Quadratic variations of rough generalized Hermite processes: hidden long memory and a universal Gaussian boundary field

Abstract

We study centered quadratic variations of fractionally filtered generalized Hermite processes. For every chaos order $q\ge2$, every admissible anisotropic monomial kernel $g_{\boldsymbolγ}$, and every fixed $0<h\le1/2$, we prove that $N^{2h+θ-1}V_N$ converges to a Rosenblatt process with self-similarity exponent $1-θ$, without subtracting any chaos projection. Thus the visible roughness $h$ determines the normalization, whereas the latent singularity exponent $θ$ determines the memory of the limit. The proof uses a sharp paired-filter threshold and charged diagram power counting. At $h=0$, the bare kernel has logarithmically divergent energy. Variance-normalized $h$-regularization and a hard endpoint cutoff have the same residue and converge in finite-dimensional distributions to the universal Gaussian field $G_t=(ζ_t-ζ_0)/\sqrt2$. Although $G$ has no stochastically continuous modification, the iterated boundary-first quadratic energies converge to $\sqrt3B$. In the isotropic case, simultaneous limits $h_N\downarrow0$ are governed by $λ_N=h_NN^{1/2-θ}$: $λ_N\to0$ yields a Brownian limit, $λ_N\toλ\in(0,\infty)$ yields an independent Brownian-Rosenblatt sum, and $λ_N\to\infty$ yields a Rosenblatt limit after division by $λ_N$. We also obtain an explicit polynomial rate in fixed Malliavin-Sobolev norms and functional convergence in little-Hölder spaces.