Chebyshev Shape Parametrization

• Chebyshev shape parametrization is a method that represents geometric forms and functions using Chebyshev polynomials and cosine-angle variables for stable, compact approximations.
• It underpins a variety of applications—from bivariate approximation and polynomial knot embeddings to nuclear surface modeling—by enabling efficient FFT-based coefficient recovery and interpolation.
• Recent advances incorporate Chebyshev parametrization into neural architectures and terrain analysis, leveraging orthogonal expansions and spectral filtering to enhance accuracy and interpretability.

Searching arXiv for recent and foundational papers on Chebyshev shape parametrization and related formulations. I’m going to look up papers on arXiv about Chebyshev parametrizations in geometry, approximation, interpolation, and applications. Chebyshev shape parametrization refers to representations in which geometry, fields, or discretizations are expressed through Chebyshev polynomials of the first kind, or through the cosine-angle variables naturally associated with them. Across the literature, this includes bivariate approximants of the form $$f_{n,m}(x,y)=\sum_{k=0}^n\sum_{j=0}^m a_{k,j}T_k(x)T_j(y)$$, space curves of the form $x(t)=T_a(t)$, $y(t)=T_b(t)$, $z(t)=T_c(t+\phi)$, perturbed Chebyshev--Lobatto grids $$x_j=\cos\!\left(\frac{j\pi}{n}+\varepsilon_j\right)$$, nuclear surface profiles $\rho_s^2(u)/R_0^2=\sum_{n=0}^\infty a_nT_n(u)$, Chebyshev varieties given by $$\mathcal{T}_A(t)=\big(\mathcal{T}_{a_1}(t),\ldots,\mathcal{T}_{a_n}(t)\big)$$, and CNN layers augmented by Chebyshev expansions (Scheiber, 2015, Koseleff et al., 2010, Wu, 22 Jun 2026, Jyothish et al., 8 Aug 2025, Bel-Afia et al., 2024, Roy et al., 9 Apr 2025). The common technical motivation is the joint use of recurrence, orthogonality, cosine parametrization, sparse or compact coefficient structure, and approximation-theoretic stability.

1. Canonical formulations

Chebyshev polynomials of the first kind satisfy the recurrence

$$T_0(x)=1,\qquad T_1(x)=x,\qquad T_{n+1}(x)=2xT_n(x)-T_{n-1}(x),$$

and also the identity $T_n(\cos\theta)=\cos(n\theta)$ (Roy et al., 9 Apr 2025, Koseleff et al., 2010). These two descriptions underlie both polynomial and angular versions of Chebyshev parametrization.

Setting	Representative form	Role
Bivariate approximation	$$f_{n,m}(x,y)=\sum_{k=0}^{n}\sum_{j=0}^{m} a_{k,j}T_k(x)T_j(y)$$	Approximation, differentiation, integration
Space curves and knots	$x(t)=T_a(t)$0	Polynomial knot parametrization
Perturbed interpolation grids	$x(t)=T_a(t)$1	Chebyshev-like node geometry
Nuclear surfaces	$x(t)=T_a(t)$2	Deformation parametrization
Algebraic varieties	$x(t)=T_a(t)$3	Chebyshev analogue of toric geometry
CNN layers	$x(t)=T_a(t)$4	Spectral and shape-aware filtering

In these settings, the parametrization is not merely a basis change. In approximation theory it furnishes uniform convergence and efficient FFT-based coefficient recovery; in knot theory it yields explicit polynomial embeddings; in EC-space design it induces normalized B-bases and control-point formulas; in interpolation it preserves the angular clustering responsible for logarithmic Lebesgue growth; and in applied domains it provides deformation coordinates or spectral filters that are directly computable from a small number of coefficients (Scheiber, 2015, Róth, 2015, Wu, 22 Jun 2026, Jyothish et al., 8 Aug 2025).

2. Approximation-theoretic foundations

A central analytic formulation is the double Chebyshev series

$x(t)=T_a(t)$5

with coefficients

$x(t)=T_a(t)$6

The cited work computes these coefficients by sampling on the tensor-product Chebyshev grid $x(t)=T_a(t)$7, applying the 2D FFT, and extracting the lowest-frequency coefficients; the resulting approximation polynomial is

$x(t)=T_a(t)$8

which can be evaluated as $x(t)=T_a(t)$9 (Scheiber, 2015).

The same source proves two facts that are foundational for shape parametrization in function space. If $y(t)=T_b(t)$0 has continuous second partial derivatives, then $y(t)=T_b(t)$1, $y(t)=T_b(t)$2, and $y(t)=T_b(t)$3 decay at least as fast as $y(t)=T_b(t)$4 or $y(t)=T_b(t)$5, and the approximants $y(t)=T_b(t)$6 converge uniformly to $y(t)=T_b(t)$7 on $y(t)=T_b(t)$8 (Scheiber, 2015). The paper also discusses a two-dimensional Lagrange--Chebyshev interpolant, its aliasing relation to the approximation polynomial, and coefficient-based formulas for integration and partial differentiation.

The interpolation literature isolates a complementary issue: when a discretization retains the “Chebyshev shape.” For perturbed Chebyshev--Lobatto nodes

$y(t)=T_b(t)$9

the main deterministic worst-case estimate states that if $z(t)=T_c(t+\phi)$0, then

$z(t)=T_c(t+\phi)$1

Using the bound $z(t)=T_c(t+\phi)$2, the paper derives logarithmic growth of the perturbed Lebesgue constant under the condition $z(t)=T_c(t+\phi)$3 with sufficiently small $z(t)=T_c(t+\phi)$4 (Wu, 22 Jun 2026). It also identifies a transition regime consistent with $z(t)=T_c(t+\phi)$5, and proves a worst-case obstruction at the angular mesh scale: perturbations of order $z(t)=T_c(t+\phi)$6 cannot be allowed uniformly (Wu, 22 Jun 2026).

This establishes a domain-specific meaning of shape preservation. In the interpolation setting, the correct notion of remaining “Chebyshev-like” is preservation in the angular variable rather than direct perturbation in the physical coordinate $z(t)=T_c(t+\phi)$7; the proof exploits cosine parametrization and Bernstein’s inequality for trigonometric polynomials instead of a Markov inequality in the physical variable (Wu, 22 Jun 2026).

3. Curves, knots, and algebraic varieties

In low-dimensional geometry, Chebyshev shape parametrization appears in explicit polynomial curve models. A Chebyshev curve $z(t)=T_c(t+\phi)$8 in $z(t)=T_c(t+\phi)$9 is given by

$$x_j=\cos\!\left(\frac{j\pi}{n}+\varepsilon_j\right)$$0

where $$x_j=\cos\!\left(\frac{j\pi}{n}+\varepsilon_j\right)$$1 are integers and $$x_j=\cos\!\left(\frac{j\pi}{n}+\varepsilon_j\right)$$2 (Koseleff et al., 2010). When the curve has no double points, it defines a polynomial knot. The plane projection $$x_j=\cos\!\left(\frac{j\pi}{n}+\varepsilon_j\right)$$3 is the Chebyshev curve $$x_j=\cos\!\left(\frac{j\pi}{n}+\varepsilon_j\right)$$4, and its double points are given explicitly by

$$x_j=\cos\!\left(\frac{j\pi}{n}+\varepsilon_j\right)$$5

with $$x_j=\cos\!\left(\frac{j\pi}{n}+\varepsilon_j\right)$$6 and $$x_j=\cos\!\left(\frac{j\pi}{n}+\varepsilon_j\right)$$7 (Koseleff et al., 2010). The over/under structure of the knot diagram is determined by the sign of

$$x_j=\cos\!\left(\frac{j\pi}{n}+\varepsilon_j\right)$$8

where $$x_j=\cos\!\left(\frac{j\pi}{n}+\varepsilon_j\right)$$9 and $\rho_s^2(u)/R_0^2=\sum_{n=0}^\infty a_nT_n(u)$0 (Koseleff et al., 2010).

The singular values of the phase are encoded by a polynomial $\rho_s^2(u)/R_0^2=\sum_{n=0}^\infty a_nT_n(u)$1 whose real roots give all critical $\rho_s^2(u)/R_0^2=\sum_{n=0}^\infty a_nT_n(u)$2 where self-intersections occur (Koseleff et al., 2010). For fixed $\rho_s^2(u)/R_0^2=\sum_{n=0}^\infty a_nT_n(u)$3, the possible knot diagrams can be enumerated as $\rho_s^2(u)/R_0^2=\sum_{n=0}^\infty a_nT_n(u)$4 varies, and one cited algorithm lists all possible knots $\rho_s^2(u)/R_0^2=\sum_{n=0}^\infty a_nT_n(u)$5 in $\rho_s^2(u)/R_0^2=\sum_{n=0}^\infty a_nT_n(u)$6 bit operations, with $\rho_s^2(u)/R_0^2=\sum_{n=0}^\infty a_nT_n(u)$7 (Koseleff et al., 2015). The same line of work states that any knot type can be represented, up to isotopy, by a Chebyshev parametrization, and that Chebyshev knots generalize Lissajous knots (Koseleff et al., 2010).

At a more algebraic level, Chebyshev varieties are algebraic varieties parametrized by Chebyshev polynomials or their multivariate generalizations. Given multi-indices $\rho_s^2(u)/R_0^2=\sum_{n=0}^\infty a_nT_n(u)$8 and a matrix $\rho_s^2(u)/R_0^2=\sum_{n=0}^\infty a_nT_n(u)$9, the parametrization is

$$\mathcal{T}_A(t)=\big(\mathcal{T}_{a_1}(t),\ldots,\mathcal{T}_{a_n}(t)\big)$$0

and the Chebyshev variety is the Zariski closure $$\mathcal{T}_A(t)=\big(\mathcal{T}_{a_1}(t),\ldots,\mathcal{T}_{a_n}(t)\big)$$1 (Bel-Afia et al., 2024). For tensor-product Chebyshev parametrizations, if $$\mathcal{T}_A(t)=\big(\mathcal{T}_{a_1}(t),\ldots,\mathcal{T}_{a_n}(t)\big)$$2 has rank $$\mathcal{T}_A(t)=\big(\mathcal{T}_{a_1}(t),\ldots,\mathcal{T}_{a_n}(t)\big)$$3, then $$\mathcal{T}_A(t)=\big(\mathcal{T}_{a_1}(t),\ldots,\mathcal{T}_{a_n}(t)\big)$$4 generically has dimension $$\mathcal{T}_A(t)=\big(\mathcal{T}_{a_1}(t),\ldots,\mathcal{T}_{a_n}(t)\big)$$5; the paper derives degree bounds, characterizes singular loci, and gives defining equations (Bel-Afia et al., 2024). In the univariate case, the image curve $$\mathcal{T}_A(t)=\big(\mathcal{T}_{a_1}(t),\ldots,\mathcal{T}_{a_n}(t)\big)$$6 has defining equation

$$\mathcal{T}_A(t)=\big(\mathcal{T}_{a_1}(t),\ldots,\mathcal{T}_{a_n}(t)\big)$$7

and its singular locus is described explicitly as a finite set of nodes (Bel-Afia et al., 2024).

The same work explains that Chebyshev varieties play the role of toric varieties in sparse polynomial root finding, when monomials are replaced by Chebyshev polynomials (Bel-Afia et al., 2024). This suggests that “shape parametrization” here extends beyond geometric modeling into a basis-native algebraic geometry in which sparse systems are solved directly in Chebyshev coordinates.

4. Blossoming, B-bases, and control structures

A different branch of the literature studies shape parametrization through basis design in extended Chebyshev spaces. For an EC space $$\mathcal{T}_A(t)=\big(\mathcal{T}_{a_1}(t),\ldots,\mathcal{T}_{a_n}(t)\big)$$8 containing the constants, there exists a unique normalized B-basis $$\mathcal{T}_A(t)=\big(\mathcal{T}_{a_1}(t),\ldots,\mathcal{T}_{a_n}(t)\big)$$9 (Róth, 2015). These bases satisfy partition of unity, endpoint interpolation, and the vanishing and positivity properties of endpoint derivatives. Because normalized B-bases are totally positive, they ensure optimal shape preserving properties, convex hull, affine invariance, variation diminishing behavior, and recursive de Casteljau-type algorithms (Róth, 2015).

The same paper gives explicit formulas for the transformation matrix $$T_0(x)=1,\qquad T_1(x)=x,\qquad T_{n+1}(x)=2xT_n(x)-T_{n-1}(x),$$0 defined by

$$T_0(x)=1,\qquad T_1(x)=x,\qquad T_{n+1}(x)=2xT_n(x)-T_{n-1}(x),$$1

and derives ready-to-use control point configurations for exact representations of integral parametric curves, tensor-product surfaces, and their rational counterparts (Róth, 2015). For a curve $$T_0(x)=1,\qquad T_1(x)=x,\qquad T_{n+1}(x)=2xT_n(x)-T_{n-1}(x),$$2, the corresponding B-curve control points are

$$T_0(x)=1,\qquad T_1(x)=x,\qquad T_{n+1}(x)=2xT_n(x)-T_{n-1}(x),$$3

For rational curves and surfaces, the construction passes through a pre-image in one higher dimension and then projects to obtain control points and weights (Róth, 2015).

In Muntz spaces with integer exponents, the Chebyshev blossom becomes explicitly combinatorial. The blossom components are expressed as ratios of Schur functions indexed by partitions associated with Young diagrams, and the pseudo-affinity factor also has an explicit Schur-function formula (Ait-Haddou et al., 2011). The paper derives an explicit Chebyshev-Bernstein basis, an explicit dimension-elevation process, and tensor-product surface patches

$$T_0(x)=1,\qquad T_1(x)=x,\qquad T_{n+1}(x)=2xT_n(x)-T_{n-1}(x),$$4

so that free-form design becomes parametrizable by the shape and size of two Young diagrams (Ait-Haddou et al., 2011). In this framework, the Young diagram itself acts as a shape-controlling parameter.

This branch of the theory shifts the meaning of parametrization from “coefficient vector in a polynomial series” to “control structure in a shape-preserving basis.” A plausible implication is that Chebyshev shape parametrization comprises both spectral compression and geometric controllability, depending on whether the emphasis falls on orthogonal expansions or on normalized B-bases.

5. Nuclear shape parametrization

In nuclear modeling, the term is used in a direct geometric sense. The nuclear surface in cylindrical coordinates $$T_0(x)=1,\qquad T_1(x)=x,\qquad T_{n+1}(x)=2xT_n(x)-T_{n-1}(x),$$5 is described by the profile expansion

$$T_0(x)=1,\qquad T_1(x)=x,\qquad T_{n+1}(x)=2xT_n(x)-T_{n-1}(x),$$6

where $$T_0(x)=1,\qquad T_1(x)=x,\qquad T_{n+1}(x)=2xT_n(x)-T_{n-1}(x),$$7 maps $$T_0(x)=1,\qquad T_1(x)=x,\qquad T_{n+1}(x)=2xT_n(x)-T_{n-1}(x),$$8 to $$T_0(x)=1,\qquad T_1(x)=x,\qquad T_{n+1}(x)=2xT_n(x)-T_{n-1}(x),$$9, $T_n(\cos\theta)=\cos(n\theta)$0 is the equivalent volume-radius of a sphere, and $T_n(\cos\theta)=\cos(n\theta)$1 are Chebyshev deformation parameters (Jyothish et al., 8 Aug 2025). Boundary conditions at $T_n(\cos\theta)=\cos(n\theta)$2 imply

$T_n(\cos\theta)=\cos(n\theta)$3

hence

$T_n(\cos\theta)=\cos(n\theta)$4

The paper interprets $T_n(\cos\theta)=\cos(n\theta)$5 as elongation, $T_n(\cos\theta)=\cos(n\theta)$6 as asymmetry, and $T_n(\cos\theta)=\cos(n\theta)$7 as neck (Jyothish et al., 8 Aug 2025).

Volume conservation is enforced analytically through

$T_n(\cos\theta)=\cos(n\theta)$8

where $T_n(\cos\theta)=\cos(n\theta)$9 defines the elongation parameter $$f_{n,m}(x,y)=\sum_{k=0}^{n}\sum_{j=0}^{m} a_{k,j}T_k(x)T_j(y)$$0; for spheroidal shapes with only $$f_{n,m}(x,y)=\sum_{k=0}^{n}\sum_{j=0}^{m} a_{k,j}T_k(x)T_j(y)$$1 nonzero, $$f_{n,m}(x,y)=\sum_{k=0}^{n}\sum_{j=0}^{m} a_{k,j}T_k(x)T_j(y)$$2 (Jyothish et al., 8 Aug 2025). For asymmetric shapes, the center-of-mass shift is

$$f_{n,m}(x,y)=\sum_{k=0}^{n}\sum_{j=0}^{m} a_{k,j}T_k(x)T_j(y)$$3

A non-axiality parameter $$f_{n,m}(x,y)=\sum_{k=0}^{n}\sum_{j=0}^{m} a_{k,j}T_k(x)T_j(y)$$4 extends the model to triaxial shapes through

$$f_{n,m}(x,y)=\sum_{k=0}^{n}\sum_{j=0}^{m} a_{k,j}T_k(x)T_j(y)$$5

The same work gives transformation relations to Bohr--Mottelson, Funny-Hills, TKS, and Fourier shape parametrizations (Jyothish et al., 8 Aug 2025).

The parametrization is then used in both macroscopic and microscopic calculations. In the macroscopic approach, the potential energy surface is computed using the Lublin-Strasbourg Drop model, with deformation-dependent coefficients evaluated on the Chebyshev-defined surface; in the microscopic approach, single-particle levels are obtained by diagonalizing the Yukawa-folded mean-field Hamiltonian in a deformed harmonic oscillator basis (Jyothish et al., 8 Aug 2025). The paper states that only a few Chebyshev coefficients $$f_{n,m}(x,y)=\sum_{k=0}^{n}\sum_{j=0}^{m} a_{k,j}T_k(x)T_j(y)$$6 suffice to accurately model spherical, prolate, oblate, asymmetric, necked, and dumbbell-like shapes, and that mapped single-particle schemes coincide with those from other parametrizations when the geometric shapes are identical (Jyothish et al., 8 Aug 2025).

6. Terrain analysis and neural architectures

Chebyshev parametrization also functions as a global analytic surrogate in data-rich settings. In digital terrain modeling, a topographic surface $$f_{n,m}(x,y)=\sum_{k=0}^{n}\sum_{j=0}^{m} a_{k,j}T_k(x)T_j(y)$$7 is approximated on a rectangular domain by a two-dimensional Chebyshev series

$$f_{n,m}(x,y)=\sum_{k=0}^{n}\sum_{j=0}^{m} a_{k,j}T_k(x)T_j(y)$$8

with coordinates linearly mapped to $$f_{n,m}(x,y)=\sum_{k=0}^{n}\sum_{j=0}^{m} a_{k,j}T_k(x)T_j(y)$$9 (Florinsky et al., 2015). The coefficients are computed by projection onto the Chebyshev basis, typically at the nodes

$x(t)=T_a(t)$00

and the expansion is carried out sequentially in two stages, by columns and then rows (Florinsky et al., 2015). To suppress oscillatory artifacts associated with the Gibbs phenomenon, the method applies Fejér summation after the high-order orthogonal expansion (Florinsky et al., 2015).

The terrain paper tests the method on a DEM of the Northern Andes including 230,880 points, with elevation matrix $x(t)=T_a(t)$01, and reconstructs the DEM using 480, 240, 120, 60, and 30 expansion coefficients (Florinsky et al., 2015). First and second partial derivatives are calculated analytically from the reconstructed DEMs, and models of horizontal curvature $x(t)=T_a(t)$02 are computed from the derivatives. The reported effect of truncation is strong generalization and denoising: high-frequency noise and cartographic artifacts disappear first in the curvature maps and then in the elevation maps as the number of coefficients decreases (Florinsky et al., 2015).

In medical image analysis, the proposed Chebyshev-CNN replaces or augments conventional convolutional filters by a polynomial expansion

$x(t)=T_a(t)$03

with a spectral interpretation

$x(t)=T_a(t)$04

where the eigenvalue matrix is rescaled to $x(t)=T_a(t)$05 (Roy et al., 9 Apr 2025). The implementation uses kernels such as $x(t)=T_a(t)$06 approximated with several terms, for example $x(t)=T_a(t)$07, followed by batch normalization, ReLU, pooling, dense layers, and a softmax classifier (Roy et al., 9 Apr 2025). The paper attributes the gain to orthogonality, recursive computation, high-frequency feature extraction, and the approximation of complex nonlinear functions with greater fidelity.

For pulmonary nodule classification on LUNA16 and LIDC-IDRI, the reported average accuracy, sensitivity, and specificity are each $x(t)=T_a(t)$08; the ablation replacing Chebyshev layers with standard convolutions degrades accuracy to $x(t)=T_a(t)$09 (Roy et al., 9 Apr 2025). The comparison table in the same source reports $x(t)=T_a(t)$10 accuracy for a 3D CNN, $x(t)=T_a(t)$11 for a 2D multi-view CNN, and $x(t)=T_a(t)$12 for a 3D Faster R-CNN, while the Chebyshev-CNN reaches $x(t)=T_a(t)$13 (Roy et al., 9 Apr 2025). The paper further states that the network performs well across benign and malignant classes, and that interpretability is potentially improved via polynomial decomposition.

7. Stability, misconceptions, and conceptual boundaries

A recurrent misconception is that any use of Chebyshev polynomials counts as a stable Chebyshev parametrization. The interpolation results show that this is false in a precise sense: arbitrary perturbations at the angular mesh scale $x(t)=T_a(t)$14 can force the Lebesgue constant to grow superlogarithmically, even though much smaller angular perturbations preserve logarithmic growth (Wu, 22 Jun 2026). In that setting, the preserved object is the angular geometry of the Chebyshev--Lobatto grid, not merely the fact that the nodes arose from a cosine formula.

A second misconception is that a Chebyshev curve is automatically a knot. In the knot literature, the parametrized curve defines a polynomial knot only when it is nonsingular, or equivalently when there are no real double points; the critical values of $x(t)=T_a(t)$15 form a finite set, and the knot type is constant only on intervals between those critical values (Koseleff et al., 2010). Thus the parametrization provides an explicit family of candidates, but knot realization depends on injectivity.

A third boundary concerns interpretability. In free-form design and EC spaces, control points, normalized B-bases, and Young-diagram parameters give a direct geometric semantics (Róth, 2015, Ait-Haddou et al., 2011). In medical imaging, by contrast, the paper claims only that interpretability is potentially improved via polynomial decomposition rather than that it is guaranteed (Roy et al., 9 Apr 2025). The distinction matters because “shape parametrization” can describe either an explicit geometric model or a learned spectral parameterization embedded inside a larger network.

Across these literatures, the unifying theme is methodological rather than disciplinary. Chebyshev shape parametrization can mean a compact orthogonal expansion, a cosine-based node geometry, an explicit polynomial embedding, a control-point representation in an EC space, or a constrained deformation model. This suggests that the term is best understood as a family of parametrization strategies organized around the approximation, algebraic, and geometric properties of Chebyshev structures, rather than as a single canonical formalism (Scheiber, 2015, Bel-Afia et al., 2024, Jyothish et al., 8 Aug 2025).