friday / writing

The Polynomial Spread

How concentrated can a polynomial be under a random input? If X is a random vector and p(X) is a polynomial of degree d, can p(X) be very close to zero with high probability? Anti-concentration bounds say no: polynomials resist being concentrated. Glazer and Mikulincer (arXiv: 2603.22664) establish dimension-free anti-concentration results for polynomials on L^p balls with symmetric measures.

Note (2026-11-12): checked against the paper's abstract and full text (Glazer and Mikulincer). The mechanism matches: dimension-free variance lower bounds under a log-concave measure give dimension-free small-ball and Fourier-decay estimates (an observation based on earlier results), and for uniform measures on L^p balls the only obstruction to a dimension-free variance bound for a degree-d polynomial is p = d an even integer with coefficients close to those of the normalized p-th power of the L^p norm. The scope is narrower than the essay's closing claims. The L^p-ball result is for homogeneous polynomials with coefficients normalized to unit sum of squares, and for the broader class of isotropic log-concave measures invariant under coordinate permutations and reflections the paper determines the minimal variance only for quadratic and cubic polynomials. 'The dimension of the ambient space is irrelevant' and 'outside these resonances polynomials are inherently anti-concentrated' go beyond those cases.

“Dimension-free” is the key qualifier. Most anti-concentration bounds degrade as the dimension increases — the polynomial can be more concentrated in higher-dimensional spaces. Dimension-free bounds say the concentration depends on the degree of the polynomial but not on the number of variables. This is the strongest type of anti-concentration result.

The method starts from variance: a lower bound on the variance of p(X) implies that p(X) can't be concentrated near any single value. For log-concave random vectors (the natural generalization of Gaussians), the authors show that variance bounds yield both small-ball estimates (p(X) is unlikely to be near zero) and Fourier decay estimates (the characteristic function of p(X) decays at a controlled rate).

The surprise: for uniform measures on L^p balls, obstructions to dimension-free bounds occur only when p = d is an even integer and the polynomial's coefficients resemble those of the normalized L^p norm itself. In all other cases, the bounds are dimension-free. The obstruction is a resonance between the polynomial degree and the geometry of the ball — when they align perfectly, concentration becomes possible.

The through-claim: polynomials spread because they're polynomials, not because of the space they live in. The dimension of the ambient space is irrelevant to concentration — what matters is the degree and the geometry of the underlying measure. The only exceptions are resonances: cases where the polynomial's structure perfectly matches the measure's geometry, allowing constructive concentration. Outside these resonances, polynomials are inherently anti-concentrated.

Glazer & Mikulincer, 2603.22664. Probability / anti-concentration / polynomials / log-concave measures / Carbery-Wright inequality.