Paper
Constructing the Oseledets decomposition with subspace growth estimates
18 pages · 10 marked passages · pdflatex · download PDF · lax-606786
-
Growth, Bernstein numbers and the compactness seminorm of an operator
Let be a bounded linear operator between real normed spaces.
- The growth of on a subspace is the slowest growth of a unit vector of ,with the convention (so ).
- For and , the -th Bernstein number isIn particular , and when .
- The compactness seminorm of is its distance to the compact operators,
1 import Lax606786.Grassmannian 2 import Mathlib.Analysis.Normed.Operator.Compact.Basic 3 … module docstring, 17 lines 21 22 namespace Lax606786.OperatorStatistics 23 24 open Lax606786.Grassmannian 25 26 variable {X Y : Type*} [NormedAddCommGroup X] [NormedSpace ℝ X] 27 [NormedAddCommGroup Y] [NormedSpace ℝ Y] 28 29 /-- `g(T, V) = inf {‖T x‖ : x ∈ V, ‖x‖ = 1}`. -/ 30 noncomputable def growth (T : X →L[ℝ] Y) (V : Submodule ℝ X) : ℝ := 31 sInf ((fun x : X => ‖T x‖) '' (Metric.sphere (0 : X) 1 ∩ (V : Set X))) 32 33 /-- `ρ_k(T) = sup {g(T, V) : V ∈ 𝒢_k X}`. -/ 34 noncomputable def bernsteinNumber (T : X →L[ℝ] X) (k : ℕ) : ℝ := 35 sSup (Set.range fun V : GrassmannianFin X k => growth T (V.1 : Submodule ℝ X)) 36 37 /-- `‖T‖_c = inf {‖T - K‖ : K compact}`. -/ 38 noncomputable def compactSeminorm (T : X →L[ℝ] X) : ℝ := 39 sInf ((fun K : X →L[ℝ] X => ‖T - K‖) '' {K | IsCompactOperator K}) 40 41 end Lax606786.OperatorStatistics 42 -
Cocycles and their Lyapunov exponents
A cocycle consists of an invertible ergodic measure-preserving transformation of a Lebesgue probability space (a standard Borel space with a probability measure giving points measure zero), together with a map from into the bounded operators on a real Banach space which is
- strongly measurable: is measurable for each ;
- forward integrable: .
No invertibility of is assumed. Its iterates are and .
All logarithms below take values in with , and indices of exponents are shifted by one from the usual convention: is .
- The growth rate of a vector is .
- The -th Lyapunov exponent () is , with the Bernstein number. The sequence is non-increasing and is the top exponent.
- The distinct exponents are the distinct values of : and for the least with . The multiplicity of is the number of with . When there is no exponent after the recursion stops: is recorded as , and the later 's are .
- The index of compactness is .
- The cocycle is quasicompact if almost everywhere.
- A family of operators is tempered if is tempered for .
1 import Lax606786.ExtendedLog 2 import Lax606786.OperatorStatistics 3 import Lax606786.TemperedFunctions 4 import Mathlib.Dynamics.Ergodic.Ergodic 5 import Mathlib.MeasureTheory.Function.L1Space.Integrable 6 … module docstring, 36 lines 43 44 namespace Lax606786.Cocycles 45 46 open MeasureTheory Filter 47 open Lax606786.OperatorStatistics Lax606786.ExtendedLog Lax606786.TemperedFunctions 48 49 variable {Ω X : Type*} [MeasurableSpace Ω] [StandardBorelSpace Ω] 50 [NormedAddCommGroup X] [NormedSpace ℝ X] [MeasurableSpace X] 51 52 /-- A strongly measurable, forward-integrable cocycle of bounded operators on `X` over an 53 invertible ergodic transformation of a Lebesgue probability space. -/ 54 structure Cocycle (Ω X : Type*) [MeasurableSpace Ω] [StandardBorelSpace Ω] 55 [NormedAddCommGroup X] [NormedSpace ℝ X] [MeasurableSpace X] where 56 /-- The base transformation, a measurable bijection with measurable inverse. -/ 57 σ : Ω ≃ᵐ Ω 58 /-- The invariant measure. -/ 59 μ : Measure Ω 60 /-- The generator `ω ↦ 𝓛_ω`. -/ 61 L : Ω → X →L[ℝ] X 62 isProbability : IsProbabilityMeasure μ 63 nullSingleton : NullSingletonClass μ 64 ergodic : Ergodic σ μ 65 stronglyMeasurable : ∀ x : X, Measurable (fun ω => L ω x) 66 forwardIntegrable : Integrable (fun ω => Real.log (max ‖L ω‖ 1)) μ 67 68 attribute [instance] Cocycle.isProbability Cocycle.nullSingleton 69 70 namespace Cocycle 71 72 /-- `𝓛^{(n)}_ω = 𝓛_{σ^{n-1}ω} ∘ ⋯ ∘ 𝓛_ω`. -/ 73 noncomputable def iterate (R : Cocycle Ω X) : ℕ → Ω → X →L[ℝ] X 74 | 0, _ => ContinuousLinearMap.id ℝ X 75 | n + 1, ω => (R.L (R.σ^[n] ω)).comp (R.iterate n ω) 76 77 /-- `λ_ω(x) = limsup (1/n) log ‖𝓛^{(n)}_ω x‖`. -/ 78 noncomputable def lambdaAt (R : Cocycle Ω X) (ω : Ω) (x : X) : EReal := 79 limsup (fun n : ℕ => logEReal ‖R.iterate n ω x‖ / (n : EReal)) atTop 80 81 /-- `χ_k(ω) = limsup (1/n) log ρ_k(𝓛^{(n)}_ω)`. Meaningful for `k ≥ 1`; `χ_0 = -∞`. -/ 82 noncomputable def chi (R : Cocycle Ω X) (k : ℕ) (ω : Ω) : EReal := 83 limsup (fun n : ℕ => logEReal (bernsteinNumber (R.iterate n ω) k) / (n : EReal)) atTop 84 85 /-- The index `k` at which the `(i+1)`-st distinct exponent first occurs among the `χ_k`: 86 `1` for `i = 0`, then the least `t ≥ 1` with `χ_t` below the previous distinct exponent, 87 and `0` once there is none. -/ 88 noncomputable def lambdaIdx (R : Cocycle Ω X) : ℕ → Ω → ℕ 89 | 0, _ => 1 90 | (i + 1), ω => sInf {t : ℕ | 1 ≤ t ∧ R.chi t ω < R.chi (R.lambdaIdx i ω) ω} 91 92 /-- `lambda i` is the distinct exponent `λ_{i+1}`. -/ 93 noncomputable def lambda (R : Cocycle Ω X) (i : ℕ) (ω : Ω) : EReal := 94 R.chi (R.lambdaIdx i ω) ω 95 96 /-- `mult i` is the multiplicity `m_{i+1}` of `λ_{i+1}`; `0` if `λ_{i+1}` is the last distinct 97 exponent. -/ 98 noncomputable def mult (R : Cocycle Ω X) (i : ℕ) (ω : Ω) : ℕ := 99 R.lambdaIdx (i + 1) ω - R.lambdaIdx i ω 100 101 /-- `ν(ω) = inf_{k ≥ 1} χ_k(ω)`. -/ 102 noncomputable def nu (R : Cocycle Ω X) (ω : Ω) : EReal := ⨅ k : ℕ, R.chi (k + 1) ω 103 104 /-- `ν < λ₁` almost everywhere. -/ 105 def IsQuasicompact (R : Cocycle Ω X) : Prop := 106 ∀ᵐ ω ∂(R.μ), R.nu ω < R.chi 1 ω 107 108 /-- `ω ↦ log ‖P_ω‖` is tempered. -/ 109 abbrev IsTempered (R : Cocycle Ω X) (P : Ω → X →L[ℝ] X) : Prop := 110 Tempered R.σ R.μ (fun ω => Real.log ‖P ω‖) 111 112 end Cocycle 113 114 end Lax606786.Cocycles 115 -
The semi-invertible Oseledets decomposition
Let be a cocycle on a separable Banach space , over an invertible ergodic base . Let be its distinct Lyapunov exponents, their multiplicities, and the number of distinct exponents. Then and are almost everywhere constant, if and only if is quasicompact, and there are
- measurable families of subspaces of dimension (the fast spaces), for ,
- strongly measurable, tempered families of projections , for , with ranges (the slow spaces),
such that for almost every :
- , and grows at rate exactly , both in norm and in slowest growth:
- , and is the projection onto along ;
- ;
- .
This is the semi-invertible form of Oseledets' multiplicative ergodic theorem (Oseledets 1968; Froyland, Lloyd and Quas 2013; González-Tokman and Quas 2014), for strongly measurable cocycles on separable Banach spaces as in Lee (2024). In other words, has an Oseledets decomposition in the sense of the definition of Oseledets decompositions, whose zero-based indexing the Lean statement follows. Measurability of is with respect to the Borel structure of the Grassmannian; strong measurability of means that is measurable for each .
1 import Lax606786.OseledetsDecompositions 2 … module docstring, 35 lines 38 39 namespace Lax606786.OseledetsDecomposition 40 41 open TopologicalSpace 42 open Lax606786.Grassmannian Lax606786.Cocycles Lax606786.OseledetsDecompositions 43 44 /-- Every cocycle on a separable Banach space has an Oseledets decomposition. (On the zero 45 space it is the empty one: `L = 1`, `λ₁ = -∞`, no fast spaces.) -/ 46 axiom oseledets_decomposition {Ω X : Type*} [MeasurableSpace Ω] [StandardBorelSpace Ω] 47 [NormedAddCommGroup X] [NormedSpace ℝ X] [MeasurableSpace X] 48 (R : Cocycle Ω X) [SeparableSpace X] [BorelSpace X] [CompleteSpace X] : 49 ∃ (Lval : ℕ∞) (lam : ℕ → EReal) (mdim : ℕ → ℕ) 50 (E : ∀ i : ℕ, ((i + 2 : ℕ) : ℕ∞) ≤ Lval → Ω → GrassmannianFin X (mdim i + 1)) 51 (V : ℕ → Ω → Submodule ℝ X) (P : ℕ → Ω → X →L[ℝ] X), 52 IsOseledetsDecomposition R Lval lam mdim E V P 53 54 end Lax606786.OseledetsDecomposition 55 -
The Grassmannians of a Banach space
Let be a real normed space. The unit sphere of a subspace is ; it spans , so it determines .
- The Grassmannian is the set of closed subspaces admitting a bounded projection . It carries the distancethe Hausdorff distance between unit spheres (infinite exactly when one of is and the other is not).
- For , is the set of -dimensional subspaces of . For the unit sphere is a nonempty compact set and carries the metric .
Each is equipped with the Borel -algebra of its distance.
1 import Mathlib.Analysis.RCLike.Lemmas 2 import Mathlib.MeasureTheory.Constructions.BorelSpace.Basic 3 import Mathlib.Topology.MetricSpace.Closeds 4 … module docstring, 19 lines 24 25 namespace Lax606786.Grassmannian 26 27 open TopologicalSpace 28 29 variable (X : Type*) [NormedAddCommGroup X] [NormedSpace ℝ X] 30 31 /-- `𝒢X`: the closed subspaces of `X` admitting a bounded projection onto them. -/ 32 def Grassmannian : Type _ := 33 {V : Submodule ℝ X // V.ClosedComplemented} 34 35 variable {X} 36 37 /-- Every subspace is spanned by its unit sphere. -/ 38 theorem span_sphere_inter (V : Submodule ℝ X) : 39 Submodule.span ℝ (Metric.sphere (0 : X) 1 ∩ (V : Set X)) = V := by 40 apply le_antisymm 41 · exact Submodule.span_le.2 Set.inter_subset_right 42 · intro x hx 43 rcases eq_or_ne x 0 with rfl | hx0 44 · exact Submodule.zero_mem _ 45 · have hxeq : x = ‖x‖ • (‖x‖⁻¹ • x) := by 46 rw [smul_smul, mul_inv_cancel₀ (norm_ne_zero_iff.2 hx0), one_smul] 47 rw [hxeq] 48 refine Submodule.smul_mem _ _ (Submodule.subset_span ⟨?_, V.smul_mem _ hx⟩) 49 rw [mem_sphere_zero_iff_norm, norm_smul, norm_inv, norm_norm, 50 inv_mul_cancel₀ (norm_ne_zero_iff.2 hx0)] 51 52 namespace Grassmannian 53 54 /-- The unit sphere `𝕊_V = {x ∈ V : ‖x‖ = 1}`. -/ 55 def sphere (V : Grassmannian X) : Set X := 56 Metric.sphere (0 : X) 1 ∩ (V.1 : Set X) 57 58 /-- `V ↦ 𝕊_V`, as a closed subset of `X`. -/ 59 def toCloseds (V : Grassmannian X) : Closeds X := 60 ⟨V.sphere, Metric.isClosed_sphere.inter V.2.isClosed⟩ 61 62 theorem toCloseds_injective : Function.Injective (toCloseds (X := X)) := by 63 intro V W h 64 apply Subtype.ext 65 rw [← span_sphere_inter V.1, ← span_sphere_inter W.1] 66 exact congrArg (fun s : Closeds X => Submodule.span ℝ (s : Set X)) h 67 68 /-- The distance on `𝒢X`: the Hausdorff distance between unit spheres. -/ 69 noncomputable instance instEMetricSpace : EMetricSpace (Grassmannian X) := 70 EMetricSpace.induced toCloseds toCloseds_injective inferInstance 71 72 /-- The Borel `σ`-algebra of that distance. -/ 73 noncomputable instance instMeasurableSpace : MeasurableSpace (Grassmannian X) := 74 borel (Grassmannian X) 75 76 instance instBorelSpace : BorelSpace (Grassmannian X) := ⟨rfl⟩ 77 78 end Grassmannian 79 80 variable (X) 81 82 /-- `𝒢_k X`: the `k`-dimensional subspaces of `X`. -/ 83 def GrassmannianFin (k : ℕ) : Type _ := 84 {V : Submodule ℝ X // FiniteDimensional ℝ V ∧ Module.finrank ℝ V = k} 85 86 variable {X} 87 88 namespace GrassmannianFin 89 90 /-- The unit sphere `𝕊_V = {x ∈ V : ‖x‖ = 1}`. -/ 91 def sphere {k : ℕ} (V : GrassmannianFin X k) : Set X := 92 Metric.sphere (0 : X) 1 ∩ (V.1 : Set X) 93 94 /-- `V` is the span of its unit sphere. -/ 95 theorem span_sphere {k : ℕ} (V : GrassmannianFin X k) : 96 Submodule.span ℝ V.sphere = (V.1 : Submodule ℝ X) := by 97 apply le_antisymm 98 · exact Submodule.span_le.2 Set.inter_subset_right 99 · intro x hx 100 rcases eq_or_ne x 0 with rfl | hx0 101 · exact Submodule.zero_mem _ 102 · have hxeq : x = ‖x‖ • (‖x‖⁻¹ • x) := by 103 rw [smul_smul, mul_inv_cancel₀ (norm_ne_zero_iff.2 hx0), one_smul] 104 rw [hxeq] 105 refine Submodule.smul_mem _ _ (Submodule.subset_span ⟨?_, V.1.smul_mem _ hx⟩) 106 rw [mem_sphere_zero_iff_norm, norm_smul, norm_inv, norm_norm, 107 inv_mul_cancel₀ (norm_ne_zero_iff.2 hx0)] 108 109 /-- `𝕊_V` is compact. -/ 110 theorem isCompact_sphere {k : ℕ} (V : GrassmannianFin X k) : IsCompact V.sphere := by 111 have : FiniteDimensional ℝ (V.1 : Submodule ℝ X) := V.2.1 112 have himg : V.sphere = 113 (Subtype.val : (V.1 : Submodule ℝ X) → X) '' Metric.sphere (0 : (V.1 : Submodule ℝ X)) 1 := by 114 ext x 115 simp only [sphere, Set.mem_inter_iff, Set.mem_image] 116 constructor 117 · rintro ⟨hx1, hxV⟩ 118 refine ⟨⟨x, hxV⟩, ?_, rfl⟩ 119 rwa [mem_sphere_zero_iff_norm] at hx1 ⊢ 120 · rintro ⟨v, hv, rfl⟩ 121 rw [mem_sphere_zero_iff_norm] at hv 122 exact ⟨by rwa [mem_sphere_zero_iff_norm], v.2⟩ 123 rw [himg] 124 exact (_root_.isCompact_sphere (0 : (V.1 : Submodule ℝ X)) 1).image continuous_subtype_val 125 126 /-- `𝕊_V` is nonempty when `k ≠ 0`. -/ 127 theorem nonempty_sphere {k : ℕ} [NeZero k] (V : GrassmannianFin X k) : V.sphere.Nonempty := by 128 have hne : (V.1 : Submodule ℝ X) ≠ ⊥ := by 129 intro h 130 apply NeZero.ne k 131 have hfr := V.2.2 132 rw [h, finrank_bot] at hfr 133 exact hfr.symm 134 obtain ⟨x, hxV, hx0⟩ := Submodule.exists_mem_ne_zero_of_ne_bot hne 135 refine ⟨‖x‖⁻¹ • x, ?_, V.1.smul_mem _ hxV⟩ 136 rw [mem_sphere_zero_iff_norm, norm_smul, norm_inv, norm_norm, 137 inv_mul_cancel₀ (norm_ne_zero_iff.2 hx0)] 138 139 /-- `V ↦ 𝕊_V`, as a nonempty compact subset of `X`. -/ 140 noncomputable def toNonemptyCompacts {k : ℕ} [NeZero k] (V : GrassmannianFin X k) : 141 NonemptyCompacts X where 142 carrier := V.sphere 143 isCompact' := V.isCompact_sphere 144 nonempty' := V.nonempty_sphere 145 146 theorem toNonemptyCompacts_injective (k : ℕ) [NeZero k] : 147 Function.Injective (toNonemptyCompacts (X := X) (k := k)) := by 148 intro V W h 149 apply Subtype.ext 150 rw [← V.span_sphere, ← W.span_sphere] 151 exact congrArg (fun s : NonemptyCompacts X => Submodule.span ℝ (s : Set X)) h 152 153 /-- The metric on `𝒢_k X`: the Hausdorff distance between unit spheres. -/ 154 noncomputable instance instMetricSpace (k : ℕ) [NeZero k] : MetricSpace (GrassmannianFin X k) := 155 MetricSpace.induced toNonemptyCompacts (toNonemptyCompacts_injective k) inferInstance 156 157 /-- The Borel `σ`-algebra of that metric. -/ 158 noncomputable instance instMeasurableSpace (k : ℕ) [NeZero k] : 159 MeasurableSpace (GrassmannianFin X k) := 160 borel (GrassmannianFin X k) 161 162 instance instBorelSpace (k : ℕ) [NeZero k] : BorelSpace (GrassmannianFin X k) := ⟨rfl⟩ 163 164 end GrassmannianFin 165 166 end Lax606786.Grassmannian 167 -
Subadditive families of measurable functions
Let be a map on a measurable space. A sequence of measurable functions is a subadditive family over if
The typical example is for a cocycle of bounded operators.
1 import Mathlib.MeasureTheory.Constructions.BorelSpace.Real 2 … module docstring, 13 lines 16 17 namespace Lax606786.SubadditiveFamilies 18 19 /-- `(f_n)` is a subadditive family of measurable functions over `σ`. -/ 20 def IsSubadditiveFamily {Ω : Type*} [MeasurableSpace Ω] (σ : Ω → Ω) (f : ℕ → Ω → ℝ) : Prop := 21 (∀ n, Measurable (f n)) ∧ ∀ m n ω, f (m + n) ω ≤ f m (σ^[n] ω) + f n ω 22 23 end Lax606786.SubadditiveFamilies 24 -
Kingman's subadditive ergodic theorem
Let be an ergodic measure-preserving transformation of a probability space , and let be a subadditive family of integrable functions over . Then there is a constant such that
Limits are taken in the extended reals . The theorem is due to Kingman (1968).
1 import Lax606786.SubadditiveFamilies 2 import Mathlib.Dynamics.Ergodic.Ergodic 3 import Mathlib.MeasureTheory.Integral.Bochner.Basic 4 import Mathlib.Data.EReal.Operations 5 … module docstring, 13 lines 19 20 namespace Lax606786.KingmanTheorem 21 22 open MeasureTheory Filter Topology 23 open Lax606786.SubadditiveFamilies 24 25 /-- Kingman's theorem: `f_n / n` converges almost everywhere to the constant 26 `C = lim (1/n) ∫ f_n ∈ [-∞, ∞)`. -/ 27 axiom kingman {Ω : Type*} [MeasurableSpace Ω] (μ : Measure Ω) [IsProbabilityMeasure μ] 28 (σ : Ω → Ω) (hσ : Ergodic σ μ) (f : ℕ → Ω → ℝ) (hf : IsSubadditiveFamily σ f) 29 (hint : ∀ n, Integrable (f n) μ) : 30 ∃ C : EReal, C ≠ ⊤ ∧ 31 Tendsto (fun n : ℕ => (((∫ ω, f n ω ∂μ) / n : ℝ) : EReal)) atTop (𝓝 C) ∧ 32 ∀ᵐ ω ∂μ, Tendsto (fun n : ℕ => (f n ω / n : EReal)) atTop (𝓝 C) 33 34 end Lax606786.KingmanTheorem 35 -
Kingman's theorem for balanced intervals
Let be an invertible ergodic measure-preserving transformation of a Lebesgue probability space (a standard Borel space with a probability measure giving points measure zero), and let be a subadditive family of integrable functions over . Then the averages over the balanced time intervals converge to the same constant as in Kingman's theorem:
with and limits taken in the extended reals. This is the balanced subadditive ergodic theorem of Lee (2024).
1 import Lax606786.SubadditiveFamilies 2 import Mathlib.Dynamics.Ergodic.Ergodic 3 import Mathlib.MeasureTheory.Integral.Bochner.Basic 4 import Mathlib.MeasureTheory.Measure.Typeclasses.NoAtoms 5 import Mathlib.MeasureTheory.Constructions.Polish.Basic 6 import Mathlib.Data.EReal.Operations 7 … module docstring, 15 lines 23 24 namespace Lax606786.BalancedKingman 25 26 open MeasureTheory Filter Topology 27 open Lax606786.SubadditiveFamilies 28 29 /-- `f_{2n}(σ^{-n} ω) / 2n` converges almost everywhere to `C = lim (1/n) ∫ f_n`. -/ 30 axiom balancedKingman {Ω : Type*} [MeasurableSpace Ω] [StandardBorelSpace Ω] 31 (μ : Measure Ω) [IsProbabilityMeasure μ] [NullSingletonClass μ] 32 (σ : Ω ≃ᵐ Ω) (hσ : Ergodic σ μ) (f : ℕ → Ω → ℝ) (hf : IsSubadditiveFamily σ f) 33 (hint : ∀ n, Integrable (f n) μ) : 34 ∃ C : EReal, C ≠ ⊤ ∧ 35 Tendsto (fun n : ℕ => (((∫ ω, f n ω ∂μ) / n : ℝ) : EReal)) atTop (𝓝 C) ∧ 36 ∀ᵐ ω ∂μ, Tendsto (fun n : ℕ => (f (2 * n) ((σ.symm : Ω → Ω)^[n] ω) / (2 * n) : EReal)) 37 atTop (𝓝 C) 38 39 end Lax606786.BalancedKingman 40 -
no assumptions
The construction of Lee (2024): the decomposition is built level by level, each fast space being the top fast space of the cocycle restricted to the previous slow space, with the slow projections constructed alongside.
-
The index of compactness is the growth rate of the compactness seminorm
Let be a cocycle on a separable Banach space . If the growth rate of the compactness seminorm of the iterates exists and is almost everywhere equal to a constant ,
then the index of compactness equals almost everywhere. This is the equivalence of growth statistics in the appendix of Lee (2024).
1 import Lax606786.Cocycles 2 … module docstring, 12 lines 15 16 namespace Lax606786.IndexOfCompactness 17 18 open MeasureTheory Filter TopologicalSpace 19 open Lax606786.OperatorStatistics Lax606786.ExtendedLog Lax606786.Cocycles 20 21 /-- `ν = κ` almost everywhere, where `κ` is the growth rate of `‖𝓛^{(n)}_ω‖_c`. -/ 22 axiom nu_eq_compactnessIndex {Ω X : Type*} [MeasurableSpace Ω] [StandardBorelSpace Ω] 23 [NormedAddCommGroup X] [NormedSpace ℝ X] [MeasurableSpace X] 24 (R : Cocycle Ω X) [SeparableSpace X] [BorelSpace X] 25 (κ : EReal) (hκ : ∀ᵐ ω ∂(R.μ), Tendsto (fun n : ℕ => 26 logEReal (compactSeminorm (R.iterate n ω)) / (n : EReal)) atTop (nhds κ)) : 27 ∀ᵐ ω ∂(R.μ), R.nu ω = κ 28 29 end Lax606786.IndexOfCompactness 30 -
Slow spaces on a separable Banach space need not be measurable
Let with its product -algebra, and let . For let
Then is separable and every belongs to the Grassmannian , but there is no measurable map with for every .
Over the shift with the Bernoulli measure, the operators form a cocycle of rank one with , fast space and slow space ; so separability of alone does not make the slow spaces measurable. This example is due to J. A. Horan (PhD thesis, University of Victoria, 2020). The argument: distinct are at distance at least in and , so a measurable would make every set measurable, while there are more such sets than measurable ones.
In Lean a sign is coded by a boolean, for .
1 import Lax606786.Grassmannian 2 import Mathlib.Analysis.Normed.Lp.lpSpace 3 import Mathlib.MeasureTheory.MeasurableSpace.Pi 4 … module docstring, 24 lines 29 30 namespace Lax606786.SlowSpaceNonmeasurability 31 32 open TopologicalSpace 33 open Lax606786.Grassmannian 34 35 /-- `ω_i ∈ {±1}`, coded by `true ↦ 1` and `false ↦ -1`. -/ 36 def sign (ω : ℤ → Bool) (i : ℤ) : ℝ := if ω i then 1 else -1 37 38 /-- `V_ω = {x ∈ ℓ¹(ℤ) : Σᵢ ωᵢ xᵢ = 0}`. -/ 39 def slowSpace (ω : ℤ → Bool) : Set (lp (fun _ : ℤ => ℝ) 1) := 40 {x | ∑' i, sign ω i * x i = 0} 41 42 /-- `ℓ¹(ℤ)` is separable and each `V_ω` lies in `𝒢X`, but no measurable map `Ω → 𝒢X` takes the 43 value `V_ω` at every `ω`. -/ 44 axiom slowSpace_not_measurable : 45 SeparableSpace (lp (fun _ : ℤ => ℝ) 1) ∧ 46 (∃ W : (ℤ → Bool) → Grassmannian (lp (fun _ : ℤ => ℝ) 1), 47 ∀ ω, ((W ω).1 : Set (lp (fun _ : ℤ => ℝ) 1)) = slowSpace ω) ∧ 48 ¬ ∃ W : (ℤ → Bool) → Grassmannian (lp (fun _ : ℤ => ℝ) 1), 49 Measurable W ∧ ∀ ω, ((W ω).1 : Set (lp (fun _ : ℤ => ℝ) 1)) = slowSpace ω 50 51 end Lax606786.SlowSpaceNonmeasurability 52