Yonsei University
Chae-Yeon Xon
Yonsei University, School of Economics
Home/Resources/Geometry, Dynamics, and Operators for Economic Models/III-14
Function Spaces, Inverse Problems, and Risk Certificates · Chapter III-14

Conditional Generators, Poincare Inequalities, and Graph Certificates

Sufficiency can be tested through variation within an observation fiber. A conditional energy penalizes derivatives or differences along hidden directions. Under a spectral gap, small energy controls distance from a fiberwise constant function; on a finite graph the same logic becomes an exact Laplacian certificate.

Conceptual map

  1. III-14.01Conditional energy
  2. III-14.02Generator assumptions
  3. III-14.03Spectral gaps
  4. III-14.04Residual bounds
  5. III-14.05Finite-graph analogues

1. Fiberwise variation as an energy problem

When an observed report fixes a fiber of latent states, a generator can measure how much a target varies along that fiber. A spectral gap converts derivative energy into a bound on unresolved conditional variance.

Let \(\pi\) be a probability law on a fiber and \(L\) a densely defined self-adjoint Markov generator on \(L^{2}(\pi)\), with \(L\)1=0 and spectrum in (\(-\infty\),0]. Its Dirichlet form is

\[E(f,f)=-\langle f,Lf\rangle _{L^{2}(\pi )}.\](1)
Definition 1 · Spectral-gap Poincaré inequality

The fiber has gap \(\lambda _{1}>0\) when every \(f\) in the form domain satisfies

\[\operatorname{Var}_{\pi }(f)\le \lambda _{1}^{-1}E(f,f).\](2)

The null space must consist of constants on the connected fiber. Multiple connected components create multiple zero modes.

2. Finite-graph certificate

For a reversible finite graph, let \(Q=-L\) be its positive semidefinite Laplacian in \(L^{2}(\pi)\). Write its orthonormal eigenvectors as \(\phi _{0}=1,\phi _{1},\ldots\) with eigenvalues \(0=\lambda _{0}<\lambda _{1}\le\cdots\).

Proposition 1 · Graph Poincaré bound

If the graph is connected, equation (2) holds with the sharp constant \(\lambda _{1}^{-1}\).

Proof. Center \(f\) and expand it as \(\sum _{j\ge 1}a_{j}\phi _{j}\). Then \(\operatorname{Var}_{\pi}(f)=\sum _{j\ge 1}a_{j}^{2}\), while \(E(f,f)=\sum _{j\ge 1}\lambda _{j}a_{j}^{2}\ge \lambda _{1}\operatorname{Var}_{\pi}(f)\). Equality holds for \(\phi _{1}\). ∎

3. A three-node path and a disconnected fiber

Take uniform \(\pi\) on three nodes and the path Laplacian

\[Q=\begin{bmatrix}1 & -1 & 0 \\ -1 & 2 & -1 \\ 0 & -1 & 1\end{bmatrix}.\](3)

Its eigenvalues are 0, 1, and 3. For \(f=(1,0,-1)^{\top}\), the mean is zero, \(\operatorname{Var}_{\pi}(f)=2/3\), and \(\langle f,Qf\rangle _{\pi}=2/3\). The Poincaré bound is attained because \(f\) is a gap eigenvector.

Delete one edge. The graph splits into two components and the zero eigenvalue has multiplicity two. A function that is constant on each component but takes different component values has zero energy and positive global variance. Component means must then enter the certificate explicitly.

4. Continuous generators and certification residuals

For the standard normal law, the Ornstein–Uhlenbeck generator \(Lf=f^{\prime\prime}-xf^{\prime}\) has gap one on its natural \(L^{2}\) domain. The inequality \(\operatorname{Var}(f)\le\mathbb{E}[(f^{\prime})^{2}]\) follows, with equality for \(f(x)=x\). Applying this template to another conditional diffusion requires verification of invariance, self-adjointness or reversibility, operator domain, and the claimed gap.

Suppose a decoder assigns one predicted target value to each observed report. The residual target on a fiber can be decomposed into its component means and a mean-zero part. Bound the latter by (2), report the former directly, and add numerical generator or graph-approximation error as a separate term.

Failure case · A discrete spectrum assumed from notation

A differential expression called a generator need not have compact resolvent or a positive spectral gap. Boundary conditions, domain, invariant measure, and connectedness determine whether the eigenvalue expansion used by the certificate exists.

5. Implementation, exercises, and sources

For a graph approximation, preserve reversibility in the weights, identify connected components, compute component means, and use the smallest positive eigenvalue within each component. Under grid refinement, report changes in the energy, gap, and target bound separately.

Download the volume verification script →

Exercises

  1. Diagonalize (3) and verify the equality case by direct multiplication.
  2. Remove each edge in turn and construct a zero-energy function with positive global variance.
  3. For \(f(x)=x^{2}\) under the standard normal law, compute variance and Ornstein–Uhlenbeck energy and compare them with (2).
Partial solutions

1. The vectors (1,1,1), (1,0,\(-1\)), and (1,\(-2\),1) have eigenvalues 0, 1, and 3. Normalize them under the uniform inner product. 3. For a standard normal variable, \(\operatorname{Var}(X^{2})=3-1=2\) and E[\((2X)^{2}\)]=4. The gap-one Poincaré bound gives \(2\le4\); equality is reserved for the first nonconstant eigenspace.

  1. Dominique Bakry, Ivan Gentil, and Michel Ledoux, Analysis and Geometry of Markov Diffusion Operators, Chapters 2–4.Generators, Dirichlet forms, and Poincaré inequalities.
  2. Fan Chung, Spectral Graph Theory, Chapter 1.Graph Laplacians and spectral gaps.
  3. Mu-Fa Chen, From Markov Chains to Non-Equilibrium Particle Systems.Gap inequalities for Markov processes.

6. Audit checkpoint

State the fiber law, generator or graph, domain, boundary conditions, reversibility, connected components, component means, energy, positive gap, residual, and discretization error.

7. Scope boundary

The chapter proves finite-graph Poincaré certification and records the assumptions needed for a conditional-generator analogue. General elliptic regularity and PDE spectral theory remain outside scope.

Prerequisites