Many conjectured inequalities about how entropy and Fisher information change as a random vector is blurred by Gaussian noise (the "heat flow") are now attacked by computer. The method converts each conjecture into a polynomial identity and asks a semidefinite program (SDP), a convex optimization over positive semidefinite matrices, to find a sum-of-squares (SOS) certificate, that is, a rewriting of the quantity as a sum of squared polynomials plus terms that integrate to zero. This approach settled third-order cases that had resisted hand proofs since McKean's 1966 conjecture. It then stalled. Some cases run out of memory, others return "no certificate" with no explanation, and none of the certificates has been checked exactly. When a search fails, we cannot tell whether the inequality is false or whether the search was simply too weak.
This note proposes using the extremal distributions of the log-concave class to steer the search. These are the log-affine profiles (exponential-type densities on segments and cones) and their Gaussian-smoothed and linearly transformed versions, including anisotropic Gaussians with unequal variances across axes. Any valid certificate must be tight, or nearly tight, on the distributions where the inequality is sharp. The derivative values of these profiles therefore tell us in advance which directions the SOS Gram matrix must annihilate. That information can serve three purposes. It can shrink the SDP by restricting it to the correct face of the SOS cone. It can remove the ill-conditioning that blocks exact rational certificates. When the inequality fails on one of these profiles, it can produce a rigorous counterexample. The result would be a pipeline that ends every case either with a formally checked proof or with a certified counterexample, instead of an unexplained solver failure.
Core Claim
The claim is that sharpness witnesses can be computed before the SDP is solved, and that they predict the kernel of the certificate. Here are the objects involved. Consider a quantity $Q(p_t)$, for example $\frac{d^2}{dt^2}\log I(X_t)$, where $I$ is Fisher information and $p_t$ is the density of $X_t = X + \sqrt{t}Z$ with $Z$ standard Gaussian. Guo, Yuan and Gao write $Q$ as an integral of a polynomial $F(v)$ in the "jet" $v$, the vector of partial derivatives of $\log p_t$ up to some order. They then search for
$$F(v) = m(v)^\top G\, m(v) + \sum_j \lambda_j R_j(v) + \sum_k \sigma_k(v)\, L_k(v), \qquad G \succeq 0.$$
In this decomposition, $m(v)$ is a vector of monomials, $G$ is the Gram matrix, the $R_j$ are integration-by-parts identities that integrate to zero, and the $L_k$ encode log-concavity ($\nabla^2 \log p_t \preceq 0$) with SOS multipliers $\sigma_k$.
Suppose a family of densities $p^{(\theta)}$ makes $\int F$ vanish. Then on the jets those densities produce, every term on the right must integrate to zero. Because $G$ is positive semidefinite, $m(v)$ must then lie in the kernel of $G$ almost everywhere along those jets. The candidate family is Gaussian-smoothed log-affine profiles together with their products and linear images. Its jets can be computed in closed form from error functions. Those jets therefore give explicit vectors that $G$ must annihilate before any solve.
The hypothesis is that these extremal profiles are exactly where the log-concave versions of the conjectures are sharp. The $1/n$ scaling in Guo et al.'s weakened conjecture hints at this, since it points to near-extremizers that concentrate in fewer dimensions or spread unequally across axes. The hypothesis can be falsified directly. The predicted kernel can be compared with the numerical kernel of the Gram matrices that already-successful SDPs return.
Position Relative to Prior Work
Three groups of papers play different roles here.
The Guo–Yuan–Gao line and Liu–Gao supply the method and also show the gap. Prove Costa's Entropy Power Inequality and High Order Inequality for Differential Entropy with Semidefinite Programming certifies third-order sign conjectures $D(3,n)$ for $n=2,3,4$ but overflows memory on $D(3,5)$. Lower Bound for Derivatives of Costa's Differential Entropy and Lower Bounds on Multivariate Higher Order Derivatives of Differential Entropy refute McKean's conjecture for $n>1$, $m>2$. They prove dimension-scaled versions under log-concavity, but solvers return negative answers on cases such as $C_1(4,2)$. Square Root Convexity of Fisher Information along Heat Flow in Dimension Two could not certify log-convexity of Fisher information for $n=2$ and fell back to square-root convexity. All of these papers enforce log-concavity only through pointwise multipliers. None asks which distributions make the inequality tight, and none produces exact certificates. We keep their encoding unchanged and add the extremal information.
The extreme-point papers motivate the choice of witness family. They also limit what we can claim. Geometric and Functional Inequalities for Log-Concave Probability Sequences uses Krein–Milman to show that extreme points of linearly constrained log-concave sets are log-affine. On a Conjecture of Feige for Discrete Log-Concave Distributions uses that fact to reduce a sharp $1/e$ bound to the parametric family $p(k)=Cp^k$. Both apply only to functionals that are linear in the density. Heat-flow derivatives are not linear in the density, so we do not claim a Krein–Milman reduction. We use extremal profiles as witnesses that shape and test the certificate, not as a proof that checking them suffices. On the Entropy Power Inequality for the Rényi Entropy of Order [0, 1] (On the Entropy Power Inequality for the Rényi Entropy of Order [0, 1]) is analytic and limited to scalar entropy power. It shows that log-concave structure organizes sharp constants, which supports the choice of witness family.
The vector entropy-power work fixes the application regime. A Vector Generalization of Costa's Entropy-Power Inequality with Applications shows that matrix, non-proportional perturbations behave differently from scalar ones. The failure of McKean's conjecture is a scalar-versus-vector effect of the same kind. Our witness family therefore deliberately includes anisotropic products, so the scalar assumption that variance is spread equally across axes is never imposed on the search.
Two further papers supply tooling. Efficient Data Structures for Representation of Polynomial Optimization Problems: Implementation in SOSTOOLS provides the problem-setup tooling. From LLM-Generated Conjectures to Lean Formalizations provides the stage that turns a numerical certificate into exact rational numbers and checks it in Lean. The ill-conditioning that the Zuo et al. paper reports on boundary polynomials is exactly what knowing the kernel in advance should cure.
Research Program
The first component is a witness library. We will compute closed-form or rigorously bounded jets of $\log p_t$ for Gaussian-smoothed one-sided exponentials and uniform segments, which are log-affine with zero slope. We will extend these to products with independent per-axis parameters and to linear images, and include anisotropic Gaussians as limiting members. The integrals $\int F$ along each family will be evaluated with interval arithmetic, so that a negative value is a proof of a counterexample.
The second component is extremal facial reduction. For each conjecture we will locate the parameter values where $\int F$ reaches zero or its infimum over the family. We sample the jets at those values and derive linear constraints $G\,m(v)=0$ and $\sigma_k$-vanishing conditions. We then re-pose the SDP on the reduced face, which has a smaller Gram matrix and fewer free multipliers.
The third component is exact recovery. On the reduced, well-conditioned face, we round the solution to rational numbers using the rational-recovery approach of the Zuo et al. paper. The resulting identity is checked in Lean 4. The kernel constraints are imposed exactly rather than numerically, which is why rounding should succeed where it failed before.
The fourth component is diagnosis. When the reduced SDP is still infeasible and the witness family shows no violation, the dual SDP solution gives pseudo-moments over jets. We test whether these are realized by any member of the family. This separates three outcomes: the inequality is false outside the family, the encoding is missing integral identities (a gap the 2020 paper already suspected), or the true polynomial is nonnegative but has no SOS certificate. The last outcome is the convex-but-not-SOS-convex situation that Ahmadi and Parrilo showed can occur (A convex polynomial that is not sos-convex).
Baselines
The naive reference point is to rerun the published Guo et al. formulations unchanged with a modern solver. The reported figures are the bar: $D(3,4)$ with 80 variables and about 1,974 constraints solved in 1.8 seconds, a configuration with 38 variables and 512 constraints solved in under 10 seconds, and $D(3,5)$ overflowing memory beyond 7,000 constraints (Prove Costa's EPI…, Lower Bound for Derivatives…).
The strongest existing approach to the scaling problem is the same formulation built with SOSTOOLS 4.00's dpvar structures. That paper reports roughly a 100x reduction in setup ("parsing") time and setup of a 50-variable robust-stability problem in 374 seconds, where YALMIP exceeded 3,600 seconds at 13 variables (SOSTOOLS data structures). It speeds up problem construction but not the solve, so it may still fail on $D(3,5)$. Both baselines will run on the same machine with the same solver, the same time and memory limits, and the same monomial degree bounds as our reduced programs. Beating them means one of two things: certifying a case they cannot, such as $D(3,5)$, $C_1(4,2)$, or log-convexity for $n=2$, or solving the solved cases at least several times faster with smaller Gram matrices.
For exact certification, the strongest baseline is NSPI's numerical refinement plus rational recovery applied directly to the unreduced certificates. On its own benchmark NSPI reports a 25.95% pass rate on PolyIneq-Synth and 11.7% on 10-variable problems, where Maple, Z3 and DeepSeek-Prover-V2 reach 0% (Zuo et al.). There is no reported result on Lacuna for NSPI or any exact-recovery method on entropy-derivative certificates, so we establish this number ourselves. Success means our reduced certificates pass Lean checking on cases where direct recovery fails.
The key ablation replaces extremal kernels with kernels taken from the numerical rank of an already-solved Gram matrix. This is generic facial reduction after the fact. It isolates whether the a priori extremal information matters, or whether any kernel estimate would do. A second ablation drops the anisotropic members from the witness family, to test whether the scalar-versus-matrix distinction actually affects which kernel is found.
Why not just…?
Why not just buy more memory and a better solver? For $D(3,5)$ this could work, and we will try it as the naive baseline. It cannot explain the cases that solvers mark infeasible, such as $C_1(4,2)$ and Fisher log-convexity for $n=2$. In those cases the SOS program is the problem, not the hardware. More compute also does not address the floating-point ill-conditioning that blocks exact certificates.
Why not just search for counterexamples numerically over random or neural-network-parametrized log-concave densities? This is a real threat for the falsification half of the project. We will run it at a matched compute budget against the witness family. Our advantage would show up in two ways: certified (interval-arithmetic) violations, and kernel vectors that feed back into the SDP. Generic numerical search provides neither. If unstructured search finds violations that the extremal family misses, our central hypothesis about where sharpness lives is wrong, and we will report that.
Why not just apply the Krein–Milman reduction and skip SDPs? The reduction applies only to functionals linear in the density, as in the Feige and dilation results. Heat-flow derivatives are quotients and products of derivatives of $p_t$. Checking only the extremal family therefore proves nothing about the general case, and we use it solely to shape and test certificates.
Why not just run NSPI end to end? NSPI's language-model-proposed decompositions target competition-style inequalities in up to 10 variables. Entropy-derivative certificates involve 38 to 80 or more jet variables plus integral identities that NSPI has no notion of. Its rational recovery stage is part of our pipeline, and the ablation above shows whether it succeeds without extremal kernel constraints.
Proposed Experiments
The first decisive check tests the hypothesis at no cost before anything new is solved. On already-certified cases ($D(3,n)$ for $n=2,3,4$, $C_3(3,n)$ for $n=2,3,4$, $C_3(4,2)$, and square-root convexity for $n=2$), we compare the subspace predicted by the witness family against the numerical kernel of the Gram matrices returned by re-solving the published SDPs. If the predicted subspace does not substantially overlap the observed kernel, the extremal approach is abandoned. The fallback is generic facial reduction, reported honestly.
The second check is scaling. Using the reduced formulation we attempt $D(3,5)$, $D(4,2)$, and $D(5,1)$, recording Gram size, constraint count, and wall time against both SDP baselines.
The third check is exactness. Every reproduced and new certificate goes through rational recovery and Lean 4 checking. Our pass rate is compared with direct NSPI-style recovery on the unreduced versions.
The fourth check is diagnosis on the open targets: Fisher-information log-convexity for $n=2$, and $C_1(4,2)$ and $C_2(3,2)$ under log-concavity. Each of these ends one of three ways: an interval-certified counterexample within the witness family, a Lean-checked certificate, or a dual pseudo-moment analysis that names the failure as a missing identity, a non-SOS obstruction, or a possible violation outside the family. Any one of these outcomes is more informative than the current "solver returned negative."
References
- Prove Costa's Entropy Power Inequality and High Order Inequality for Differential Entropy with Semidefinite Programming
- Lower Bound for Derivatives of Costa's Differential Entropy
- Lower Bounds on Multivariate Higher Order Derivatives of Differential Entropy
- Square Root Convexity of Fisher Information along Heat Flow in Dimension Two
- Geometric and Functional Inequalities for Log-Concave Probability Sequences
- On a Conjecture of Feige for Discrete Log-Concave Distributions
- On the Entropy Power Inequality for the Rényi Entropy of Order [0, 1]
- A vector generalization of costa's entropy-power inequality with applications
- Efficient Data Structures for Representation of Polynomial Optimization Problems: Implementation in SOSTOOLS
- From LLM-Generated Conjectures to Lean Formalizations: Automated Polynomial Inequality Proving via Sum-of-Squares Certificates
- A convex polynomial that is not sos-convex