Documentation

TauCeti.Analysis.ODE.Gronwall

Grönwall-type estimates #

Mathlib's Grönwall inequality (le_gronwallBound_of_liminf_deriv_right_le) bounds a function by a bound on its derivative. Energy estimates for evolution equations instead produce a bound on a function by an integral of itself: a continuous φ with φ t ≤ c + K ∫ₐᵗ φ on [a, b], for some K ≥ 0. This file proves that such a φ satisfies φ t ≤ c exp (K (t - a)) (TauCeti.le_mul_exp_of_le_add_mul_integral).

The hypothesis K ≥ 0 cannot be dropped: a very negative stretch of φ makes c + K ∫ₐᵗ φ large when K < 0.

For a monotone g the differential form needs neither continuity nor differentiability: if k g ≤ g' at almost every point of [s, t] and g s ≥ 0, then g s exp (∫ₛᵗ k) ≤ g t (TauCeti.mul_exp_integral_le_of_mul_le_deriv), because the derivative of the monotone function log g integrates to at most its increment.

It also writes Mathlib's gronwallBound with a negative rate -λ in terms of the exponential integral E_λ(T) = ∫₀ᵀ e^{λr} dr (TauCeti.exp_mul_gronwallBound_neg), and records that E_λ(T) > 0 for T > 0 (TauCeti.integral_exp_mul_pos), that z ↦ E_λ(z - t) has derivative 1 at t (TauCeti.hasDerivAt_integral_exp_mul_sub), and E_λ(T) ≤ T e^{|λ|T} (TauCeti.integral_exp_mul_le).

Finally, it specializes Mathlib's inequality to exponential decay: a function whose derivative satisfies g' ≤ -λ g on [T, ∞) is bounded by g T exp (-λ (t - T)) there (TauCeti.le_mul_exp_of_hasDerivAt_le_neg_mul), hence by C exp (-λ t) for large t (TauCeti.exists_le_mul_exp_atTop_of_hasDerivAt_le_neg_mul).

Mathlib's uniqueness statement eq_zero_of_abs_deriv_le_mul_abs_self_of_eq_zero_right propagates the vanishing of a function with ‖f'‖ ≤ K ‖f‖ forward from the left endpoint of an interval. Reversing time gives the two-sided version: on [a, b] such a function vanishes identically as soon as it vanishes at one point (TauCeti.eq_zero_of_norm_deriv_le_mul_norm_self_of_mem_Icc). The general closed-interval bound is supplied by TauCeti.norm_le_gronwallBound_of_norm_derivWithin_le; its interior-derivative variant is TauCeti.norm_le_gronwallBound_of_norm_deriv_le. This is the uniqueness argument for linear differential equations by energy estimates. Averaging the reversed Grönwall bound over [s, s + 1] shows that such a function, if it is square integrable on the line, is bounded by its L² norm: ‖f s‖² ≤ exp (2 K) ∫ ‖f‖² (TauCeti.norm_sq_le_exp_mul_integral_norm_sq).

References #

theorem TauCeti.le_mul_exp_of_le_add_mul_integral {φ : ℝ → ℝ} {a b c K : ℝ} (hφ : ContinuousOn φ (Set.Icc a b)) (hK : 0 ≤ K) (h : ∀ t ∈ Set.Icc a b, φ t ≤ c + K * ∫ (s : ℝ) in a..t, φ s) (t : ℝ) :
t ∈ Set.Icc a b → φ t ≤ c * Real.exp (K * (t - a))

Grönwall's inequality, integral form. If φ is continuous on [a, b], 0 ≤ K and φ t ≤ c + K ∫ₐᵗ φ for every t ∈ [a, b], then φ t ≤ c exp (K (t - a)) on [a, b].

theorem TauCeti.mul_exp_integral_le_of_mul_le_deriv {g k : ℝ → ℝ} {s t : ℝ} (hst : s ≤ t) (hg : MonotoneOn g (Set.Icc s t)) (hgs : 0 ≤ g s) (h : ∀ᵐ (σ : ℝ), σ ∈ Set.Ioo s t → k σ * g σ ≤ deriv g σ) :
g s * Real.exp (∫ (σ : ℝ) in s..t, k σ) ≤ g t

Grönwall's inequality for monotone functions, almost-everywhere form. Let g be monotone on [s, t] with g s ≥ 0. If k g ≤ g' at almost every point of (s, t), then g s · exp (∫ₛᵗ k) ≤ g t. No continuity of g is needed. If k is not interval integrable, the integral is the junk value 0 and the bound is just monotonicity.

theorem TauCeti.exp_mul_gronwallBound_neg (δ lam c T : ℝ) :
Real.exp (lam * T) * gronwallBound δ (-lam) c T = δ + (∫ (r : ℝ) in 0..T, Real.exp (lam * r)) * c

The closed form of e^{λT} times the Grönwall bound with rate -λ: e^{λT} gronwallBound δ (-λ) c T = δ + E_λ(T) c, where E_λ(T) = ∫₀ᵀ e^{λr} dr.

theorem TauCeti.integral_exp_mul_pos (lam : ℝ) {T : ℝ} (hT : 0 < T) :
0 < ∫ (r : ℝ) in 0..T, Real.exp (lam * r)

The function E_λ(T) = ∫₀ᵀ e^{λr} dr is positive for T > 0.

theorem TauCeti.hasDerivAt_integral_exp_mul_sub (lam t : ℝ) :
HasDerivAt (fun (z : ℝ) => ∫ (r : ℝ) in 0..z - t, Real.exp (lam * r)) 1 t

The map z ↦ ∫ r in 0..(z - t), exp (lam * r) has derivative 1 at t.

theorem TauCeti.integral_exp_mul_le (lam T : ℝ) :
∫ (r : ℝ) in 0..T, Real.exp (lam * r) ≤ T * Real.exp (|lam| * T)

The function E_λ(T) = ∫₀ᵀ e^{λr} dr satisfies E_λ(T) ≤ T e^{|λ|T}.

theorem TauCeti.le_mul_exp_of_hasDerivAt_le_neg_mul {g g' : ℝ → ℝ} {T lam : ℝ} (hg : ∀ s ∈ Set.Ici T, HasDerivAt g (g' s) s) (hle : ∀ s ∈ Set.Ici T, g' s ≤ -lam * g s) {t : ℝ} (ht : T ≤ t) :
g t ≤ g T * Real.exp (-(lam * (t - T)))

Exponential decay from a differential inequality. If g' ≤ -lam * g on [T, ∞), then g t ≤ g T * exp (-(lam * (t - T))) for every t ≥ T.

theorem TauCeti.exists_le_mul_exp_atTop_of_hasDerivAt_le_neg_mul {g g' : ℝ → ℝ} {T lam : ℝ} (hg : ∀ s ∈ Set.Ici T, HasDerivAt g (g' s) s) (hle : ∀ s ∈ Set.Ici T, g' s ≤ -lam * g s) :
∃ C > 0, ∀ᶠ (t : ℝ) in Filter.atTop, g t ≤ C * Real.exp (-(lam * t))

Exponential decay from a differential inequality, eventual form. If g' ≤ -lam * g on [T, ∞), then g t ≤ C * exp (-(lam * t)) for some C > 0 and all large t.

theorem TauCeti.eq_zero_of_norm_deriv_le_mul_norm_self_of_mem_Icc {E : Type u_1} [NormedAddCommGroup E] [NormedSpace ℝ E] {f f' : ℝ → E} {K a b c : ℝ} (hf : ∀ x ∈ Set.Icc a b, HasDerivWithinAt f (f' x) (Set.Icc a b) x) (bound : ∀ x ∈ Set.Icc a b, ‖f' x‖ ≤ K * ‖f x‖) (hc : c ∈ Set.Icc a b) (hfc : f c = 0) (x : ℝ) :
x ∈ Set.Icc a b → f x = 0

Two-sided uniqueness from a Grönwall bound. If f is differentiable within [a, b], satisfies ‖f'‖ ≤ K ‖f‖ there and vanishes at one point of [a, b], then it vanishes on [a, b]. Mathlib's eq_zero_of_abs_deriv_le_mul_abs_self_of_eq_zero_right is the case where the point is a.

theorem TauCeti.norm_sq_le_exp_mul_integral_norm_sq {F : Type u_1} [NormedAddCommGroup F] [NormedSpace ℝ F] {u u' : ℝ → F} {K : ℝ} (hK : 0 ≤ K) (hu : ∀ (s : ℝ), HasDerivAt u (u' s) s) (hu' : ∀ (s : ℝ), ‖u' s‖ ≤ K * ‖u s‖) (hu2 : MeasureTheory.MemLp u 2 MeasureTheory.volume) (s : ℝ) :
‖u s‖ ^ 2 ≤ Real.exp (2 * K) * ∫ (t : ℝ), ‖u t‖ ^ 2

A square-integrable path with derivative controlled by its norm is bounded. If ‖u'‖ ≤ K ‖u‖ with K ≥ 0 and u is square integrable on the line, then ‖u s‖² ≤ exp (2 K) ∫ ‖u‖² for every s.