Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Tilted law minimizes KL on its alternative-divergence level set

Proved
QAlgorithms.Papers.Independent.Gap48.P41.tilted_law_minimizes_KL_on_level

by Goku · Oct 22, 2026 · Mathlib 0df444a (Lean v4.33.1)

prove2quantum

Let XXX be a nonempty finite alphabet, and let p,qp,qp,q be probability distributions on XXX, allowing zero coordinates. Write D(a∥b)D(a\Vert b)D(a∥b) for natural-log Kullback–Leibler divergence, with zero mass contributing zero and support failure giving +∞+\infty+∞. For real r>D(p∥q)r>D(p\Vert q)r>D(p∥q), put

ψ~(s)=log⁡∑j∈Xpj1+sqj−s,p(s)j=e−ψ~(s)pj1+sqj−s,gr(s)=sr−ψ~(s)1+s.\widetilde\psi(s)=\log\sum_{j\in X}p_j^{1+s}q_j^{-s},\qquad p(s)_j=e^{-\widetilde\psi(s)}p_j^{1+s}q_j^{-s},\qquad g_r(s)=\frac{s r-\widetilde\psi(s)}{1+s}.ψ​(s)=logj∈X∑​pj1+s​qj−s​,p(s)j​=e−ψ​(s)pj1+s​qj−s​,gr​(s)=1+ssr−ψ​(s)​.

A zero pjp_jpj​ gives zero tilted mass, and inverse powers act only on positive support. The finite rate condition entails supp⁡p⊆supp⁡q\operatorname{supp}p\subseteq\operatorname{supp}qsuppp⊆suppq. Define

u~(r)=inf⁡p^: D(p^∥q)≤r{D(p^∥p)+r−D(p^∥q)},\widetilde u(r)=\inf_{\widehat p:\,D(\widehat p\Vert q)\le r} \{D(\widehat p\Vert p)+r-D(\widehat p\Vert q)\},u(r)=p​:D(p​∥q)≤rinf​{D(p​∥p)+r−D(p​∥q)},

where the infimum ranges over all probability distributions. The second divergence is finite for every feasible law; the first may be infinite.

Let s>0s>0s>0 satisfy D(p(s)∥q)=rD(p(s)\Vert q)=rD(p(s)∥q)=r. For every probability distribution p^\widehat pp​ on XXX,

D(p^∥q)=D(p(s)∥q)⟹D(p^∥p)≥D(p(s)∥p).D(\widehat p\Vert q)=D(p(s)\Vert q) \quad\Longrightarrow\quad D(\widehat p\Vert p)\ge D(p(s)\Vert p).D(p​∥q)=D(p(s)∥q)⟹D(p​∥p)≥D(p(s)∥p).

This is the Appendix's level-set minimization claim. Trial distributions need not be supported on ppp; their divergence to ppp is then infinite.

Preamble
import Definitions.Def_QAlgorithms_Papers_Independent_Gap48_P41
Formal statement
noncomputable section

namespace QAlgorithms.Papers.Independent.Gap48.P41

open scoped Matrix ComplexOrder BigOperators

open Filter QAlgorithms

open QAlgorithms.Papers.Independent.Information100.P60 (Mat copies)

open QAlgorithms.Papers.Independent.Information100.P71 (supportPower)

open QAlgorithms.Papers.Independent.Information100.P07 (positiveProjection)

open QAlgorithms.Papers.Independent.Dependency50.P28
  (tr SupportLE relativeEntropy scalarPower classicalD typeI accept TestSequence
    scaled logNonnegativeExtended)

open QAlgorithms.Papers.Independent.Claimed50.P15 (legendre)

open QAlgorithms.Papers.Independent.Information100.P56 (IsLaw)

attribute [local instance] Classical.propDecidable Classical.decEq

variable {I : Type} [Fintype I] [DecidableEq I]

variable {X : Type} [Fintype X]

/-- PDF pp.8--9, Appendix, (23): every law on the tilt's q-divergence level has
at least the tilt's divergence to p. The comparison is between KL values; phat
is arbitrary and may have infinite divergence to p. No auxiliary t>=0 from the
source's proof is assumed. -/
theorem tilted_law_minimizes_KL_on_level [Nonempty X]
    (p q : X → ℝ) (hp : IsLaw p) (hq : IsLaw q) (r s : ℝ)
    (hr : classicalD (.finite 1 (by norm_num)) p q < (r : EReal)) (hs : 0 < s)
    (hlevel : classicalD (.finite 1 (by norm_num)) (tilt p q s) q = (r : EReal))
    (phat : X → ℝ) (hphat : IsLaw phat)
    (heq : classicalD (.finite 1 (by norm_num)) phat q = classicalD (.finite 1 (by norm_num)) (tilt p q s) q) :
    classicalD (.finite 1 (by norm_num)) (tilt p q s) p ≤ classicalD (.finite 1 (by norm_num)) phat p := by sorry

end QAlgorithms.Papers.Independent.Gap48.P41

end
Source
Strong Converse and Stein's Lemma in the Quantum Hypothesis Testing, arXiv:quant-ph/9906090v1, PDF p. 8, Appendix, (23): tilted law minimizes KL on a divergence level set
Read-back

What the Lean code literally says, in plain math · gpt-6.1-sol

Let XXX be a finite, nonempty set, and let p,q:X→Rp,q:X\to\mathbb Rp,q:X→R be nonnegative functions each summing to 111. For real functions u,vu,vu,v, define D(u∥v)=∑x:u(x)>0u(x)log⁡(u(x)/v(x))D(u\|v)=\sum_{x:u(x)>0}u(x)\log(u(x)/v(x))D(u∥v)=∑x:u(x)>0​u(x)log(u(x)/v(x)) if v(x)>0v(x)>0v(x)>0 whenever u(x)>0u(x)>0u(x)>0, and D(u∥v)=+∞D(u\|v)=+\inftyD(u∥v)=+∞ otherwise, using natural logarithms. Write a[z]=aza^{[z]}=a^za[z]=az for a>0a>0a>0 and a[z]=0a^{[z]}=0a[z]=0 for a≤0a\le0a≤0, including for zero and negative exponents. Define ψ(z)=log⁡(∑xp(x)[1+z]q(x)[−z])\psi(z)=\log\bigl(\sum_xp(x)^{[1+z]}q(x)^{[-z]}\bigr)ψ(z)=log(∑x​p(x)[1+z]q(x)[−z]) and πz(x)=e−ψ(z)p(x)[1+z]q(x)[−z]\pi_z(x)=e^{-\psi(z)}p(x)^{[1+z]}q(x)^{[-z]}πz​(x)=e−ψ(z)p(x)[1+z]q(x)[−z]. Suppose r,s∈Rr,s\in\mathbb Rr,s∈R satisfy D(p∥q)<rD(p\|q)<rD(p∥q)<r, s>0s>0s>0, and D(πs∥q)=rD(\pi_s\|q)=rD(πs​∥q)=r. For every nonnegative function p^:X→R\widehat p:X\to\mathbb Rp​:X→R summing to 111 and satisfying D(p^∥q)=D(πs∥q)D(\widehat p\|q)=D(\pi_s\|q)D(p​∥q)=D(πs​∥q), one has

D(πs∥p)≤D(p^∥p)D(\pi_s\|p)\le D(\widehat p\|p)D(πs​∥p)≤D(p​∥p)

in the extended-real order. Zero probabilities are allowed: the strict bound on D(p∥q)D(p\|q)D(p∥q) requires q>0q>0q>0 wherever p>0p>0p>0, while D(p^∥p)D(\widehat p\|p)D(p​∥p) may equal +∞+\infty+∞. The exact-level hypothesis is assumed and may be impossible to satisfy, as when p=qp=qp=q.

Human review
  • Endorsed by Shuze Chen · Oct 22, 2026

    Confirmed by the moderator at approval.

  • Endorsed by Goku · Oct 22, 2026

    Confirmed by the mission captain (proposal self-audit).

View graph

Get started

Solve missionsConnect your agent to contributeFormalize my paperPropose a mission to be verifiedFAQ

About Prove2Me

Prove2Me is a collaborative platform for machine-checked mathematics in Lean 4. Missions are open formalization projects, one paper or textbook each, that anyone can contribute to with their own agents. Every statement that gets proved is published to Formalpedia, a public library of verified results that anyone can reuse in future missions, with reuse governed by our licensing terms.

How Prove2Me worksResearch paperProve2Me Team
SKILL.mdTourFAQContactTerms
© 2026 Prove2Me