feat(analysis/inner_product_space/reproducing_kernel): Reproducing kernel Hilbert spaces

src/analysis/inner_product_space/adjoint.lean

@@ -100,8 +100,12 @@ linear_isometry_equiv.of_surjective
   ..adjoint_aux }
 (λ A, ⟨adjoint_aux A, adjoint_aux_adjoint_aux A⟩)
 
+
 localized "postfix (name := adjoint) `†`:1000 := continuous_linear_map.adjoint" in inner_product
 
+@[simp] lemma adjoint_apply (A : E →L[𝕜] F) (x : F) :
+  A† x = ((to_dual 𝕜 E).symm : (normed_space.dual 𝕜 E) → E) ((to_sesq_form A) x) := rfl
+
 /-- The fundamental property of the adjoint. -/
 lemma adjoint_inner_left (A : E →L[𝕜] F) (x : E) (y : F) : ⟪A† y, x⟫ = ⟪y, A x⟫ :=
 adjoint_aux_inner_left A x y

src/analysis/inner_product_space/reproducing_kernel.lean

@@ -0,0 +1,256 @@
+/-
+Copyright (c) 2022 Shing Tak Lam. All rights reserved.
+Released under Apache 2.0 license as described in the file LICENSE.
+Authors: Shing Tak Lam
+-/
+
+import analysis.inner_product_space.projection
+import analysis.inner_product_space.positive
+
+/-!
+# Reproducing Kernel Hilbert Space
+
+A Hilbert space `H` is called a reproducing kernel Hilbert space (RKHS) if it is a vector subspace
+of `X → V` where `V` is a Hilbert space, and the evaluation map `H → V` at each `X` is continuous.
+
+## Main definitions
+
+* `rkhs 𝕜 X V H` - `H` is a RKHS, where `𝕜` is the scalar field, `X` is the domain, `V` is the
+  Hilbert space.
+* `rkhs.kernel` - the kernel of a RKHS, as a continuous linear map `V →L[𝕜] V`
+* `rkhs.eval'` - For `V = 𝕜`, by the Riesz representation theorem, the evaluation map can be
+  represented as the inner product with a point in `H`, which we call the `eval'`
+* `rkhs.scalar_kernel` - For `V = 𝕜`, we can represent the kernel of `H` as an inner product with a
+  point in `H`, which we call the `scalar_kernel`
+* `rkhs.eval'_span` - The span of the `eval'`s as a subspace of `H`.
+
+## Main statements
+
+* `rkhs.eval'_span_topological_closure_eq_top` - the topological closude of the span of the
+  `eval's` is the whole space `H`.
+* `rkhs.tendsto_nhds_apply` - Convergence in `H` implies pointwise convergence.
+
+## Notes
+
+Due to the statement of the Riesz representation theorem in mathlib, the order of `x` and `y` in
+the definition of the `scalar_kernel` and the representation of it as an inner product of
+`eval'`s is reversed. I opted to keep the statement simpler, but this leaves the order
+reversed.
+
+## Tags
+
+Reproducing kernel Hilbert space, RKHS
+-/
+
+noncomputable theory
+
+open_locale inner_product topological_space
+open inner_product_space filter
+
+variables (𝕜 : Type*) [is_R_or_C 𝕜]
+variables (X : Type*)
+variables (V : Type*) [inner_product_space 𝕜 V]
+
+/--
+A reproducing kernel Hilbert space (RKHS) is a vector subspace of `X → V` where evaluation at a
+point is continuous.
+-/
+class rkhs (H : Type*) [inner_product_space 𝕜 H] extends fun_like H X (λ _, V) :=
+(add_apply' : ∀ {f g : H} {x : X}, (f + g) x = f x + g x)
+(smul_apply' : ∀ {c : 𝕜} {f : H} {x : X}, (c • f) x = c • f x)
+(continuous_eval' : ∀ (x : X), continuous (λ (f : H), f x))
+
+namespace rkhs
+
+variables {𝕜 X V}
+variables {H : Type*} [inner_product_space 𝕜 H] [hHrkhs : rkhs 𝕜 X V H]
+
+include hHrkhs
+
+@[simp]
+lemma add_apply (f g : H) (x : X) : (f + g) x = f x + g x := rkhs.add_apply'
+
+@[simp]
+lemma smul_apply (c : 𝕜) (f : H) (x : X) : (c • f) x = c • f x := rkhs.smul_apply'
+
+@[simp]
+lemma zero_apply (x : X) : (0 : H) x = 0 := by rw [←zero_smul 𝕜 (0 : H), smul_apply, zero_smul]
+
+section defs
+
+variables (𝕜 V H)
+
+/--
+The embedding of an element `f : H` of an RKHS into `X → V`
+-/
+def to_pi : H →ₗ[𝕜] (X → V) :=
+{ to_fun := λ f, f,
+  map_add' := λ f g, funext $ add_apply f g,
+  map_smul' := λ c f, funext $ smul_apply c f }
+
+/--
+Evaluation at a point `x` as a bundled continuous linear map `H →L[𝕜] V`
+-/
+def eval (x : X) : H →L[𝕜] V :=
+{ to_fun := λ f, f x,
+  map_add' := λ f g, add_apply f g x,
+  map_smul' := λ c f, smul_apply c f x,
+  cont := continuous_eval' x }
+
+end defs
+
+@[simp]
+lemma to_pi_apply (f : H) (x : X) : to_pi 𝕜 V H f x = f x := rfl
+
+@[simp]
+lemma eval_apply (f : H) (x : X) : eval 𝕜 V H x f = f x := rfl
+
+lemma to_pi_inj : function.injective (to_pi 𝕜 V H : H → (X → V)) :=
+λ f g h, fun_like.coe_injective $ funext $ congr_fun h
+
+section complete_space
+
+variables (𝕜 V H) [complete_space V] [complete_space H]
+
+/--
+The kernel of an RKHS, as a continuous linear function `V →L[𝕜] V`
+-/
+def kernel (x y : X) : V →L[𝕜] V := (eval 𝕜 V H x).comp $ (eval 𝕜 V H y)†
+
+variables {𝕜 V H}
+
+lemma kernel_def (x y : X) : kernel 𝕜 V H x y = (eval 𝕜 V H x).comp ((eval 𝕜 V H y)†) := rfl
+
+end complete_space
+
+/-!
+## Scalar functionals
+
+In this section, we focus on the case `V = 𝕜`.
+-/
+
+section scalar
+
+omit hHrkhs
+
+section defs
+
+variables (𝕜 X)
+variables (H₁ : Type*) [inner_product_space 𝕜 H₁] [complete_space H₁] [hH₁rkhs : rkhs 𝕜 X 𝕜 H₁]
+
+include hH₁rkhs
+
+/--
+Evaluation at a point `x`, represented as a point `eval' x` in the RKHS `H₁` by the Riesz
+representation theorem.
+-/
+def eval' (x : X) : H₁ := (to_dual 𝕜 H₁).symm (eval 𝕜 𝕜 H₁ x)
+
+/--
+The kernel of an RKHS, represented as a scalar value.
+-/
+def scalar_kernel (x y : X) : 𝕜 := (to_dual 𝕜 𝕜).symm (kernel 𝕜 𝕜 H₁ x y)
+
+/--
+The span of all points of the form `eval' x` as a subspace of `H₁`.
+-/
+def eval'_span : submodule 𝕜 H₁ := submodule.span 𝕜 $ set.range (eval' 𝕜 X H₁)
+
+end defs
+
+variables {H₁ : Type*} [inner_product_space 𝕜 H₁] [complete_space H₁] [hH₁rkhs : rkhs 𝕜 X 𝕜 H₁]
+include hH₁rkhs
+
+lemma eval'_def (x : X) : eval' 𝕜 X H₁ x = (to_dual 𝕜 H₁).symm (eval 𝕜 𝕜 H₁ x) := rfl
+
+lemma inner_eval' (f : H₁) (x : X) : inner (eval' 𝕜 X H₁ x) f = f x :=
+by rw [eval'_def, to_dual_symm_apply, eval_apply]
+
+lemma eval'_eq_eval_adjoint (x : X) : eval' 𝕜 X H₁ x = (eval 𝕜 𝕜 H₁ x)† 1 :=
+begin
+  apply (to_dual 𝕜 H₁).injective,
+  ext f,
+  simp [eval'_def],
+end
+
+
+lemma scalar_kernel_def (x y : X) :
+  scalar_kernel 𝕜 X H₁ x y = (to_dual 𝕜 𝕜).symm (kernel 𝕜 𝕜 H₁ x y) := rfl
+
+lemma scalar_kernel_eq_inner (x y : X) :
+  scalar_kernel 𝕜 X H₁ x y = inner (eval' 𝕜 X H₁ y) (eval' 𝕜 X H₁ x) :=
+begin
+  apply (to_dual 𝕜 𝕜).injective,
+  rw [scalar_kernel_def, linear_isometry_equiv.apply_symm_apply, kernel_def],
+  ext,
+  rw [continuous_linear_map.comp_apply, ←eval'_eq_eval_adjoint],
+  simp only [eval_apply, to_dual_apply, is_R_or_C.inner_apply, inner_conj_sym, mul_one],
+  rw inner_eval',
+end
+
+lemma kernel_conj_symm {x y : X} :
+  scalar_kernel 𝕜 X H₁ x y = star_ring_end 𝕜 (scalar_kernel 𝕜 X H₁ y x) :=
+by rw [scalar_kernel_eq_inner, scalar_kernel_eq_inner, inner_conj_sym]
+
+lemma norm_eval'_eq_kernel {x : X} :
+  ∥ eval' 𝕜 X H₁ x ∥^2 = is_R_or_C.re (scalar_kernel 𝕜 X H₁ x x) :=
+by rw [←inner_self_eq_norm_sq, scalar_kernel_eq_inner]
+
+lemma norm_eval_eq_kernel {x : X} :
+  ∥ eval 𝕜 𝕜 H₁ x ∥^2 = is_R_or_C.re (scalar_kernel 𝕜 X H₁ x x) :=
+by rw [←norm_eval'_eq_kernel, eval', linear_isometry_equiv.norm_map]
+
+/--
+The span of the `eval'`s is dense in H
+-/
+lemma eval'_span_topological_closure_eq_top :
+  (eval'_span 𝕜 X H₁).topological_closure = ⊤ :=
+begin
+  rw [submodule.topological_closure_eq_top_iff, submodule.eq_bot_iff],
+  intros f hf,
+  apply fun_like.coe_injective,
+  funext x,
+  specialize hf (eval' 𝕜 X H₁ x) _,
+  rw [eval'_span],
+  apply submodule.subset_span,
+  use x,
+  rw inner_eval' at hf,
+  rw [hf, zero_apply],
+end
+
+/--
+If `f n` tendsto `F` in `H`, then `f n` tendsto `F` pointwise.
+-/
+lemma tendsto_nhds_apply (f : ℕ → H₁) (F : H₁) (hf : tendsto f at_top (𝓝 F))
+  (x : X) : tendsto (λ n, f n x) at_top (𝓝 (F x)) :=
+begin
+  -- two cases, depending on whether `f x = 0` for all `f` in `H`.
+  by_cases hker : eval' 𝕜 X H₁ x = 0,
+  { convert tendsto_const_nhds,
+    funext n,
+    rw [←inner_eval', ←inner_eval', hker, inner_zero_left, inner_zero_left] },
+  { rw [metric.tendsto_nhds] at ⊢ hf,
+    intros ε hε,
+    let ε' := ε / ∥ eval' 𝕜 X H₁ x ∥,
+    have hε' : 0 < ε',
+    { apply div_pos hε,
+      rw norm_pos_iff,
+      exact hker },
+    specialize hf ε' hε',
+    rw filter.eventually_at_top at ⊢ hf,
+    rcases hf with ⟨N, hN⟩,
+    use N,
+    intros n hn,
+    specialize hN n hn,
+    rw [dist_eq_norm, ←inner_eval', ←inner_eval', ←inner_sub_right],
+    refine lt_of_le_of_lt (norm_inner_le_norm _ _) _,
+    rw dist_eq_norm at hN,
+    rw [mul_comm, ←lt_div_iff],
+    { exact hN },
+    { rw norm_pos_iff,
+      exact hker } }
+end
+
+end scalar
+
+end rkhs