Documentation

Linglib.Core.MeasureTheory.Measure.Tilted

Tilted measures #

Tilting is invariant under scaling the measure, (c • μ).tilted f = μ.tilted f for a finite nonzero c, since the normalization absorbs the factor; so tilting the uniform measure on a finite type is tilting the counting measure. On a finite type the tilt reweights each atom by the exponential of f and renormalizes, a probability measure is a tilt exactly when its atoms are proportional to the reweighted ones, and the counting measure tilted by f gives each point the mass Real.softmax f, which is positive, so that every measure is absolutely continuous with respect to it.

theorem MeasureTheory.tilted_smul {α : Type u_1} {mα : MeasurableSpace α} {μ : Measure α} {f : α → ℝ} {c : ENNReal} (hc0 : c ≠ 0) (hc : c ≠ ⊤) :
(c • μ).tilted f = μ.tilted f

Tilting is invariant under scaling the measure by a finite nonzero constant.

theorem MeasureTheory.uniformOn_univ_tilted {α : Type u_1} {mα : MeasurableSpace α} {f : α → ℝ} [Fintype α] [Nonempty α] :
(ProbabilityTheory.uniformOn Set.univ).tilted f = Measure.count.tilted f

On a finite type, tilting the uniform measure is tilting the counting measure.

theorem MeasureTheory.tilted_real_singleton {α : Type u_1} [MeasurableSpace α] [Fintype α] [MeasurableSingletonClass α] (μ : Measure α) [IsFiniteMeasure μ] (f : α → ℝ) (a : α) :
(μ.tilted f).real {a} = μ.real {a} * Real.exp (f a) / ∑ b : α, μ.real {b} * Real.exp (f b)

On a finite type, the tilted measure at an atom is the atom's mass reweighted by the exponential of f and renormalized.

theorem MeasureTheory.eq_tilted_iff {α : Type u_1} [MeasurableSpace α] [Fintype α] [MeasurableSingletonClass α] (μ ν : Measure α) [IsProbabilityMeasure μ] [IsFiniteMeasure ν] [NeZero ν] (f : α → ℝ) :
μ = ν.tilted f ↔ ∃ Z > 0, ∀ (a : α), μ.real {a} = ν.real {a} * Real.exp (f a) / Z

On a finite type, a probability measure is the tilt of ν by f exactly when its atoms are proportional to those of ν reweighted by the exponential of f.

theorem MeasureTheory.absolutelyContinuous_count_tilted {α : Type u_1} [MeasurableSpace α] [Fintype α] [MeasurableSingletonClass α] (μ : Measure α) (f : α → ℝ) :
μ.AbsolutelyContinuous (Measure.count.tilted f)

On a finite type every measure is absolutely continuous with respect to a tilt of the counting measure, which gives every point positive mass.

theorem MeasureTheory.count_tilted_real_singleton {α : Type u_1} [MeasurableSpace α] [Fintype α] [MeasurableSingletonClass α] (f : α → ℝ) (a : α) :
(Measure.count.tilted f).real {a} = Real.softmax f a

On a finite type, the counting measure tilted by f gives each point its softmax probability.