Samplinglib
Lean gate not recorded for this source state main · 9bb7a7394ab0

An explicit separation ratio has an exponential floor

Huanjian Zhou and Taiji Suzuki, On Complexity of Gaussian Sampling. Scalar comparison from Proposition B.7; the authored Lean proof uses two canonical logarithm tangent estimates instead of the source integral estimate.

arXiv:2610.20492v1 · Proposition B.7 Eq19 second inequality and Appendix B.2.2 final proof paragraph A2.SS2.SSS2.p6 · faithful paraphrase

Complete source statement and conditions

For every natural dimension \(n\ge3\) and real \(\kappa\ge4\), set \(\rho_\kappa=(\sqrt{\kappa}-1)/(\sqrt{\kappa}+1)\). Then \(\frac12e^{-5n/\sqrt{\kappa}}\le\rho_\kappa^n/(2+4n\sqrt{\kappa}/(\sqrt{\kappa}+1)^2)\). This is the second scalar inequality of (19); the first inequality from actual TV to the ratio is separate.

  • Natural \(n\ge3\) and real \(\kappa\ge4\), including \(\kappa=4\).
  • All square roots, exponentials, powers and divisions are literal real operations. Positivity is proved internally.

Scalar exponential floor

\[\frac12 e^{-5n/\sqrt{\kappa}}\le\frac{\rho_\kappa^n}{2+4n\sqrt{\kappa}/(\sqrt{\kappa}+1)^2}.\]

Read the formalized proofs

Each statement and proof below has its own closed Lean disclosure. Samplinglib parents, Mathlib calls and external mathematical sources are distinguished in each proof.

Samplinglib mathematical lesson

An explicit separation ratio has an exponential floor

Read this first

Mathematical statement

The explicit single-block ratio is at least the exponential scale used in the lower-bound argument.

\[\frac12 e^{-5n/\sqrt{\kappa}}\le\frac{\rho_\kappa^n}{2+4n\sqrt{\kappa}/(\sqrt{\kappa}+1)^2}.\]
Lean statement · explicit_ratio_ge_exponential

Exactly the two source numerical assumptions; no estimate is supplied as a hypothesis.

Braces mark parameters Lean can infer; square brackets request structures such as a measurable space or probability measure. Named hypotheses are mathematical premises, not facts established by this declaration. Section parameters are described in the mathematical hypotheses above; the module link retains their exact source context.

theorem explicit_ratio_ge_exponential {n : ℕ} (hn : 3 ≤ n) {κ : ℝ} (hκ : 4 ≤ κ) :
    (1 / 2 : ℝ) * Real.exp (-(5 * (n : ℝ)) / Real.sqrt κ) ≤
      (((Real.sqrt κ - 1) / (Real.sqrt κ + 1)) ^ n) /
        (2 + 4 * (n : ℝ) * Real.sqrt κ / (Real.sqrt κ + 1) ^ 2)

Exact module and namespace context

Scalar ingredient for Proposition B.7; the actual TV-to-ratio bound remains open.

Full assumptions and notation

All objects and hypotheses

  • Natural \(n\ge3\) and real \(\kappa\ge4\), including \(\kappa=4\).
  • All square roots, exponentials, powers and divisions are literal real operations. Positivity is proved internally.

Notation and interpretation

\(s\) and \(\rho_\kappa\)

Short notation for the square root and spectral ratio.

\[s=\sqrt{\kappa},\qquad\rho_\kappa=(s-1)/(s+1).\]

Mathematical proof

Open a step to read its argument and formula.

1. Check the logarithm domains

Write s for the square root of the condition number. The source assumption makes every denominator and logarithm argument positive.

\[s=\sqrt{\kappa}\ge2,\qquad s-1>0,\quad s+1>0,\quad \rho_\kappa>0.\]
Corresponding Lean step

hs, hs0, hm, hp, ha, hb and hr establish all domains internally via Real.le_sqrt_of_sq_le and positivity.

2. Bound the reciprocal log ratio

The source obtains this final bound through an integral and the sharper intermediate constant 8/3. The Lean route instead splits the ratio and applies the same canonical log(x) ≤ x−1 theorem twice. It proves the displayed constant 3 without claiming the source integral identity.

\[\log\frac{s+1}{s-1}=\log\frac{s+1}{s}+\log\frac{s}{s-1}\le\frac1s+\frac1{s-1}\le\frac3s.\]
Corresponding Lean step

hla and hlb call Real.log_le_sub_one_of_pos; hprod, hfra and hlog combine them. The original integral route and this sufficient alternative remain distinct.

3. Keep the numerator from decaying too quickly

Negate the reciprocal logarithm estimate, exponentiate and take the natural power n.

\[\log\rho_\kappa\ge-3/s,\qquad \rho_\kappa^n\ge e^{-3n/s}.\]
Corresponding Lean step

hlower and hexp use Real.log_div and Real.le_log_iff_exp_le; hpower uses Real.exp_nat_mul and pow_le_pow_left₀.

4. Control the denominator

Use the elementary rational comparison, then the canonical exponential tangent bound.

\[\frac{s}{(s+1)^2}\le\frac1s,\qquad 2+\frac{4ns}{(s+1)^2}\le2+\frac{4n}s\le2e^{2n/s}.\]
Corresponding Lean step

hkernel clears positive denominators; hden uses Real.add_one_le_exp with 2n/s. hdenpos proves that division preserves the comparison.

5. Divide and combine the exponents

The positive numerator and denominator bounds give the scalar floor. Its use with the actual TV ratio bound is the next separate assembly.

\[\frac{\rho_\kappa^n}{2+4ns/(s+1)^2}\ge\frac{e^{-3n/s}}{2e^{2n/s}}=\frac12e^{-5n/s}.\]
Corresponding Lean step

hquot calls div_le_div₀, and heq uses Real.exp_sub. No TV or Gaussian law is a premise of this scalar theorem.

Lean verification

Lean proof · explicit_ratio_ge_exponential

Two shared logarithm tangent bounds, monotonicity, the exponential tangent bound, then positive division.

Braces mark parameters Lean can infer; square brackets request structures such as a measurable space or probability measure. Named hypotheses are mathematical premises, not facts established by this declaration. Section parameters are described in the mathematical hypotheses above; the module link retains their exact source context.

Download module (.lean)
theorem explicit_ratio_ge_exponential {n : ℕ} (hn : 3 ≤ n) {κ : ℝ} (hκ : 4 ≤ κ) :
    (1 / 2 : ℝ) * Real.exp (-(5 * (n : ℝ)) / Real.sqrt κ) ≤
      (((Real.sqrt κ - 1) / (Real.sqrt κ + 1)) ^ n) /
        (2 + 4 * (n : ℝ) * Real.sqrt κ / (Real.sqrt κ + 1) ^ 2) := by
  let s := Real.sqrt κ
  have hs : (2 : ℝ) ≤ s := Real.le_sqrt_of_sq_le (by nlinarith)
  have hs0 : 0 < s := by linarith
  have hm : 0 < s - 1 := by linarith
  have hp : 0 < s + 1 := by linarith
  have ha : 0 < (s + 1) / s := div_pos hp hs0
  have hb : 0 < s / (s - 1) := div_pos hs0 hm
  have hla : Real.log ((s + 1) / s) ≤ 1 / s := by
    convert Real.log_le_sub_one_of_pos ha using 1; field_simp; ring
  have hlb : Real.log (s / (s - 1)) ≤ 1 / (s - 1) := by
    convert Real.log_le_sub_one_of_pos hb using 1; field_simp; ring
  have hprod : ((s + 1) / s) * (s / (s - 1)) = (s + 1) / (s - 1) := by
    field_simp
  have hfra : 1 / (s - 1) ≤ 2 / s := by
    apply (div_le_div_iff₀ hm hs0).2
    linarith
  have hlog : Real.log ((s + 1) / (s - 1)) ≤ 3 / s := by
    rw [← hprod, Real.log_mul (ne_of_gt ha) (ne_of_gt hb)]
    have hsum : 1 / s + 1 / (s - 1) ≤ 3 / s := by
      convert add_le_add_left hfra (1 / s) using 1 <;> ring
    exact (add_le_add hla hlb).trans hsum
  have hr : 0 < (s - 1) / (s + 1) := div_pos hm hp
  have hlower : -(3 : ℝ) / s ≤ Real.log ((s - 1) / (s + 1)) := by
    rw [Real.log_div (ne_of_gt hm) (ne_of_gt hp)]
    rw [Real.log_div (ne_of_gt hp) (ne_of_gt hm)] at hlog
    simpa only [neg_div, neg_sub] using neg_le_neg hlog
  have hexp : Real.exp (-(3 : ℝ) / s) ≤ (s - 1) / (s + 1) :=
    (Real.le_log_iff_exp_le hr).1 hlower
  have hpower : Real.exp (-(3 * (n : ℝ)) / s) ≤ ((s - 1) / (s + 1)) ^ n := by
    have heq : -(3 * (n : ℝ)) / s = (n : ℝ) * (-(3 : ℝ) / s) := by ring
    rw [heq, Real.exp_nat_mul]
    exact pow_le_pow_left₀ (Real.exp_nonneg _) hexp n
  have hkernel : s / (s + 1) ^ 2 ≤ 1 / s := by
    apply (div_le_div_iff₀ (sq_pos_of_pos hp) hs0).2
    nlinarith
  have hden : 2 + 4 * (n : ℝ) * s / (s + 1) ^ 2 ≤
      2 * Real.exp (2 * (n : ℝ) / s) := by
    have hscale := mul_le_mul_of_nonneg_left hkernel
      (show 0 ≤ 4 * (n : ℝ) by positivity)
    have he := Real.add_one_le_exp (2 * (n : ℝ) / s)
    have hscale' : 4 * (n : ℝ) * s / (s + 1) ^ 2 ≤ 4 * (n : ℝ) / s := by
      simpa only [← mul_div_assoc, mul_one] using hscale
    simp only [div_eq_mul_inv] at hscale' he ⊢
    linarith only [hscale', he]
  have hdenpos : 0 < 2 + 4 * (n : ℝ) * s / (s + 1) ^ 2 := by positivity
  have hquot := div_le_div₀ (pow_nonneg (le_of_lt hr) n)
    hpower hdenpos hden
  have heq : Real.exp (-(3 * (n : ℝ)) / s) /
      (2 * Real.exp (2 * (n : ℝ) / s)) =
      (1 / 2 : ℝ) * Real.exp (-(5 * (n : ℝ)) / s) := by
    rw [mul_comm (2 : ℝ), div_mul_eq_div_div, ← Real.exp_sub]
    have harg : -(3 * (n : ℝ)) / s - 2 * (n : ℝ) / s =
        -(5 * (n : ℝ)) / s := by ring
    rw [harg]
    ring
  rw [heq] at hquot
  exact hquot

end AutoSamplingTheory.ExampleCases.GaussianComplexity.NumericSeparation

Exact module and namespace context

Source correspondence and review

Source assumptions versus formal assumptions

SourceActual LeanDifference kindWhy it matters
Natural n>=3 and real kappa>=4 in Proposition B.7.Exactly these assumptions; all rational/logarithm positivity is derived internally.sameThe source scalar second inequality is isolated without assuming the actual TV first inequality or Poisson identities.

Exact scalar second comparison of B.7 Eq19 at natural n>=3 and real kappa>=4, including kappa4. Only literal sqrt/exp/Nat power/real division; positivity and logarithm domains derived internally. Actual single-block TV first ratio bound, selected tilt, sphere/Poisson Eq24, product amplification, exact-oracle transcript/query reductions, fixed-spectrum hidden-eigenbasis, both main theorems and purification remain separate and open.

Encoder–denoiser: accepted · equivalent-after-elaboration

A generalization is not a source correction. Proposed missing conditions require separate independent repair review. No proposed repair silently changes the original theorem.

Scope and omitted-condition boundaries
  • Exact scalar second comparison of B.7 Eq19 at natural n≥3 and real kappa≥4, including kappa4. Only literal sqrt/exp/Nat power/real division; positivity and logarithm domains derived internally. Actual single-block TV first ratio bound, selected tilt, sphere/Poisson Eq24, product amplification, exact-oracle transcript/query reductions, fixed-spectrum hidden-eigenbasis, both main theorems and purification remain separate and open.
Sources, dependencies, and reuse

Source and reuse

Samplinglib parents called

    Mathlib API called (external library)

    • Real.le_sqrt_of_sq_le
    • Real.log_le_sub_one_of_pos
    • Real.log_mul
    • Real.log_div
    • Real.le_log_iff_exp_le
    • Real.exp_nat_mul
    • pow_le_pow_left₀
    • Real.add_one_le_exp
    • div_le_div₀
    • Real.exp_sub

    Mathematical sources

    Samplinglib prose is not a quotation or a source-equivalence certificate. Definitions and aliases are explained as constructions, not counted as new mathematical proofs.

    Which proof obligations and prerequisites are covered?

    • TODO — not closed by these contributions Explicit B7 ratio dominates the exponential floor with coefficient 1/2 and exponent constant5