Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Theorem 4.1, p. 11 — Ridge fit-and-transfer error on a new question: structural plus estimation error

Proved
SynDigits.Ridge.theorem_4_1

by mikedeng1 · Oct 24, 2026 · Mathlib 0df444a (Lean v4.33.1)

digital-twinerror-boundp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-strengthenedp2o-v1ridge-regressionsynthetic-control

Setting (latent factor model, §4.1). There are nnn users, mmm existing questions and a latent dimension ddd. User embeddings U∈Rn×dU \in \mathbb{R}^{n\times d}U∈Rn×d, question embeddings V∈Rm×dV \in \mathbb{R}^{m\times d}V∈Rm×d (rows vjv_jvj​) and a noise matrix E∈Rn×m\mathcal E \in \mathbb{R}^{n\times m}E∈Rn×m give the human responses Y=UV⊤+EY = UV^\top + \mathcal EY=UV⊤+E. The new question has embedding v∈Rdv \in \mathbb{R}^dv∈Rd and noise Ev∈Rn\mathcal E_v \in \mathbb{R}^nEv​∈Rn, and the (unobserved) human responses to it are Yv=Uv+EvY_v = Uv + \mathcal E_vYv​=Uv+Ev​. Decompose v=V⊤b+ev = V^\top b + ev=V⊤b+e with b∈Rmb \in \mathbb{R}^mb∈Rm and e∈Rde \in \mathbb{R}^de∈Rd.

Estimator (Ridge fit-and-transfer, §3.2). The digital-twin responses are Y~∈Rn×m\tilde Y \in \mathbb{R}^{n\times m}Y~∈Rn×m on the existing questions and Y~v∈Rn\tilde Y_v \in \mathbb{R}^nY~v​∈Rn on the new one. With Σ~^=1nY~⊤Y~\hat{\tilde\Sigma} = \frac1n\tilde Y^\top\tilde YΣ~^=n1​Y~⊤Y~, γ~^=1nY~⊤Y~v\hat{\tilde\gamma} = \frac1n\tilde Y^\top\tilde Y_vγ~​^​=n1​Y~⊤Y~v​ and penalty λ≥0\lambda \ge 0λ≥0, the Ridge fit on the twins is β^=(Σ~^+λI)−1γ~^\hat\beta = (\hat{\tilde\Sigma} + \lambda I)^{-1}\hat{\tilde\gamma}β^​=(Σ~^+λI)−1γ~​^​, and it is transferred to the humans as Y^v=Yβ^\hat Y_v = Y\hat\betaY^v​=Yβ^​.

Population target. Let Σ~∈Rm×m\tilde\Sigma \in \mathbb{R}^{m\times m}Σ~∈Rm×m be symmetric positive semidefinite and γ~∈Rm\tilde\gamma \in \mathbb{R}^mγ~​∈Rm (in the paper, Σ~=1nE[Y~⊤Y~]\tilde\Sigma = \frac1n\mathbb E[\tilde Y^\top\tilde Y]Σ~=n1​E[Y~⊤Y~] and γ~=1nE[Y~⊤Y~v]\tilde\gamma = \frac1n\mathbb E[\tilde Y^\top\tilde Y_v]γ~​=n1​E[Y~⊤Y~v​]). Put β∗=(Σ~+λI)−1γ~\beta^* = (\tilde\Sigma + \lambda I)^{-1}\tilde\gammaβ∗=(Σ~+λI)−1γ~​ and r=Yv−Yβ∗r = Y_v - Y\beta^*r=Yv​−Yβ∗. Assume σmin⁡(Σ~^)+λ>0\sigma_{\min}(\hat{\tilde\Sigma}) + \lambda > 0σmin​(Σ~^)+λ>0 and σmin⁡(Σ~)+λ>0\sigma_{\min}(\tilde\Sigma) + \lambda > 0σmin​(Σ~)+λ>0, where σmin⁡\sigma_{\min}σmin​ is the smallest singular value.

Theorem. With ∥⋅∥2\|\cdot\|_2∥⋅∥2​ the Euclidean norm on vectors and the spectral norm on matrices,

∥Y^v−Yv∥2≤∥r∥2⏟structural error+∥Y∥2σmin⁡(Σ~^)+λ(∥Σ~^−Σ~∥2 ∥β∗∥2+∥γ~^−γ~∥2)⏟estimation error,\|\hat Y_v - Y_v\|_2 \le \underbrace{\|r\|_2}_{\text{structural error}} + \underbrace{\frac{\|Y\|_2}{\sigma_{\min}(\hat{\tilde\Sigma}) + \lambda}\Big(\|\hat{\tilde\Sigma} - \tilde\Sigma\|_2\,\|\beta^*\|_2 + \|\hat{\tilde\gamma} - \tilde\gamma\|_2\Big)}_{\text{estimation error}},∥Y^v​−Yv​∥2​≤structural error∥r∥2​​​+estimation errorσmin​(Σ~^)+λ∥Y∥2​​(∥Σ~^−Σ~∥2​∥β∗∥2​+∥γ~​^​−γ~​∥2​)​​,

with

∥r∥2≤∥Ue∥2+∥Ev−Eb∥2+∥Y∥2σmin⁡(Σ~)+λ(λ∥b∥2+∥γ~−Σ~b∥2).\|r\|_2 \le \|Ue\|_2 + \|\mathcal E_v - \mathcal E b\|_2 + \frac{\|Y\|_2}{\sigma_{\min}(\tilde\Sigma) + \lambda}\Big(\lambda\|b\|_2 + \|\tilde\gamma - \tilde\Sigma b\|_2\Big).∥r∥2​≤∥Ue∥2​+∥Ev​−Eb∥2​+σmin​(Σ~)+λ∥Y∥2​​(λ∥b∥2​+∥γ~​−Σ~b∥2​).

The theorem explains why calibrating on digital-twin responses transfers to humans: the estimation error vanishes as the sample moments concentrate, and the structural error is small when the new question is nearly representable by existing ones and the twins' inter-question covariance is aligned with the humans'. Exact alignment of the two systems is not needed.

Formalization Note Both inequalities are conclusions (a conjunction). The two positivity hypotheses make Σ~^+λI\hat{\tilde\Sigma} + \lambda IΣ~^+λI and Σ~+λI\tilde\Sigma + \lambda IΣ~+λI invertible, so β^\hat\betaβ^​ and β∗\beta^*β∗ are not Lean's junk value 000 and the denominators are positive. At n=0n = 0n=0 the factor 1/n1/n1/n is 000 in Lean, but then Y^v−Yv\hat Y_v - Y_vY^v​−Yv​ and rrr live in R0\mathbb R^0R0, both left-hand sides are 000, and the statement holds trivially.

Differences from the source:

  • generalization: Σ̃ = (1/n)𝔼[ỸᵀỸ], γ̃ = (1/n)𝔼[ỸᵀỸ_v] for a random twin matrix → any positive semidefinite Sig, any vector gam; reason: every such expectation is symmetric PSD, so the statement specializes to the paper's; the argument uses only (Σ̃+λI)β* = γ̃ and Σ̃ ⪰ 0
  • encoding: v = Vb + e → v = Vᵀ *ᵥ b + e; reason: V ∈ ℝ^{m×d} has the question embeddings as rows and b ∈ ℝ^m, so only Vᵀb type-checks; the proof's Yb = UVᵀb + ℰb (p. 32) confirms it
  • reading: Y_v = Uv (§4.1, p. 9, noiseless); ℰ_v appears in Theorem 4.1 without a definition → hYv : Yv = U *ᵥ v + Ev; reason: Theorem 4.1 is stated "in the noisy setting" (p. 11) and its proof writes Y_v − Yb = Uv + ℰ_v − UVb − ℰb (p. 32), so Y_v = Uv + ℰ_v; Ev = 0 recovers the noiseless model of p. 9
  • reading: Ỹ ∈ ℝ^{n×(m+1)}; Σ̂̃, γ̂̃, β̂, Ŷ_v used but not defined → Yt = Ỹ_{1:m} : Matrix (Fin n) (Fin m) ℝ, ytv = Ỹ_v : Fin n → ℝ, sampleCov Yt = (1/n)ỸᵀỸ, sampleCross Yt ytv = (1/n)ỸᵀỸ_v, ridgeFit = (Σ̂̃+λI)⁻¹γ̂̃, ridgePredict = Y β̂; reason: Algorithm 1 fits Ỹ_{m+1} on Ỹ_{1:m}, and the m×m Σ̃ of p. 11 is the moment of Ỹ_{1:m}; the hat quantities are the sample analogues of the printed Σ̃, γ̃; the proof's normal equation (Σ̂̃+λI)β̂ = γ̂̃ (p. 32); Algorithm 1's Predict Ŷ_{m+1} = M(Y_{1:m}) (p. 6)
  • encoding: (Σ̂̃+λI)⁻¹ and (Σ̃+λI)⁻¹ exist (implicit) → 0 < sMin (sampleCov Yt) + lam and 0 < sMin Sig + lam; reason: for PSD S these exclude exactly the singular S + λI, where the paper's β̂, β* and the bound's denominators are undefined; Lean's inverse returns 0 there
  • encoding: λ ≥ 0 (Ridge penalty, implicit) → 0 ≤ lam; reason: Ridge penalties are nonnegative; the printed λ‖b‖₂ equals ‖λb‖₂ only for λ ≥ 0
  • encoding: vectors in ℝⁿ, n×m matrices, ‖·‖₂, spectral norm, σ_min → Fin n → ℝ, Matrix (Fin n) (Fin m) ℝ, vnorm, GaussianMatrix.specNorm, GaussianMatrix.sMin; reason: Lean representation of the same objects; same instances
Preamble
import Mathlib
import Definitions.Def_GaussianMatrix_basic
import Definitions.Def_SynDigits_Ridge_Model
open Matrix GaussianMatrix
Formal statement
namespace SynDigits.Ridge

theorem theorem_4_1 {n m d : ℕ} (U : Matrix (Fin n) (Fin d) ℝ)
    (V : Matrix (Fin m) (Fin d) ℝ) (Eps Y : Matrix (Fin n) (Fin m) ℝ) (v e : Fin d → ℝ)
    (Ev Yv : Fin n → ℝ) (b : Fin m → ℝ) (Yt : Matrix (Fin n) (Fin m) ℝ) (ytv : Fin n → ℝ)
    (Sig : Matrix (Fin m) (Fin m) ℝ) (gam : Fin m → ℝ) (lam : ℝ)
    (hY : Y = U * Vᵀ + Eps) (hYv : Yv = U *ᵥ v + Ev) (hv : v = Vᵀ *ᵥ b + e)
    (hSig : Sig.PosSemidef) (hlam : 0 ≤ lam)
    (hposHat : 0 < sMin (sampleCov Yt) + lam) (hpos : 0 < sMin Sig + lam) :
    vnorm (ridgePredict Y Yt ytv lam - Yv)
        ≤ vnorm (resid Y Yv Sig gam lam)
          + specNorm Y / (sMin (sampleCov Yt) + lam)
            * (specNorm (sampleCov Yt - Sig) * vnorm (betaStar Sig gam lam)
              + vnorm (sampleCross Yt ytv - gam)) ∧
      vnorm (resid Y Yv Sig gam lam)
        ≤ vnorm (U *ᵥ e) + vnorm (Ev - Eps *ᵥ b)
          + specNorm Y / (sMin Sig + lam) * (lam * vnorm b + vnorm (gam - Sig *ᵥ b)) := by sorry

end SynDigits.Ridge
Source
Fan, Huang, Peng, Wang, Wu, SYN-DIGITS: A Synthetic Control Framework for Calibrated Digital Twin Simulation, arXiv:2604.07513v1, p. 11, Theorem 4.1 (with the preamble on p. 11; setting §3.1–§3.2, pp. 5–6 and §4.1, p. 9; proof Appendix B.1, pp. 32–33)
Human review
  • Endorsed by Shuze Chen · Oct 24, 2026

    Confirmed by the moderator at approval.

  • Endorsed by mikedeng1 · Oct 24, 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