On the Local Structure and Approximation Stability of Block Isotropic Gaussian Fields

Skew-symmetric functions are a class of functions defined on a product space $M \times M$ that are antisymmetric with respect to the order of their inputs. In [13], the authors proved that non-deterministic skew-symmetric Gaussian fields cannot be stationary or isotropic and proposed an alternative notion: stationarity (isotropy) in each component space. Our work focuses on local quadratic approximations of the associated Gaussian fields. Local quadratic approximations to random fields are random polynomials parametrized by a jointly sampled gradient vector and Hessian matrix. We characterize the distribution of the corresponding random vectors and random matrices. Then, we study the error in the quadratic approximation, which is also a Gaussian field. We investigate the error induced by the quadratic approximation in three senses: the pointwise error, the maximal error over an ellipsoidal region, and the worst-case error for multivariate Gaussian inputs at a given confidence level. Next, we explore the limiting behavior of the worst-case error as the distance between an expansion point and evaluation points approaches zero and infinity. Finally, we study how, as the input dimension increases, the variance of multivariate Gaussian distributions must be restricted to keep the worst-case error bound constant.

💡 Research Summary

The paper investigates the local quadratic approximation of a special class of Gaussian processes called Block Isotropic Gaussian Processes (BIGPs), which arise from skew‑symmetric functions defined on a product space M × M. Classical results show that non‑deterministic skew‑symmetric Gaussian fields cannot be stationary or isotropic in the usual sense. To overcome this limitation the authors introduce “block stationarity (isotropy)”, a weaker symmetry that requires each component space (the two copies of M) to be stationary and isotropic while preserving the overall skew‑symmetry across the two blocks.

The authors first lay out the standard theory of Gaussian random fields, defining mean and covariance functions, weak/strict stationarity, isotropy, and mean‑square differentiability. They prove that for an isotropic field the covariance depends only on the squared Euclidean distance, and that differentiability of the covariance kernel guarantees the existence of Gaussian derivatives (gradient and Hessian) with explicit covariance formulas.

In the block‑isotropic setting the index set is ℝ^{2D} and points are written as (z₁, z₂). The kernel takes the form
k((z₁,z₂),(z₁′,z₂′)) = h(‖z₁−z₁′‖² + ‖z₂−z₂′‖²).
Evaluating the process at a “diagonal” point (z,z) respects the skew‑symmetry because f(z₁,z₂)=−f(z₂,z₁). The key technical contribution is the characterization of the first‑ and second‑order derivatives at such a point:

• The gradient ∇f(z) is an isotropic Gaussian vector: ∇f(z) ∼ N(0,σ²I_D).
• The Hessian ∇²f(z) is a symmetric random matrix drawn from the Gaussian Orthogonal Ensemble (GOE(D)).
• The gradient and Hessian are independent.

Consequently the second‑order Taylor expansion
\tilde f(x)=f(z)+∇f(z)·(x−z)+½(x−z)^T∇²f(z)(x−z)
is itself a Gaussian random field, parametrized by the random gradient and Hessian.

The error field e(x)=f(x)−\tilde f(x) is also Gaussian. By differentiating the original kernel three and four times, the authors obtain a closed‑form expression for the covariance of e(x). This yields:

1. Pointwise error: e(x) ∼ N(0,Var_e(‖x−z‖)), where Var_e is a polynomial in the distance that starts at order ‖x−z‖⁶, reflecting the fact that the quadratic truncation discards terms of order three and higher.
2. Uniform error over an ellipsoid: For a region ℰ={x:(x−z)^TΛ⁻¹(x−z)≤r²}, a high‑probability bound (1−δ) on sup_{x∈ℰ}|e(x)| is derived using Gaussian concentration and covering‑number arguments. The bound scales with r³ and the eigenvalues of Λ, which encode the anisotropy of the underlying kernel.
3. Worst‑case error for Gaussian inputs: When the evaluation point X is drawn from N(z,Σ), the authors consider the maximal error within a confidence ellipsoid of X. They show that the worst‑case error grows with the spectral norm of Σ relative to Λ, and they provide an explicit formula for the required confidence radius to keep the error below a prescribed threshold.

The asymptotic analysis of the error as the distance d=‖x−z‖ varies yields two regimes:

• Near‑field (d→0): The error behaves like O(d³), confirming that the quadratic approximation is locally accurate.
• Far‑field (d→∞): Higher‑order derivatives dominate and the error approaches a constant plateau determined by the kernel’s tail behavior.

Finally, the paper studies the high‑dimensional regime. Because the eigenvalues of a GOE matrix scale as √D, the variance of the gradient and the entries of the Hessian grow with dimension. To keep the uniform error bound constant as D increases, the input covariance Σ must be scaled down proportionally to 1/D. This result provides a practical guideline for designing experiments or algorithms that rely on quadratic surrogates of high‑dimensional skew‑symmetric Gaussian fields.

In summary, the authors deliver a complete probabilistic description of the local quadratic approximation of block‑isotropic Gaussian fields, quantify the approximation error in several practically relevant settings, and reveal how dimensionality influences the required scaling of input variance. Their work bridges random field theory, random matrix theory (via GOE), and approximation theory, offering a solid theoretical foundation for applications such as preference learning, network flow modeling, and Bayesian optimization where skew‑symmetric Gaussian priors are employed.