AutoSamplingTheory.ExampleCases.SmoothedPicardHMC.IdealRGOIdentification

No declaration docstring.

private theorem smooth_lower (f : E → ℝ) (hf : Differentiable ℝ f)
    (beta : NNReal) (hlip : LipschitzWith beta (gradient f)) (x0 u : E) :
    f x0 + inner ℝ (gradient f x0) u - (beta : ℝ)/2 * ‖u‖^2 ≤ f (x0+u) := by
  let H : ℝ → ℝ := fun t => f (x0+t • u) - t * inner ℝ (gradient f x0) u +
    (beta : ℝ)/2 * t^2 * ‖u‖^2
  let D : ℝ → ℝ := fun t => inner ℝ (gradient f (x0+t • u)) u -
    inner ℝ (gradient f x0) u + (beta : ℝ)*t*‖u‖^2
  have hd (t : ℝ) : HasDerivAt H (D t) t := by
    have hp : HasDerivAt (fun r : ℝ => x0+r • u) u t := by
      simpa only [Pi.add_def, zero_add, one_smul, id_eq] using
        (hasDerivAt_const t x0).add ((hasDerivAt_id t).smul_const u)
    have hg := (hf (x0+t • u)).hasGradientAt
    have hc := hg.hasFDerivAt.comp_hasDerivAt t hp
    have hlin := (hasDerivAt_id t).mul_const (inner ℝ (gradient f x0) u)
    have hquad := (((hasDerivAt_id t).pow 2).const_mul ((beta : ℝ)/2)).mul_const (‖u‖^2)
    convert! (hc.sub hlin).add hquad using 1
    all_goals first | rfl | (dsimp [H,D,Function.comp_def,Pi.add_def,Pi.sub_def]; ring)
  have hnon (t : ℝ) (ht : 0 ≤ t) : 0 ≤ D t := by
    have hn := hlip.norm_sub_le (x0+t • u) x0
    have he : ‖gradient f (x0+t • u)-gradient f x0‖ ≤ (beta : ℝ)*t*‖u‖ := by
      simpa [norm_smul, Real.norm_eq_abs, abs_of_nonneg ht, mul_assoc] using hn
    have hb := norm_inner_le_norm (𝕜 := ℝ) (gradient f (x0+t • u)-gradient f x0) u
    have hh : |inner ℝ (gradient f (x0+t • u)-gradient f x0) u| ≤
        (beta : ℝ)*t*‖u‖^2 := by
      rw [← Real.norm_eq_abs]
      calc
        _ ≤ _ := hb
        _ ≤ ((beta : ℝ)*t*‖u‖)*‖u‖ := mul_le_mul_of_nonneg_right he (norm_nonneg _)
        _ = _ := by ring
    have hx := (abs_le.mp hh).1
    dsimp [D]
    rw [inner_sub_left] at hx
    linarith
  have hmono : MonotoneOn H (Set.Icc (0 : ℝ) 1) :=
    monotoneOn_of_deriv_nonneg (convex_Icc _ _)
      ((continuous_iff_continuousAt.mpr (fun t => (hd t).continuousAt)).continuousOn)
      (fun t _ => (hd t).differentiableAt.differentiableWithinAt)
      (fun t ht => by
        rw [(hd t).deriv]
        exact hnon t ((interior_subset ht).1))
  have h01 := hmono (by norm_num : (0 : ℝ) ∈ Set.Icc 0 1)
    (by norm_num : (1 : ℝ) ∈ Set.Icc 0 1) zero_le_one
  simp only [H, zero_smul, add_zero, zero_mul, zero_pow (by decide : 2 ≠ 0),
    mul_zero, sub_zero, one_smul, one_mul, one_pow, mul_one] at h01
-- Source excerpt truncated; follow the exact source link.

AutoSamplingTheory/ExampleCases/SmoothedPicardHMC/IdealRGOIdentification.lean:30published source at 0e31a3cda412

Excerpt truncated; the exact source link is authoritative.

No declaration docstring.

private theorem regularized_lower (f : E → ℝ) (hf : Differentiable ℝ f)
    (beta : NNReal) (hlip : LipschitzWith beta (gradient f))
    (eta : ℝ) (heta : 0 < eta) (hstep : (beta : ℝ)*eta < 1) (x0 u : E) :
    let k := 1/eta - (beta : ℝ)
    0 < k ∧ f x0 - ‖gradient f x0‖^2/k + k/4*‖u‖^2 ≤
      f (x0+u) + ‖u‖^2/(2*eta) := by
  let k := 1/eta - (beta : ℝ)
  have hk : 0 < k := sub_pos.mpr ((lt_div_iff₀ heta).mpr hstep)
  refine ⟨hk, ?_⟩
  have hs := smooth_lower f hf beta hlip x0 u
  have hn := norm_inner_le_norm (𝕜 := ℝ) (gradient f x0) u
  have hin : -(‖gradient f x0‖*‖u‖) ≤ inner ℝ (gradient f x0) u := by
    exact (abs_le.mp (by simpa only [Real.norm_eq_abs] using hn)).1
  have hy := sq_nonneg (‖gradient f x0‖ - k/2*‖u‖)
  have hc : (‖gradient f x0‖^2/k)*k = ‖gradient f x0‖^2 :=
    div_mul_cancel₀ _ hk.ne'
  have hyoung : ‖gradient f x0‖*‖u‖ ≤ ‖gradient f x0‖^2/k + k/4*‖u‖^2 := by
    nlinarith
  have hweight : ‖u‖^2/(2*eta) = ((beta : ℝ)+k)/2*‖u‖^2 := by
    dsimp [k]
    field_simp
    ring
  change f x0 - ‖gradient f x0‖^2/k + k/4*‖u‖^2 ≤ _
  rw [hweight]
  nlinarith

variable [MeasurableSpace E] [BorelSpace E]

AutoSamplingTheory/ExampleCases/SmoothedPicardHMC/IdealRGOIdentification.lean:76published source at 0e31a3cda412

No declaration docstring.

private theorem ideal_volume_integrable (f : E → ℝ) (hf : Differentiable ℝ f)
    (beta : NNReal) (hlip : LipschitzWith beta (gradient f))
    (eta : ℝ) (heta : 0 < eta) (hstep : (beta : ℝ)*eta < 1) (x0 : E) :
    Integrable (fun x => Real.exp (-(f x + ‖x-x0‖^2/(2*eta)))) volume ∧
      0 < ∫ x, Real.exp (-(f x + ‖x-x0‖^2/(2*eta))) ∂volume := by
  let k := 1/eta - (beta : ℝ)
  have hk : 0 < k := (regularized_lower f hf beta hlip eta heta hstep x0 0).1
  have henv := AutoSamplingTheory.TechnicalLemmas.Analysis.Integrability.integrable_exp_neg_add_mul_norm_sub_sq
    (E := E) (a := k/4) (b := f x0 - ‖gradient f x0‖^2/k) x0 (by positivity)
  have hi : Integrable (fun x => Real.exp (-(f x + ‖x-x0‖^2/(2*eta)))) volume := by
    apply henv.mono'
    · exact (Real.continuous_exp.comp (hf.continuous.add (by fun_prop)).neg).aestronglyMeasurable
    · apply Filter.Eventually.of_forall
      intro x
      rw [Real.norm_eq_abs, abs_of_pos (Real.exp_pos _)]
      apply Real.exp_le_exp.mpr
      have hl := (regularized_lower f hf beta hlip eta heta hstep x0 (x-x0)).2
      have hx : x0 + (x-x0) = x := by abel
      rw [hx] at hl
      dsimp only [k] at *
      linarith
  exact ⟨hi, integral_exp_pos hi⟩

AutoSamplingTheory/ExampleCases/SmoothedPicardHMC/IdealRGOIdentification.lean:104published source at 0e31a3cda412

No declaration docstring.

private theorem translated_density (a : E) (eta : ℝ) (heta : 0 < eta) :
    (stdGaussian E).map (fun z => a + Real.sqrt eta • z) =
      (volume : Measure E).withDensity (fun z => ENNReal.ofReal
        (((Real.sqrt (2 * Real.pi * eta))⁻¹) ^ Module.finrank ℝ E *
          Real.exp (-‖z-a‖ ^ 2 / (2 * eta)))) := by
  have hmap : (stdGaussian E).map (fun z => a + Real.sqrt eta • z) =
      ((stdGaussian E).map (fun z : E => Real.sqrt eta • z)).map
        (MeasurableEquiv.addLeft a) := by
    rw [Measure.map_map (by fun_prop) (by fun_prop)]
    rfl
  rw [hmap,
    AutoSamplingTheory.TechnicalLemmas.Measure.IsotropicGaussianDensity.map_sqrt_smul_stdGaussian_eq_withDensity eta heta,
    AutoSamplingTheory.TechnicalLemmas.Measure.RadonNikodym.measurableEquiv_map_withDensity (MeasurableEquiv.addLeft a) _ (by fun_prop)]
  have hvol : (volume : Measure E).map (MeasurableEquiv.addLeft a) = volume :=
    Measure.IsAddLeftInvariant.map_add_left_eq_self a
  rw [hvol]
  congr 1
  funext z
  simp [MeasurableEquiv.addLeft, sub_eq_add_neg, add_comm]

AutoSamplingTheory/ExampleCases/SmoothedPicardHMC/IdealRGOIdentification.lean:127published source at 0e31a3cda412

No declaration docstring.

private theorem ideal_gaussian_integrable (f : E → ℝ) (hf : Differentiable ℝ f)
    (beta : NNReal) (hlip : LipschitzWith beta (gradient f))
    (eta : ℝ) (heta : 0 < eta) (hstep : (beta : ℝ)*eta < 1) (x0 : E) :
    let G := (stdGaussian E).map (fun z => x0 + Real.sqrt eta • z)
    Integrable (fun x => Real.exp (-f x)) G ∧
      0 < ∫ x, Real.exp (-f x) ∂G := by
  let G := (stdGaussian E).map (fun z => x0 + Real.sqrt eta • z)
  let c := ((Real.sqrt (2 * Real.pi * eta))⁻¹) ^ Module.finrank ℝ E
  have hc : 0 ≤ c := by positivity
  have hi : Integrable (fun x => Real.exp (-f x)) G := by
    dsimp only [G]
    rw [translated_density x0 eta heta]
    apply (integrable_withDensity_iff_integrable_smul' (by fun_prop)
      (Filter.Eventually.of_forall (fun _ => ENNReal.ofReal_lt_top))).mpr
    have hv := (ideal_volume_integrable f hf beta hlip eta heta hstep x0).1.const_mul c
    convert! hv using 1
    funext x
    rw [ENNReal.toReal_ofReal (mul_nonneg hc (Real.exp_pos _).le)]
    change c * Real.exp _ * Real.exp _ = c * Real.exp _
    rw [mul_assoc, ← Real.exp_add]
    congr 2
    ring
  let : IsProbabilityMeasure G := Measure.isProbabilityMeasure_map (by fun_prop)
  exact ⟨hi, integral_exp_pos hi⟩

omit [NormedAddCommGroup E] [InnerProductSpace ℝ E] [FiniteDimensional ℝ E] [BorelSpace E] in

AutoSamplingTheory/ExampleCases/SmoothedPicardHMC/IdealRGOIdentification.lean:147published source at 0e31a3cda412

No declaration docstring.

private theorem tilt_change_density (mu : Measure E) (w m n : E → ℝ)
    (hw : Measurable w) (hm : Measurable m)
    (hwpos : ∀ x, 0 ≤ w x) (A : ℝ) (hA : 0 < A)
    (heq : ∀ x, w x * Real.exp (m x) = A * Real.exp (n x))
    (hi : Integrable (fun x => Real.exp (n x)) mu)
    (hZ : 0 < ∫ x, Real.exp (n x) ∂mu) :
    let q := mu.withDensity (fun x => ENNReal.ofReal (w x))
    Integrable (fun x => Real.exp (m x)) q ∧
      (∫ x, Real.exp (m x) ∂q) = A * ∫ x, Real.exp (n x) ∂mu ∧
      0 < ∫ x, Real.exp (m x) ∂q ∧ q.tilted m = mu.tilted n := by
  let q := mu.withDensity (fun x => ENNReal.ofReal (w x))
  have hfun : (fun x => (ENNReal.ofReal (w x)).toReal • Real.exp (m x)) =
      (fun x => A * Real.exp (n x)) := by
    funext x
    simpa only [ENNReal.toReal_ofReal (hwpos x), smul_eq_mul] using heq x
  have hqi : Integrable (fun x => Real.exp (m x)) q := by
    apply (integrable_withDensity_iff_integrable_smul' hw.ennreal_ofReal
      (Filter.Eventually.of_forall (fun _ => ENNReal.ofReal_lt_top))).mpr
    rw [hfun]
    exact hi.const_mul A
  have hZi : (∫ x, Real.exp (m x) ∂q) = A * ∫ x, Real.exp (n x) ∂mu := by
    dsimp only [q]
    rw [integral_withDensity_eq_integral_toReal_smul hw.ennreal_ofReal
      (Filter.Eventually.of_forall (fun _ => ENNReal.ofReal_lt_top)), hfun,
      integral_const_mul]
  refine ⟨hqi, hZi, by rw [hZi]; exact mul_pos hA hZ, ?_⟩
  rw [Measure.tilted, ← withDensity_mul _ hw.ennreal_ofReal
    ((hm.exp.div_const _).ennreal_ofReal)]
  change mu.withDensity _ = mu.withDensity _
  congr 1
  funext x
  simp only [Pi.mul_apply]
  rw [← ENNReal.ofReal_mul (hwpos x)]
  congr 1
  change w x * (Real.exp (m x) / ∫ y, Real.exp (m y) ∂q) = _
  rw [hZi, ← mul_div_assoc, heq x]
  exact mul_div_mul_left _ _ hA.ne'

AutoSamplingTheory/ExampleCases/SmoothedPicardHMC/IdealRGOIdentification.lean:173published source at 0e31a3cda412

No declaration docstring.

private theorem affine_gaussian_tilt (f : E → ℝ) (hf : Differentiable ℝ f)
    (beta : NNReal) (hlip : LipschitzWith beta (gradient f))
    (eta : ℝ) (heta : 0 < eta) (hstep : (beta : ℝ)*eta < 1)
    (x0 g : E) (C : ℝ) :
    let G := (stdGaussian E).map (fun z => x0 + Real.sqrt eta • z)
    let q := (stdGaussian E).map (fun z => (x0-eta • g) + Real.sqrt eta • z)
    let m := fun x => inner ℝ g x - f x + C
    Integrable (fun x => Real.exp (m x)) q ∧
      0 < ∫ x, Real.exp (m x) ∂q ∧ q.tilted m = G.tilted (fun x => -f x) := by
  let G := (stdGaussian E).map (fun z => x0 + Real.sqrt eta • z)
  let q := (stdGaussian E).map (fun z => (x0-eta • g) + Real.sqrt eta • z)
  let m := fun x => inner ℝ g x - f x + C
  let w := fun x => Real.exp
    (inner ℝ ((x0-eta • g)-x0) (x-x0)/eta - ‖(x0-eta • g)-x0‖^2/(2*eta))
  let A := Real.exp (C + inner ℝ g x0 - eta*‖g‖^2/2)
  have hg := AutoSamplingTheory.TechnicalLemmas.Measure.GaussianLikelihood.translated_gaussian_likelihood
    (x0-eta • g) x0 eta heta
  have hq : q = G.withDensity (fun x => ENNReal.ofReal (w x)) := hg.2.1
  have heq (x : E) : w x * Real.exp (m x) = A * Real.exp (-f x) := by
    dsimp only [w, m, A]
    rw [← Real.exp_add, ← Real.exp_add]
    congr 1
    have hs : (x0-eta • g)-x0 = (-eta) • g := by module
    rw [hs, real_inner_smul_left, inner_sub_right, norm_smul,
      Real.norm_eq_abs, mul_pow, sq_abs]
    field_simp
    ring
  have hbase := ideal_gaussian_integrable f hf beta hlip eta heta hstep x0
  have ht := tilt_change_density G w m (fun x => -f x) hg.1
    (by dsimp [m]; exact ((by fun_prop : Measurable (fun x => inner ℝ g x)).sub hf.continuous.measurable).add_const C)
    (fun x => (Real.exp_pos _).le) A (Real.exp_pos _) heq hbase.1 hbase.2
  rw [← hq] at ht
  exact ⟨ht.1, ht.2.2.1, ht.2.2.2⟩

AutoSamplingTheory/ExampleCases/SmoothedPicardHMC/IdealRGOIdentification.lean:211published source at 0e31a3cda412

No declaration docstring.

private theorem ideal_volume_law (f : E → ℝ) (hf : Differentiable ℝ f)
    (beta : NNReal) (hlip : LipschitzWith beta (gradient f))
    (eta : ℝ) (heta : 0 < eta) (hstep : (beta : ℝ)*eta < 1) (x0 : E) :
    ((stdGaussian E).map (fun z => x0 + Real.sqrt eta • z)).tilted (fun x => -f x) =
      (volume : Measure E).withDensity (fun x => ENNReal.ofReal
        (Real.exp (-(f x + ‖x-x0‖^2/(2*eta))) /
          ∫ y, Real.exp (-(f y + ‖y-x0‖^2/(2*eta))) ∂volume)) := by
  let c := ((Real.sqrt (2 * Real.pi * eta))⁻¹) ^ Module.finrank ℝ E
  let w := fun x : E => c * Real.exp (-‖x-x0‖^2/(2*eta))
  have hc : 0 < c := by dsimp [c]; positivity
  have hprod (x : E) : w x * Real.exp (-f x) =
      c * Real.exp (-(f x + ‖x-x0‖^2/(2*eta))) := by
    dsimp [w]
    rw [mul_assoc, ← Real.exp_add]
    congr 2
    ring
  have hv := ideal_volume_integrable f hf beta hlip eta heta hstep x0
  have ht := tilt_change_density (volume : Measure E) w (fun x => -f x)
    (fun x => -(f x + ‖x-x0‖^2/(2*eta))) (by dsimp [w]; fun_prop)
    hf.continuous.measurable.neg (fun x => mul_nonneg hc.le (Real.exp_pos _).le)
    c hc hprod hv.1 hv.2
  rw [translated_density x0 eta heta]
  exact ht.2.2.2

AutoSamplingTheory/ExampleCases/SmoothedPicardHMC/IdealRGOIdentification.lean:245published source at 0e31a3cda412

No declaration docstring.

private def estimator (f : E → ℝ) (h xp x : E) (s : ℝ × E) : ℝ :=
  inner ℝ ((Real.pi / 2) • (Real.cos (Real.pi / 2 * s.1) • (x-h) -
    Real.sin (Real.pi / 2 * s.1) • s.2))
    (gradient f xp - gradient f (h + Real.sin (Real.pi / 2 * s.1) • (x-h) +
      Real.cos (Real.pi / 2 * s.1) • s.2))

AutoSamplingTheory/ExampleCases/SmoothedPicardHMC/IdealRGOIdentification.lean:269published source at 0e31a3cda412

No declaration docstring.

private theorem actual_mean_tilt (f : E → ℝ) (hf : Differentiable ℝ f)
    (beta : NNReal) (hlip : LipschitzWith beta (gradient f))
    (eta : ℝ) (heta : 0 < eta) (hstep : (beta : ℝ)*eta < 1) (x0 xp : E) :
    let h := x0-eta • gradient f xp
    let P := (stdGaussian E).map (fun z => Real.sqrt eta • z)
    let nu := (volume.restrict (Set.Ioc (0 : ℝ) 1)).prod P
    let q := (stdGaussian E).map (fun z => h + Real.sqrt eta • z)
    let m := fun x => ∫ s, estimator f h xp x s ∂nu
    Integrable (fun x => Real.exp (m x)) q ∧
      0 < ∫ x, Real.exp (m x) ∂q ∧
      q.tilted m = ((stdGaussian E).map (fun z => x0 + Real.sqrt eta • z)).tilted
        (fun x => -f x) := by
  let h := x0-eta • gradient f xp
  let P := (stdGaussian E).map (fun z => Real.sqrt eta • z)
  let nu := (volume.restrict (Set.Ioc (0 : ℝ) 1)).prod P
  let C := (∫ z, f (h+z) ∂P) - inner ℝ (gradient f xp) h
  have hmean := AutoSamplingTheory.ExampleCases.SmoothedPicardHMC.GradientArcMean.gradient_arc_mean
    f hf eta (beta : ℝ) heta beta.coe_nonneg hlip h xp
  have hm : (fun x => ∫ s, estimator f h xp x s ∂nu) =
      (fun x => inner ℝ (gradient f xp) x - f x + C) := by
    funext x
    exact (hmean.2.2.2 x).2.2.2
  change Integrable (fun x => Real.exp ((fun x => ∫ s, estimator f h xp x s ∂nu) x)) _ ∧ _
  rw [hm]
  exact affine_gaussian_tilt f hf beta hlip eta heta hstep x0 (gradient f xp) C

/-- Positive finite normalization and equality of the actual ideal RGO laws. -/

AutoSamplingTheory/ExampleCases/SmoothedPicardHMC/IdealRGOIdentification.lean:275published source at 0e31a3cda412

- Positive finite normalization and equality of the actual ideal RGO laws.

theorem ideal_rgo_identification (f : E → ℝ) (hf : Differentiable ℝ f)
    (beta : NNReal) (hlip : LipschitzWith beta (gradient f))
    (eta : ℝ) (heta : 0 < eta) (hstep : (beta : ℝ)*eta < 1) (x0 xp : E) :
    let h := x0-eta • gradient f xp
    let P := (stdGaussian E).map (fun z => Real.sqrt eta • z)
    let nu := (volume.restrict (Set.Ioc (0 : ℝ) 1)).prod P
    let G := (stdGaussian E).map (fun z => x0 + Real.sqrt eta • z)
    let q := (stdGaussian E).map (fun z => h + Real.sqrt eta • z)
    let m := fun x => ∫ s, estimator f h xp x s ∂nu
    let V := fun x => f x + ‖x-x0‖^2/(2*eta)
    Integrable (fun x => Real.exp (-f x)) G ∧
      0 < ∫ x, Real.exp (-f x) ∂G ∧
      Integrable (fun x => Real.exp (m x)) q ∧
      0 < ∫ x, Real.exp (m x) ∂q ∧
      Integrable (fun x => Real.exp (-V x)) volume ∧
      0 < ∫ x, Real.exp (-V x) ∂volume ∧
      q.tilted m = G.tilted (fun x => -f x) ∧
      q.tilted m = (volume : Measure E).withDensity (fun x => ENNReal.ofReal
        (Real.exp (-V x) / ∫ y, Real.exp (-V y) ∂volume)) := by
  have hg := ideal_gaussian_integrable f hf beta hlip eta heta hstep x0
  have hm := actual_mean_tilt f hf beta hlip eta heta hstep x0 xp
  have hv := ideal_volume_integrable f hf beta hlip eta heta hstep x0
  exact ⟨hg.1, hg.2, hm.1, hm.2.1, hv.1, hv.2, hm.2.2,
    hm.2.2.trans (ideal_volume_law f hf beta hlip eta heta hstep x0)⟩

end AutoSamplingTheory.ExampleCases.SmoothedPicardHMC.IdealRGOIdentification

AutoSamplingTheory/ExampleCases/SmoothedPicardHMC/IdealRGOIdentification.lean:302published source at 0e31a3cda412