← All problems

Log-convexity of Fisher information along heat flow

11 years old · traced to

For every smooth positive density ff on Rd\mathbb{R}^d, must the Fisher information t↦I(f∗γt)t\mapsto I(f*\gamma_t) be log-convex along the heat flow?

References

Progress summary

Refreshed
Open · solution reported

A new preprint disproves the conjecture in every dimension from two upward, while the one-dimensional case remains true.

Cheng and Geng formulated the conjecture in 20152015: Fisher information along the heat flow should be log-convex for every smooth positive density. The assertion is true in one dimension but fails in higher dimensions.

Known results

  • Ledoux, Nair, and Wang proved log-convexity in dimension 11.
  • Liu and Gao proved the weaker square-root convexity inequality 2I[f]D[f]≥Q[f]22I[f]D[f]\ge Q[f]^2 in dimension 22.
  • Gu and Sellke disproved several related one-dimensional complete-monotonicity conjectures.

2026 multidimensional counterexample

The arXiv preprint constructs a smooth, strictly positive, Gaussian-decaying density on cmathbbR2cmathbb{R}^2 with I[f]D[f]−Q[f]2<0I[f]D[f]-Q[f]^2<0. Since I′(t)=−Q[ft]I'(t)=-Q[f_t] and I′′(t)=D[ft]I''(t)=D[f_t], log-convexity fails for sufficiently small t>0t>0; tensorization extends the counterexample to every dimension d≥2d\ge2. The preprint states that GPT-5 found the explicit counterexample.

Current status (as of May 2026): The conjecture is settled negatively for every d≥2d\ge2 and positively for d=1d=1.

Sources

Solutions 0

No solutions have been posted yet.