English
If K is complete, then D f K is contained in {x | f differentiable at x and f'(x) ∈ K}.
Русский
Если K полно, то D(f,K) ⊆ {x | f дифференцируема в x и f'(x) ∈ K}.
LaTeX
$$$$D(f,K) \subseteq \{x \mid \text{DifferentiableAt } 𝕜\; f\; x \wedge fderiv 𝕜 f x \in K\}.$$$$
Lean4
/-- Harder inclusion: at a point in `D f K`, the function `f` has a derivative, in `K`. -/
theorem D_subset_differentiable_set {K : Set (E →L[𝕜] F)} (hK : IsComplete K) :
D f K ⊆ {x | DifferentiableAt 𝕜 f x ∧ fderiv 𝕜 f x ∈ K} :=
by
have P : ∀ {n : ℕ}, (0 : ℝ) < (1 / 2) ^ n := fun {n} => pow_pos (by norm_num) n
rcases NormedField.exists_one_lt_norm 𝕜 with ⟨c, hc⟩
intro x hx
have :
∀ e : ℕ,
∃ n : ℕ,
∀ p q, n ≤ p → n ≤ q → ∃ L ∈ K, x ∈ A f L ((1 / 2) ^ p) ((1 / 2) ^ e) ∩ A f L ((1 / 2) ^ q) ((1 / 2) ^ e) :=
by
intro e
have := mem_iInter.1 hx e
rcases mem_iUnion.1 this with ⟨n, hn⟩
refine ⟨n, fun p q hp hq => ?_⟩
simp only [mem_iInter] at hn
rcases mem_iUnion.1 (hn p hp q hq) with ⟨L, hL⟩
exact
⟨L, exists_prop.mp <| mem_iUnion.1 hL⟩
/- Recast the assumptions: for each `e`, there exist `n e` and linear maps `L e p q` in `K`
such that, for `p, q ≥ n e`, then `f` is well approximated by `L e p q` at scale `2 ^ (-p)` and
`2 ^ (-q)`, with an error `2 ^ (-e)`. -/
choose! n L hn using this
have M :
∀ e p q e' p' q',
n e ≤ p → n e ≤ q → n e' ≤ p' → n e' ≤ q' → e ≤ e' → ‖L e p q - L e' p' q'‖ ≤ 12 * ‖c‖ * (1 / 2) ^ e :=
by
intro e p q e' p' q' hp hq hp' hq' he'
let r := max (n e) (n e')
have I : ((1 : ℝ) / 2) ^ e' ≤ (1 / 2) ^ e := pow_le_pow_of_le_one (by norm_num) (by norm_num) he'
have J1 : ‖L e p q - L e p r‖ ≤ 4 * ‖c‖ * (1 / 2) ^ e :=
by
have I1 : x ∈ A f (L e p q) ((1 / 2) ^ p) ((1 / 2) ^ e) := (hn e p q hp hq).2.1
have I2 : x ∈ A f (L e p r) ((1 / 2) ^ p) ((1 / 2) ^ e) := (hn e p r hp (le_max_left _ _)).2.1
exact norm_sub_le_of_mem_A hc P P I1 I2
have J2 : ‖L e p r - L e' p' r‖ ≤ 4 * ‖c‖ * (1 / 2) ^ e :=
by
have I1 : x ∈ A f (L e p r) ((1 / 2) ^ r) ((1 / 2) ^ e) := (hn e p r hp (le_max_left _ _)).2.2
have I2 : x ∈ A f (L e' p' r) ((1 / 2) ^ r) ((1 / 2) ^ e') := (hn e' p' r hp' (le_max_right _ _)).2.2
exact norm_sub_le_of_mem_A hc P P I1 (A_mono _ _ I I2)
have J3 : ‖L e' p' r - L e' p' q'‖ ≤ 4 * ‖c‖ * (1 / 2) ^ e :=
by
have I1 : x ∈ A f (L e' p' r) ((1 / 2) ^ p') ((1 / 2) ^ e') := (hn e' p' r hp' (le_max_right _ _)).2.1
have I2 : x ∈ A f (L e' p' q') ((1 / 2) ^ p') ((1 / 2) ^ e') := (hn e' p' q' hp' hq').2.1
exact norm_sub_le_of_mem_A hc P P (A_mono _ _ I I1) (A_mono _ _ I I2)
calc
‖L e p q - L e' p' q'‖ = ‖L e p q - L e p r + (L e p r - L e' p' r) + (L e' p' r - L e' p' q')‖ := by congr 1;
abel
_ ≤ ‖L e p q - L e p r‖ + ‖L e p r - L e' p' r‖ + ‖L e' p' r - L e' p' q'‖ := norm_add₃_le
_ ≤ 4 * ‖c‖ * (1 / 2) ^ e + 4 * ‖c‖ * (1 / 2) ^ e + 4 * ‖c‖ * (1 / 2) ^ e := by gcongr
_ = 12 * ‖c‖ * (1 / 2) ^ e := by
ring
/- For definiteness, use `L0 e = L e (n e) (n e)`, to have a single sequence. We claim that this
is a Cauchy sequence. -/
let L0 : ℕ → E →L[𝕜] F := fun e => L e (n e) (n e)
have : CauchySeq L0 := by
rw [Metric.cauchySeq_iff']
intro ε εpos
obtain ⟨e, he⟩ : ∃ e : ℕ, (1 / 2) ^ e < ε / (12 * ‖c‖) := exists_pow_lt_of_lt_one (by positivity) (by norm_num)
refine ⟨e, fun e' he' => ?_⟩
rw [dist_comm, dist_eq_norm]
calc
‖L0 e - L0 e'‖ ≤ 12 * ‖c‖ * (1 / 2) ^ e := M _ _ _ _ _ _ le_rfl le_rfl le_rfl le_rfl he'
_ < 12 * ‖c‖ * (ε / (12 * ‖c‖)) := by gcongr
_ = ε := by
field_simp
-- As it is Cauchy, the sequence `L0` converges, to a limit `f'` in `K`.
obtain ⟨f', f'K, hf'⟩ : ∃ f' ∈ K, Tendsto L0 atTop (𝓝 f') :=
cauchySeq_tendsto_of_isComplete hK (fun e => (hn e (n e) (n e) le_rfl le_rfl).1) this
have Lf' : ∀ e p, n e ≤ p → ‖L e (n e) p - f'‖ ≤ 12 * ‖c‖ * (1 / 2) ^ e :=
by
intro e p hp
apply le_of_tendsto (tendsto_const_nhds.sub hf').norm
rw [eventually_atTop]
exact
⟨e, fun e' he' => M _ _ _ _ _ _ le_rfl hp le_rfl le_rfl he'⟩
-- Let us show that `f` has derivative `f'` at `x`.
have : HasFDerivAt f f' x :=
by
simp only [hasFDerivAt_iff_isLittleO_nhds_zero, isLittleO_iff]
/- to get an approximation with a precision `ε`, we will replace `f` with `L e (n e) m` for
some large enough `e` (yielding a small error by uniform approximation). As one can vary `m`,
this makes it possible to cover all scales, and thus to obtain a good linear approximation in
the whole ball of radius `(1/2)^(n e)`. -/
intro ε εpos
have pos : 0 < 4 + 12 * ‖c‖ := by positivity
obtain ⟨e, he⟩ : ∃ e : ℕ, (1 / 2) ^ e < ε / (4 + 12 * ‖c‖) :=
exists_pow_lt_of_lt_one (div_pos εpos pos) (by norm_num)
rw [eventually_nhds_iff_ball]
refine
⟨(1 / 2) ^ (n e + 1), P, fun y hy => ?_⟩
-- We need to show that `f (x + y) - f x - f' y` is small. For this, we will work at scale
-- `k` where `k` is chosen with `‖y‖ ∼ 2 ^ (-k)`.
by_cases y_pos : y = 0
· simp [y_pos]
have yzero : 0 < ‖y‖ := norm_pos_iff.mpr y_pos
have y_lt : ‖y‖ < (1 / 2) ^ (n e + 1) := by simpa using mem_ball_iff_norm.1 hy
have yone : ‖y‖ ≤ 1 :=
le_trans y_lt.le
(pow_le_one₀ (by norm_num) (by norm_num))
-- define the scale `k`.
obtain ⟨k, hk, h'k⟩ : ∃ k : ℕ, (1 / 2) ^ (k + 1) < ‖y‖ ∧ ‖y‖ ≤ (1 / 2) ^ k :=
exists_nat_pow_near_of_lt_one yzero yone (by norm_num : (0 : ℝ) < 1 / 2)
(by norm_num : (1 : ℝ) / 2 < 1)
-- the scale is large enough (as `y` is small enough)
have k_gt : n e < k :=
by
have : ((1 : ℝ) / 2) ^ (k + 1) < (1 / 2) ^ (n e + 1) := lt_trans hk y_lt
rw [pow_lt_pow_iff_right_of_lt_one₀ (by norm_num : (0 : ℝ) < 1 / 2) (by norm_num)] at this
omega
set m := k - 1
have m_ge : n e ≤ m := Nat.le_sub_one_of_lt k_gt
have km : k = m + 1 := (Nat.succ_pred_eq_of_pos (lt_of_le_of_lt (zero_le _) k_gt)).symm
rw [km] at hk h'k
have J1 : ‖f (x + y) - f x - L e (n e) m (x + y - x)‖ ≤ (1 / 2) ^ e * (1 / 2) ^ m :=
by
apply le_of_mem_A (hn e (n e) m le_rfl m_ge).2.2
· simp only [mem_closedBall, dist_self]
positivity
· simpa only [dist_eq_norm, add_sub_cancel_left, mem_closedBall, pow_succ, mul_one_div] using h'k
have J2 : ‖f (x + y) - f x - L e (n e) m y‖ ≤ 4 * (1 / 2) ^ e * ‖y‖ :=
calc
‖f (x + y) - f x - L e (n e) m y‖ ≤ (1 / 2) ^ e * (1 / 2) ^ m := by simpa only [add_sub_cancel_left] using J1
_ = 4 * (1 / 2) ^ e * (1 / 2) ^ (m + 2) := by field_simp; ring
_ ≤ 4 * (1 / 2) ^ e * ‖y‖ := by
gcongr
-- use the previous estimates to see that `f (x + y) - f x - f' y` is small.
calc
‖f (x + y) - f x - f' y‖ = ‖f (x + y) - f x - L e (n e) m y + (L e (n e) m - f') y‖ := congr_arg _ (by simp)
_ ≤ 4 * (1 / 2) ^ e * ‖y‖ + 12 * ‖c‖ * (1 / 2) ^ e * ‖y‖ :=
(norm_add_le_of_le J2 <| (le_opNorm _ _).trans <| by gcongr; exact Lf' _ _ m_ge)
_ = (4 + 12 * ‖c‖) * ‖y‖ * (1 / 2) ^ e := by ring
_ ≤ (4 + 12 * ‖c‖) * ‖y‖ * (ε / (4 + 12 * ‖c‖)) := by gcongr
_ = ε * ‖y‖ := by field_simp
rw [← this.fderiv] at f'K
exact ⟨this.differentiableAt, f'K⟩
