An MCMC Method for Uncertainty Set Generation via Operator-Theoretic Metrics

The paper defines the discounted Koopman kernel and claims convergence as T→∞ when λ > 2 log ρ(K), presenting an informal argument that uses Gelfand’s formula and the assertion that a series ∑ A^t converges if lim ||A^t||_F = 0; it also appeals to a vague “product of two convergent series” claim. This line of reasoning is not mathematically sound: lim ||A_t|| → 0 is necessary but not sufficient for ∑ A_t to converge, and ||K^t||_F ≲ ρ(K)^t requires a polynomial factor and constants, not a strict inequality. Moreover, the paper states the cosine pseudo-metric d_k is bounded in [0,1] without assumptions that ensure nonnegative similarities (the linear/trace kernel they also consider can yield negative inner products). By contrast, the model’s solution supplies a rigorous Jordan–Chevalley-based growth bound, uses Cauchy–Schwarz to prove absolute convergence of the matrix series under λ > 2 max{log ρ(K_1), log ρ(K_2)}, and correctly characterizes d_k ∈ [0,√2] in general, narrowing to [0,1] when the kernel is positive-valued. The HMC construction in the paper and the model agree; the model adds correct gradient chain-rule details. Therefore, the paper’s conclusions are directionally right but the arguments are incomplete or partly incorrect, while the model’s solution is correct and complete on the points audited. Key places in the paper: discounted kernel and informal convergence claim; definition of d_k as bounded in [0,1]; HMC potential U and Hamiltonian H (, , , , ).