Probabilistic validation of homology computations for nodal domains (Q2456050)

Motivated by applications (say, evolution equations solved numerically), the paper treats accuracy of homology computations for random nodal domains \(N^{\pm}(\omega)\), replaced with their cubical approximations \(Q_M^{\pm}(\omega)\). For a random field \(u:[a,b]^d\times \Omega \to \mathbb{R}\) set \(N^{\pm}(\omega) =\{x\in [a,b]^d: \pm u(x,\omega)\geq 0\}\). Given \(M\in \mathbb{N}\) one may think that \([a,b+M^{-1}(b-a)]^d\) is divided into \((M+1)^d\) cubes included in \(Q_M^{\pm}(\omega)\) according to the sign of \(u\) on the grid of minimal vertices. The goal is to provide sharp lower bounds for \(\mathbb{P}\{H_*(N^{\pm}) \cong H_*(Q_M^{\pm})\}\) in terms of \(M\) and smoothness properties of \(u\). If \(d=1\), then under certain conditions the bound is \(1-8C_0(b-a)^3M^{-2}\), with \(C_0\) involved in the description of the local behavior of \(u\). The authors employ a deterministic validation criterion controlling trajectories' oscillations. Whence the constant \(C_0\) is determined for random Fourier series \(u\) with independent Gaussian coefficients. Specializing to random trigonometric polynomials of order \(N\) shows that accuracy with high confidence follows when \(M\sim N^{3/2}\) for \(N\to \infty\). The same program is realized with \(u\) on a square in \(\mathbb{R}^2\) \((d=2)\) demanded a laborious analysis of various admissible sign patterns of \(u\) on the grid subsquares. Note only the recommendation \(M\sim N^2\) for bivariate trigonometric polynomials \(u\). The probability estimation for periodic \(u\) required investigating certain parametric families of Gaussian vectors.