§4.1 — the expected log loss in log-likelihood space,
ProvedForecastAggLearn.logit_space_lossforecast-aggregationonline-learningp2o-gran-per-chapterp2o-plan-paperp2o-v1probability
Let and let the aggregator report a log-likelihood , i.e. the forecast . Then its expected logarithmic loss is
and .
In log-likelihood space the log loss becomes a logistic loss, which is convex in the report; this is the loss online gradient descent is run on.
Formalization Note The last identity is stated in with the loss logLoss.
Differences from the source:
- encoding: r*(s) in the display; "the aggregator's logarithmic loss is given by L̃(r̃)" → q ∈ [0, 1] in place of r*(s); the identity with the expected loss q·l(r,1) + (1 − q)·l(r,0) of r = (1 + e^{−r̃})^{−1} stated in ℝ≥0∞; reason: r*(s) ranges over [0, 1]; the last clause is the sentence introducing the display
Preamble
import Mathlib import Definitions.Def_ForecastAggLearn_logLoss
Formal statement
namespace ForecastAggLearn
/-- **The expected log loss in log-likelihood space.**
Babichenko, Garber, arXiv:1802.07107v1, §4.1, display, p.16: for a forecast given in
log-likelihood form `r̃` (the forecast is `(1 + exp(−r̃))^{−1}`),
`L̃(r̃) = −r* ln(1/(1 + exp(−r̃))) − (1 − r*) ln(exp(−r̃)/(1 + exp(−r̃)))
= −r* r̃ − ln(exp(−r̃)/(1 + exp(−r̃)))
= −(r* − 1) r̃ + ln(1 + exp(−r̃))`.
The three real expressions are equal for every `q = r* ∈ [0, 1]` and `r̃ ∈ ℝ`, and they equal the
expected log loss `q l(r, 1) + (1 − q) l(r, 0)` of the forecast `r = (1 + exp(−r̃))^{−1}`.
**Formalization Note.** The last clause links the display to `logLoss` (in `ℝ≥0∞`). -/
theorem logit_space_loss (q : ℝ) (hq0 : 0 ≤ q) (hq1 : q ≤ 1) (rt : ℝ) :
-q * Real.log (1 / (1 + Real.exp (-rt))) -
(1 - q) * Real.log (Real.exp (-rt) / (1 + Real.exp (-rt))) =
-q * rt - Real.log (Real.exp (-rt) / (1 + Real.exp (-rt))) ∧
-q * rt - Real.log (Real.exp (-rt) / (1 + Real.exp (-rt))) =
-(q - 1) * rt + Real.log (1 + Real.exp (-rt)) ∧
ENNReal.ofReal q * logLoss (1 / (1 + Real.exp (-rt))) true +
ENNReal.ofReal (1 - q) * logLoss (1 / (1 + Real.exp (-rt))) false =
ENNReal.ofReal (-(q - 1) * rt + Real.log (1 + Real.exp (-rt))) := by sorry
end ForecastAggLearn
Source
Babichenko, Garber, Learning of Optimal Forecast Aggregation in Partial Evidence Environments, arXiv:1802.07107v1, §4.1, display, p.16
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.