A polynomial method over formal power series

We give new upper bounds on families of integer vectors with prescribed coordinate multiplicities and a forbidden scalar product. Our method uses coefficient functions over formal power series and recovers the Frankl--Wilson bound for binary vectors. For vectors in $\{-1,0,1\}^n$, we obtain an exponential improvement over the ordinary polynomial count in a wide range; in certain regimes, we determine the lower limit of the exponential base. We use the new bounds to make progress on several questions in combinatorial geometry. In particular, we improve the base of the exponent in Borsuk's partition problem: we construct finite subsets of $\mathbb R^N$ requiring at least $(1.37+o(1))^{\sqrt N}$ parts of smaller diameter for all sufficiently large $N$, and obtain a base of $1.3839$ along a subsequence of dimensions.