ω=0}`. -/
theorem eq_5 {n : ℕ} (z : EuclideanSpace ℝ (Fin n)) (μt : ℝ) (h : EuclideanSpace ℝ (Fin n))
(q : ℝ) (hq0 : 0 ≤ q) (hq1 : q ≤ 1) :
q * barL z μt true h + (1 - q) * barL z μt false h =
-(q - 1) * (inner ℝ h z + μt) + Real.log (1 + Real.exp (-(inner ℝ h z + μt))) := by sorry
end ForecastAggLearn
Source
Babichenko, Garber, Learning of Optimal Forecast Aggregation in Partial Evidence Environments, arXiv:1802.07107v1, §4.2.1, eqs. (4)–(5), p.18
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.