Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Theorem 9, p. 553 — closed form of the entropy-constrained worst-case VaR

Proved
WorstCaseVaR.Entropy.entropy_constrained_var

by mikedeng1 · Sep 27, 2026 · Mathlib 0df444a (Lean v4.33.1)

p2o-batch-p100bp2o-gran-per-chapterp2o-plan-paperp2o-v1relative-entropyrobust-optimizationvalue-at-risk

Let P0=N(x^,Γ)P_0 = \mathcal N(\hat x, \Gamma)P0​=N(x^,Γ) be a Gaussian distribution on Rn\mathbb R^nRn with positive definite covariance Γ≻0\Gamma \succ 0Γ≻0, let d≥0d \ge 0d≥0, and let Pd\mathcal P_dPd​ be the set of probability distributions PPP of returns with Kullback–Leibler divergence

KL(P,P0)=∫log⁡dPdP0 dP≤d.\mathrm{KL}(P, P_0) = \int \log\frac{dP}{dP_0}\,dP \le d.KL(P,P0​)=∫logdP0​dP​dP≤d.

Let w∈Rnw \in \mathbb R^nw∈Rn be a nonzero portfolio and 0<ε<10 < \varepsilon < 10<ε<1 a loss probability level. The worst-case Value-at-Risk

VPd(w)=min⁡{γ:sup⁡P∈PdP{γ≤−w⊤x}≤ε}V_{\mathcal P_d}(w) = \min\Big\{\gamma : \sup_{P \in \mathcal P_d} P\{\gamma \le -w^\top x\} \le \varepsilon\Big\}VPd​​(w)=min{γ:P∈Pd​sup​P{γ≤−w⊤x}≤ε}

exists and equals

VPd(w)=κ(ε,d)w⊤Γw−x^⊤w,V_{\mathcal P_d}(w) = \kappa(\varepsilon,d)\sqrt{w^\top\Gamma w} - \hat x^\top w,VPd​​(w)=κ(ε,d)w⊤Γw​−x^⊤w,

where κ(ε,d)=−Φ−1(f(ε,d))\kappa(\varepsilon,d) = -\Phi^{-1}(f(\varepsilon,d))κ(ε,d)=−Φ−1(f(ε,d)), f(ε,d)=sup⁡λ>0eε/λ−d−1e1/λ−1f(\varepsilon,d) = \sup_{\lambda>0} \dfrac{e^{\varepsilon/\lambda - d} - 1}{e^{1/\lambda} - 1}f(ε,d)=supλ>0​e1/λ−1eε/λ−d−1​, and Φ\PhiΦ is the standard normal cumulative distribution function.

The theorem shows that an entropy ball around a Gaussian changes the Gaussian VaR only through its risk factor: the worst-case VaR keeps the mean–standard deviation form, with −Φ−1(ε)-\Phi^{-1}(\varepsilon)−Φ−1(ε) replaced by κ(ε,d)\kappa(\varepsilon, d)κ(ε,d), so it can be optimised over portfolios by second-order cone programming.

Formalization Note "min" is encoded as IsLeast of the feasible set {γ:P{γ≤−w⊤x}≤ε for all P∈Pd}\{\gamma : P\{\gamma \le -w^\top x\} \le \varepsilon \text{ for all } P \in \mathcal P_d\}{γ:P{γ≤−w⊤x}≤ε for all P∈Pd​}, and "sup over Pd\mathcal P_dPd​ is at most ε\varepsilonε" as "for every P∈PdP \in \mathcal P_dP∈Pd​". The paper's range ε∈(0,1]\varepsilon \in (0,1]ε∈(0,1] is narrowed to (0,1)(0,1)(0,1): at ε=1\varepsilon = 1ε=1 every γ\gammaγ is feasible and no minimum exists. The hypothesis w≠0w \neq 0w=0 is the paper's standing assumption that the admissible set of portfolios does not contain 000. ∥Γ1/2w∥2\|\Gamma^{1/2}w\|_2∥Γ1/2w∥2​ is written w⊤Γw\sqrt{w^\top \Gamma w}w⊤Γw​, and Φ−1(p)=inf⁡{t:p≤Φ(t)}\Phi^{-1}(p) = \inf\{t : p \le \Phi(t)\}Φ−1(p)=inf{t:p≤Φ(t)}.

Preamble
import Mathlib
import Definitions.Def_WorstCaseVaR_Entropy_Basic
Formal statement
namespace WorstCaseVaR.Entropy

/-- Theorem 9, p. 553: when the distribution of returns is only known to satisfy the relative
entropy constraint `KL(P, P₀) ≤ d` with respect to the Gaussian `P₀ = 𝒩(x̂, Γ)`, `Γ ≻ 0`,
the entropy-constrained worst-case Value-at-Risk (4) of a portfolio `w ≠ 0` at level
`ε ∈ (0, 1)` is `V_𝒫(w) = κ(ε, d)√(wᵀΓw) - x̂ᵀw` (Eq. (44)): this value is the least `γ`
with `P{γ ≤ -wᵀx} ≤ ε` for every such `P`. -/
theorem entropy_constrained_var {n : ℕ} (xhat w : Returns n) (Γ : Matrix (Fin n) (Fin n) ℝ)
    (hΓ : Γ.PosDef) (hw : w ≠ 0) (ε d : ℝ) (hε0 : 0 < ε) (hε1 : ε < 1) (hd : 0 ≤ d) :
    IsLeast (varFeasible (klBall xhat Γ d) w ε)
      (kappaEntropy ε d * Real.sqrt (quadForm Γ w) - inner ℝ xhat w) := by sorry

end WorstCaseVaR.Entropy
Source
El Ghaoui, Oks and Oustry, Worst-Case Value-at-Risk and Robust Portfolio Optimization: A Conic Programming Approach, Oper. Res. 51 (2003), p. 553, Theorem 9, Eqs. (43)-(45); p. 544, Eq. (4)
Read-back

What the Lean code literally says, in plain math · claude-opus-5-5

Throughout, nnn is a natural number and Returns n\mathrm{Returns}\,nReturnsn is a type defined in the imported module Definitions.Def_WorstCaseVaR_Entropy_Basic, which is not shown here. From how it is used, it has a zero element and a real inner product ⟨⋅,⋅⟩\langle\cdot,\cdot\rangle⟨⋅,⋅⟩. Four other names in the statement come from that same unseen module: klBall, varFeasible, kappaEntropy and quadForm. Their definitions are not available to this read-back. They are described below only by their inputs and the type of their output, and are not expanded. The code carries a documentation comment describing what these objects are meant to be, but that comment is not part of what the theorem asserts, and this read-back does not rely on it.

Data. The theorem takes:

  • a natural number nnn (implicit);
  • two elements x^,w∈Returns n\hat x, w \in \mathrm{Returns}\,nx^,w∈Returnsn;
  • a real n×nn \times nn×n matrix Γ\GammaΓ, indexed by {0,…,n−1}\{0, \dots, n-1\}{0,…,n−1};
  • two real numbers ε\varepsilonε and ddd.

Hypotheses.

  • Γ\GammaΓ is positive definite in Mathlib's sense: it is symmetric (Hermitian over R\mathbb{R}R), and v⊤Γv>0v^\top \Gamma v > 0v⊤Γv>0 for every nonzero v∈Rnv \in \mathbb{R}^nv∈Rn.
  • w≠0w \neq 0w=0 in Returns n\mathrm{Returns}\,nReturnsn.
  • 0<ε<10 < \varepsilon < 10<ε<1.
  • d≥0d \ge 0d≥0.

Objects built from the definitions module.

  • B:=klBall(x^,Γ,d)\mathcal{B} := \mathrm{klBall}(\hat x, \Gamma, d)B:=klBall(x^,Γ,d) is built from x^\hat xx^, Γ\GammaΓ and ddd.
  • F:=varFeasible(B,w,ε)F := \mathrm{varFeasible}(\mathcal{B}, w, \varepsilon)F:=varFeasible(B,w,ε) is a set of real numbers built from B\mathcal{B}B, www and ε\varepsilonε.
  • κ(ε,d):=kappaEntropy(ε,d)\kappa(\varepsilon, d) := \mathrm{kappaEntropy}(\varepsilon, d)κ(ε,d):=kappaEntropy(ε,d) is a real number.
  • qΓ(w):=quadForm(Γ,w)q_\Gamma(w) := \mathrm{quadForm}(\Gamma, w)qΓ​(w):=quadForm(Γ,w) is a real number.

Conclusion. Let

γ⋆  =  κ(ε,d) qΓ(w)  −  ⟨x^,w⟩.\gamma^\star \;=\; \kappa(\varepsilon, d)\,\sqrt{q_\Gamma(w)} \;-\; \langle \hat x, w\rangle .γ⋆=κ(ε,d)qΓ​(w)​−⟨x^,w⟩.

Here ⋅\sqrt{\cdot}⋅​ is Mathlib's real square root, which returns 000 for negative arguments. The theorem asserts that γ⋆\gamma^\starγ⋆ is the least element of FFF. That means both of the following hold:

  1. γ⋆∈F\gamma^\star \in Fγ⋆∈F;
  2. γ⋆≤γ\gamma^\star \le \gammaγ⋆≤γ for every γ∈F\gamma \in Fγ∈F.

This is a statement of attainment, not just an infimum: γ⋆\gamma^\starγ⋆ must itself belong to FFF. The statement does not say what membership in FFF means. That depends entirely on the unseen definitions of varFeasible and klBall, and likewise the value of γ⋆\gamma^\starγ⋆ depends on the unseen kappaEntropy and quadForm.

Degenerate cases.

  • n=0n = 0n=0. Here Γ\GammaΓ is the empty matrix, and "positive definite" holds vacuously. If Returns 0\mathrm{Returns}\,0Returns0 has only one element, as a zero-dimensional space would, then w≠0w \neq 0w=0 cannot be satisfied and the theorem is vacuous for n=0n = 0n=0. Whether Returns 0\mathrm{Returns}\,0Returns0 is a one-element type cannot be confirmed from the code shown.
  • d=0d = 0d=0. This is allowed. Then B\mathcal{B}B is whatever klBall returns with radius 000.
  • ε\varepsilonε. It is kept strictly inside (0,1)(0,1)(0,1), so the endpoints 000 and 111 are excluded.
  • Negative qΓ(w)q_\Gamma(w)qΓ​(w). If qΓ(w)q_\Gamma(w)qΓ​(w) were negative, qΓ(w)\sqrt{q_\Gamma(w)}qΓ​(w)​ would silently be 000 and γ⋆\gamma^\starγ⋆ would reduce to −⟨x^,w⟩-\langle \hat x, w\rangle−⟨x^,w⟩. Whether qΓ(w)q_\Gamma(w)qΓ​(w) can be negative under these hypotheses depends on how quadForm is defined.
  • Junk values inside the definitions. Any default values produced inside kappaEntropy, klBall or varFeasible cannot be identified from this code. Examples are division by zero, a logarithm or inverse of a boundary value, or the supremum or infimum of an empty or unbounded set.
  • Empty FFF. If FFF is empty for some admissible inputs, the conclusion γ⋆∈F\gamma^\star \in Fγ⋆∈F is false there. So the theorem, if true, implies that FFF is nonempty and contains γ⋆\gamma^\starγ⋆ for every admissible choice of n,x^,w,Γ,ε,dn, \hat x, w, \Gamma, \varepsilon, dn,x^,w,Γ,ε,d.
Human review
  • Endorsed by Shuze Chen · Sep 27, 2026

    Confirmed by the moderator at approval.

  • Endorsed by mikedeng1 · Sep 27, 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