Problem 2.4: Weak ℓ_q balls and approximation bounds #
From Rigollet Chapter 2, Problem 2.4.
A vector θ ∈ ℝᵈ is in a weak ℓ_q ball of radius R if its decreasing rearrangement satisfies |θ_{[j]}| ≤ R · j^{-1/q}. The weak ℓ_q norm is defined as |θ|{wℓ_q} = max{1 ≤ j ≤ d} j^{1/q} |θ_{[j]}|.
Part (b): |θ|_{wℓ_q} ≤ |θ|_q.
A decreasing rearrangement of (fun i => |θ i|): a function a : Fin d → ℝ
that gives the same multiset of values as (fun i => |θ i|) and is antitone
(nonincreasing). The index convention is 0-based: a ⟨0, _⟩ is the largest.
Each value is nonneg
- antitone : Antitone a
The sequence is nonincreasing
- multiset_eq : Multiset.map a Finset.univ.val = Multiset.map (fun (i : Fin d) => |θ i|) Finset.univ.val
It is a rearrangement of |θ ·| : same multiset of values
Instances For
Membership in a weak ℓ_q ball of radius R: the j-th largest absolute value
of θ is at most R · j^{-1/q}. Equivalently, weakLqNorm q a ≤ R.
Instances For
Helper: a rearrangement preserves sums under any function.
Problem 2.4(b). The weak ℓ_q norm is bounded by the strong ℓ_q norm:
‖θ‖_{wℓ_q} ≤ ‖θ‖_q, i.e.,
max_{j} (j+1)^{1/q} · a(j) ≤ (∑ᵢ |θᵢ|^q)^{1/q}.
The proof uses the fact that since a is the decreasing rearrangement,
j · a(j)^q ≤ ∑ᵢ |θᵢ|^q for each j, so
(j+1)^{1/q} · a(j) ≤ (∑ᵢ |θᵢ|^q)^{1/q}.