GPT 6 Astra Ultra attempt on Math-5794

Erdős Problem 651: convex configurations in higher dimensions

Statement.

Let $f_k(n)$ be the least number of points in general position in $\mathbb R^k$ that forces $n$ points in convex position. Must $f_k(n)>(1+c_k)^n$ hold for some $c_k>0$?

Claim and source.

For every fixed $k\geq3$, the answer is no. The external input is Pohoata and Zakharov, Convex polytopes from fewer points, https://arxiv.org/abs/2208.04878, Theorem 1.1: for every $\varepsilon>0$, all sufficiently large $n$ satisfy \[ f_3(n)\leq\lceil2^{\varepsilon n}\rceil. \] We spell out the projection and quantifier arguments that give the disproof in every higher dimension. The planar case is different; no disproof for $k=2$ is claimed. See https://www.erdosproblems.com/651.

Projection preserves the needed hypotheses.

Let $X\subset\mathbb R^k$ be a finite general-position set of at least $k+1$ points, with $k\geq3$. Every four points of $X$ are affinely independent: otherwise adding another $k-3$ points would yield $k+1$ points in an affine subspace of dimension at most $k-1$, contrary to general position.

Choose a linear map $L:\mathbb R^k\to\mathbb R^3$ for which every four-point subset remains affinely independent. Such a map exists: for each fixed four-point subset, the determinant of its three image-difference vectors is a nonzero polynomial in the entries of $L$, since those original difference vectors are independent. The zero sets of finitely many nonzero polynomials cannot cover the whole space of linear maps. The resulting image is a set of distinct points with no four coplanar.

If $Y\subseteq X$ projects to points in convex position, then $Y$ itself is in convex position. Indeed, an expression $y=\sum_{z\in Y\setminus\{y\}}\lambda_z z$, with nonnegative $\lambda_z$ summing to one, would project to the same forbidden convex combination for $L(y)$. Thus, in the asymptotic range relevant here, \[ f_k(n)\leq f_3(n). \]

Excluding every positive exponential constant.

Fix any proposed $c_k>0$, and take $\varepsilon=\tfrac12\log_2(1+c_k)>0$. For all sufficiently large $n$, \[ f_k(n)\leq\lceil2^{\varepsilon n}\rceil <2^{2\varepsilon n}=(1+c_k)^n. \] This contradicts the proposed exponential lower bound, even if it was required only eventually. Consequently no such positive constant exists in dimension $k\geq3$.

Scope.

The disproof is complete relative to the stated Pohoata--Zakharov theorem. The generic projection and the comparison of growth rates are proved here; the deep three-dimensional theorem is not re-proved or presented as a new result.

Completion estimate.

COMPLETION: 100%