Chebyshev-Weighted Collocation Scheme

• Chebyshev-weighted collocation is a numerical method that uses Chebyshev polynomial bases and associated weight functions to approximate functions and solve differential equations.
• It employs a deterministic Weil grid that clusters nodes optimally near domain endpoints, ensuring spectral (exponential) accuracy and robust convergence.
• The framework adapts to various measures via weighted least squares, providing well-conditioned, stable solutions ideal for high-dimensional surrogate modeling and parametric PDEs.

A Chebyshev-weighted collocation scheme is a class of numerical methods for function approximation, interpolation, and the solution of differential equations, which leverage Chebyshev polynomial bases along with an associated Chebyshev weight function in the design of discrete collocation or projection schemes. Such methods ensure spectral (exponential) accuracy for analytic problems, optimal node clustering near domain endpoints, and improved conditioning. The distinguishing feature is the use of discretizations, quadrature, and norm weightings naturally aligned with the Chebyshev measure, allowing both deterministic and weighted least-squares collocation frameworks. A key example is the deterministic Weil grid collocation developed for multivariate discrete least-squares approximation with rigorous stability and optimal convergence guarantees (Zhou et al., 2014).

1. Chebyshev Polynomial Spaces and Weighted Collocation

Chebyshev-weighted collocation schemes utilize polynomial spaces defined by Chebyshev polynomials $T_k(x)=\cos(k\,\arccos\,x)$ with domain $x\in[-1,1]$, and are orthogonal with respect to the Chebyshev density $\rho_c(x) = 1/(\pi\sqrt{1-x^2})$. In the multivariate case over $\Gamma = [-1,1]^d$, the tensor-product density $\rho_c(x) = \prod_{j=1}^d \rho_c^j(x_j)$ is used, with the polynomial basis $T_\alpha(x) = \prod_{j=1}^d T_{\alpha_j}(x_j)$.

Given a function $f$, the best $L^2_{\rho_c}$-approximation in some finite-dimensional space $$P^\Lambda = \mathrm{span}\{T_\alpha\}_{\alpha\in\Lambda}$$ ($|\Lambda|=N$), is defined by

$$P^\Lambda f = \arg\min_{p\in P^\Lambda} \|f - p\|_{L^2_{\rho_c}}.$$

In the discrete collocation approach, this $L^2$ minimization is replaced by weighted or unweighted discrete least squares using carefully chosen collocation nodes $y_k$, forming the surrogate

$$P^\Lambda_m f = \arg\min_{p\in P^\Lambda} \sum_{k=0}^m (p(y_k)-f(y_k))^2.$$

This formulation naturally extends to a weighted least-squares version for general target density $\rho$ via

$$P^\Lambda_{w,m} f = \arg\min_{p\in P^\Lambda} \sum_{j=0}^m w_j (p(y_j)-f(y_j))^2,$$

where the weights satisfy $w_j = \rho(y_j)/\rho_c(y_j)$ (Zhou et al., 2014).

2. Deterministic Collocation Grids and Weil Construction

An essential advance is the deterministic construction of collocation nodes ($\Theta_M$ or the "Weil grid") that achieve stability and asymptotic equidistribution with respect to the Chebyshev measure. For a given maximal polynomial degree $q$ and dimension $d$, consider a prime $M > 2q+1$:

• Set $m = \lfloor M/2 \rfloor$.
• For $j=0,\dots,m$,

$$x_j = 2\pi (j, j^2, \dots, j^d)/M \pmod{2\pi}, \qquad y_j = \cos(x_j) \in [-1,1]^d.$$

• The grid $\Theta_M = \{y_j : 0 \leq j \leq m\}$ deterministically samples the Chebyshev measure in $d$ dimensions (Zhou et al., 2014).

To ensure stability, the number of collocation points must scale quadratically in the basis dimension: $M \geq 4^{d+1} d^2 N^2$, thus $m = O(N^2)$. This scaling guarantees that the discrete normal matrix $A = D^T D$ (where $D_{k\alpha}=T_\alpha(y_k)$) is well-conditioned and the normal equations are uniquely solvable.

3. Weighted Least Squares and Measure Correction

The deterministic grid $\Theta_M$ equidistributes asymptotically to the product Chebyshev measure as $M\to\infty$. For density $\rho$ absolutely continuous with respect to $\rho_c$, the weighted least-squares approach corrects for the difference between desired and Chebyshev measures:

• Weights: $$w_j = \rho(y_j)/\rho_c(y_j) = \pi^d \rho(y_j) \prod_{l=1}^d \sqrt{1-(y_j)_l^2}$$.
• In particular, for uniform measure $\rho \equiv 2^{-d}$, $w_j = (\pi/2)^d \prod_{l=1}^d \sqrt{1-(y_j)_l^2}$.

Asymptotic equidistribution (via Weyl's criterion and Weil's estimate) legitimizes this approach: unweighted least squares approximates integration with respect to $\rho_c$; for any other $\rho$, discrete weights $w_j$ correct the quadrature (Zhou et al., 2014).

4. Stability, Convergence, and Conditioning

The scheme's theoretical foundation is rigorous:

• If $M\geq 4^{d+1} d^2 N^2$, then the normalized matrix $A$ satisfies $\| (2^{d+1}/M)A - I \|_2 \leq 1/2$.
• Therefore, $A$ is uniformly well-conditioned, guaranteeing robust numerical solvability.
• Convergence: For the continuous Chebyshev projection $P^\Lambda f$ and its discrete surrogate $P^\Lambda_m f$,

$$\|f - P^\Lambda_m f\|_{L^2_{\rho_c}} \leq (1 + 4/(d^2 N)) \|f - P^\Lambda f\|_{L^\infty(\Gamma)}.$$

• Extension to more general $\rho$ with $\rho \leq C \rho_c$ ensures

$$\|f - P^\Lambda_m f\|_{L^2_\rho} \leq \sqrt{C}(1 + 4/(d^2 N)) \|f - P^\Lambda f\|_{L^\infty}.$$

These properties remove dependence on probabilistic qualifiers commonly seen in Monte Carlo or random grid approaches (Zhou et al., 2014).

5. Algorithmic Realization

The practical implementation follows these steps:

1. Choose $M =$ smallest prime $\geq \max(2q+1, 4^{d+1} d^2 N^2)$.
2. Compute $m = \lfloor M/2 \rfloor$.
3. For $j=0,\ldots,m$, evaluate $y_j$, $f(y_j)$, and weights $w_j$ (if necessary).
4. Form the design matrix $D$ with $D_{j,\alpha} = T_\alpha(y_j)$ for all multi-indices $\alpha \in \Lambda$.
5. Solve the (possibly weighted) normal equations: $(D^T W D)c = D^T W f$; $W$ is diagonal with $w_j$ (or $I$ if unweighted).
6. Construct the polynomial surrogate $$f_\Lambda(x) = \sum_{\alpha \in \Lambda} c_\alpha T_\alpha(x)$$.

The assembly of $D$ requires $O((m+1)N)$ operations; solving the linear system is $O(N^3)$ (Zhou et al., 2014).

6. Extensions and Generalization

The framework naturally generalizes to other orthogonal polynomial bases and measures:

• Any orthonormal system $\{\phi_\alpha\}$ under measure $\rho$ can be incorporated by using the same deterministic grid $\Theta_M$ and re-weighting via $w_j = \rho(y_j)/\rho_c(y_j)$.
• The essential property is asymptotic equidistribution of the grid with respect to the desired measure; this forms the basis for weighted least squares in arbitrary orthogonal polynomial spaces.
• For example, Legendre polynomials (uniform $\rho$) use the same collocation nodes with appropriate weights, and convergence estimates analogous to those for Chebyshev bases follow.

The deterministic construction avoids the variability of random sampling, yielding guarantees on conditioning and convergence that are fully deterministic (Zhou et al., 2014).

7. Significance, Limitations, and Applications

The Chebyshev-weighted collocation scheme provides a robust approach for high-dimensional approximation in uncertainty quantification, parametric PDEs, and other multivariate settings where polynomial surrogates are used. Its main advantage is the deterministic, theoretically well-founded construction:

• Avoids probabilistic convergence language and ensures stability for polynomial degrees scaling as $O(\sqrt{m})$, with practical algorithms for high-dimensional projection.
• Key applications include surrogate modeling in stochastic problems, sparse grids, and well-conditioned discrete projections for function approximation.
• While the number-theoretic construction leads to a quadratic scaling of collocation points with respect to the basis size, this suggests a tradeoff between stability and sample complexity. Other deterministic grids may further optimize this balance if they also asymptotically equidistribute to the target measure.

These advances, as detailed by Zhou–Narayan–Xu, establish a rigorous, practical, and highly extensible method for polynomial discrete least-squares approximation with Chebyshev-weighted collocation (Zhou et al., 2014).