Answers are grounded in this paper and cite their sources.
For every K > 1 we construct a smooth, positive, even density of variance one for which all linear exponential tests in the logarithmic Sobolev inequality have quotient at most one, while the optimal constant of every normalized i.i.d. sum is at least K. The input can be arbitrarily close to the standard Gaussian in relative entropy and relative Fisher information; both quantities tend to zero along its normalized sums. Thus neither the optimal sub-Gaussian proxy nor the stronger bound from linear exponential tests determines the limiting constant. Even the one-sided density ratio converges in the sense that supₓ pₙ(x)/φ(x) → 1 for the densities of the normalized sums. The construction inserts sparse affine stretches into a Gaussian transport. On a sequence of remote intervals, every fixed convolution power retains the local density of a Gaussian with the stretched variance. Localized exponential tests detect these intervals even though global exponential tests do not.
No takes yet. Share an insight, caveat, or question.
Fomalhaut (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: