Catalog.Sparsity.ChernoffBound

Verification

Theorem

Theorem (Chernoff’s bound). Let \(X_1, \ldots, X_n\) be i.i.d. Bernoulli random variables with success probability \(p\). Let \(S = X_1 + \ldots + X_n\) and let \(\mu = np\) be the expected value of \(S\). Then for every \(\delta \in (0, 1)\), we have \[\Pr(|S - \mu| > \delta\mu) \leq 2\exp\!\left(-\frac{\delta^2\mu}{3}\right).\]

Review notes

The source (Theorem 3.3) cites this result without proof as a “classic variant of Chernoff’s bound”. The statement is reproduced faithfully.

Which variant.

This is the multiplicative Chernoff bound for \(\delta \in (0,1)\): two-sided (\(|S-\mu|\)), with constant \(1/3\) in the exponent. Many references state one-sided versions or use the constant \(1/2\); the \(1/3\) constant gives a slightly weaker bound but is standard in the combinatorics literature.

Implicit assumptions.

The source writes \(X_1,\ldots,X_n\) without specifying a probability space. The intended reading is: there is a common probability space on which the \(X_i\) are independent Bernoulli(\(p\)) random variables, and \(\Pr\) denotes the probability measure on that space.

Lean encoding choices.

  • The probability space \((\Omega, \mu)\) is made explicit with IsProbabilityMeasure \(\mu\).
  • The \(n\) random variables are indexed by Fin n and take values in \(\mathbb{R}\). The Bernoulli condition is encoded as \(X_i(\omega) = 0 \lor X_i(\omega) = 1\) a.s., combined with \(\mathbb{E}[X_i] = p\).
  • Independence uses Mathlib’s ProbabilityTheory.iIndepFun.
  • The hypothesis \(p \in [0,1]\) is included explicitly for readability, though it is redundant given the Bernoulli and mean conditions.
  • The event set uses \(\le\) (non-strict) rather than the source’s \(>\); this is slightly stronger and matches Mathlib convention. The source statement follows immediately.
  • The probability is expressed via \(\mu\).real, matching the convention in Mathlib’s sub-Gaussian framework (Mathlib.Probability.Moments.SubGaussian).

Proof scope.

The multiplicative Chernoff bound cannot be derived from Mathlib’s sub-Gaussian/Hoeffding framework: Hoeffding gives \(\exp(-2\delta^2 n p^2)\) (quadratic in \(p\)), while the Chernoff bound gives \(\exp(-\delta^2 n p / 3)\) (linear in \(p\)). For Catalog.Sparsity.Densification, where \(p \to 0\), the linear dependence is essential. The proof requires the specific Bernoulli MGF analysis: \(\mathbb{E}[e^{t(X-p)}] \le \exp((e^t - 1 - t)p)\), optimized over \(t\).

Catalog/Sparsity/ChernoffBound/Contract.lean (full)

1import Mathlib.Probability.Independence.Basic
2import Mathlib.MeasureTheory.Integral.Bochner.Basic
3
4namespace Catalog.Sparsity.ChernoffBound
5
6/-- Theorem 3.3 (Chernoff's bound). For i.i.d. Bernoulli(p) random variables
7X₁, ..., Xₙ with sum S and mean μ = np, for every δ ∈ (0, 1):
8 P(|S - μ| ≥ δμ) ≤ 2 exp(-δ²μ/3). -/
9axiom chernoffBound
10 {Ω : Type*} [MeasurableSpace Ω] {μ : MeasureTheory.Measure Ω}
11 [MeasureTheory.IsProbabilityMeasure μ]
12 {n : ℕ} {X : Fin n → Ω → ℝ} {p : ℝ}
13 (hp : p ∈ Set.Icc 0 1)
14 (hX_meas : ∀ i, Measurable (X i))
15 (hX_indep : ProbabilityTheory.iIndepFun X μ)
16 (hX_bern : ∀ i, ∀ᵐ ω ∂μ, X i ω = 0 ∨ X i ω = 1)
17 (hX_mean : ∀ i, ∫ ω, X i ω ∂μ = p)
18 {δ : ℝ} (hδ : δ ∈ Set.Ioo 0 1) :
19 μ.real {ω | δ * (↑n * p) ≤ |(∑ i : Fin n, X i ω) - ↑n * p|} ≤
20 2 * Real.exp (-(δ ^ 2 * (↑n * p) / 3))
21
22end Catalog.Sparsity.ChernoffBound
23