Legendre–Sobolev Graded Basis

• Legendre–Sobolev graded basis is a family of polynomial bases constructed via Gram–Schmidt orthogonalization using a Sobolev inner product that incorporates both function and derivative information.
• It is applied in one-dimensional handwriting recognition to produce compact, stable curve descriptors and in multi-dimensional settings via block-organized expansions for Sobolev spaces.
• The method improves numerical stability by reducing coefficient interference, with empirical findings suggesting Chebyshev–Sobolev variants for optimal accuracy and controlled norm growth.

Searching arXiv for the specified papers and closely related context. Legendre–Sobolev graded basis denotes a family of polynomial bases constructed by Gram–Schmidt orthogonalization of monomials under a Sobolev inner product, typically of the form

$$\langle f, g \rangle_{S} = \int_a^b f(s) g(s) \, ds + \mu \int_a^b f'(s) g'(s) \, ds,$$

with a real parameter $\mu$ controlling the contribution of derivative information. In the mathematical handwriting setting, such bases are used to represent parameterized plane curves of digital ink after arc-length normalization, yielding compact coefficient vectors for the coordinate functions $x(s)$ and $y(s)$ (Corless et al., 13 Sep 2025). In a broader Sobolev-space setting, closely related graded constructions organize Legendre polynomial blocks by derivative order, boundary trace, and codimension, with reconstruction operators that map derivative-and-trace data back to the original function (Jagt et al., 2023). The term therefore covers two closely connected ideas: a one-dimensional graded orthogonal basis adapted to Sobolev inner products for curve representation, and a block-structured Legendre-based organization of Sobolev data in higher-dimensional constructive representations.

1. Formal definition and basic representation

In the handwriting-recognition formulation, each stroke is modeled as a parameterized plane curve with respect to normalized arc length $s$, satisfying

$ds^2 = dx^2 + dy^2,$

with

$x(s), y(s), \quad s \in [-1,1].$

Each coordinate is approximated by a finite expansion

$$x(s) \approx \sum_{i=0}^{d} x_i B_i(s), \qquad y(s) \approx \sum_{i=0}^{d} y_i B_i(s),$$

where $\{B_i(s)\}$ is a graded basis of orthogonal polynomials obtained from the monomials $\{1,s,s^2,\dots,s^d\}$ by Gram–Schmidt orthogonalization under a chosen inner product (Corless et al., 13 Sep 2025).

For the standard Legendre construction, the inner product is

$\mu$0

For the Chebyshev variant, the weighted inner product is

$\mu$1

For Legendre–Sobolev and Chebyshev–Sobolev bases, the Sobolev inner product is

$\mu$2

The experiments reported for handwriting recognition use $\mu$3 (Corless et al., 13 Sep 2025).

The basis is orthogonal by construction. The paper states the orthogonality relation in the Sobolev case as

$\mu$4

with analogous relations for the standard and Chebyshev-weighted inner products. It does not provide explicit recurrence relations or closed-form expressions for the Legendre–Sobolev basis functions; the construction is numerical, via Gram–Schmidt (Corless et al., 13 Sep 2025).

A distinct but related formalization appears in the constructive Sobolev representation on an $\mu$5-dimensional hyperrectangle

$\mu$6

where functions in

$\mu$7

are represented in terms of their highest-order mixed derivative and lower-order boundary traces (Jagt et al., 2023). In that setting, a Legendre–Sobolev graded basis is not a single one-dimensional polynomial family, but a union of Legendre tensor-product blocks indexed by derivative order and boundary location.

2. Construction from monomials and Sobolev data

The handwriting paper constructs a graded basis directly from monomials by Gram–Schmidt with respect to the chosen functional inner product. The crucial point is that the basis is adapted to the geometry of the representation rather than imposed through a fixed analytic formula. Because coefficient computation is performed relative to an orthogonal basis, the paper states that “Because of orthogonality, there is no interfence in the calculation of coefficients” (Corless et al., 13 Sep 2025).

The constructive Sobolev-space framework makes the grading more explicit. For a multi-index $\mu$8, the data vector is

$\mu$9

where

$x(s)$0

The function is reconstructed by

$x(s)$1

with $x(s)$2 defined as a tensor product of one-dimensional multiplication and integration operators involving the kernels

$x(s)$3

This yields a bijection

$x(s)$4

(Jagt et al., 2023).

In that framework, the “graded” structure has a precise combinatorial meaning. The basis is organized blockwise by $x(s)$5, with each block corresponding to a derivative-trace component $x(s)$6 living on a lower-dimensional subdomain. The paper defines a primary grading by codimension,

$x(s)$7

and a secondary grading by polynomial degree within each block (Jagt et al., 2023). This suggests a broader interpretation of Legendre–Sobolev graded basis: grading can refer either to polynomial degree alone, as in the one-dimensional handwriting model, or to a hierarchical arrangement of derivative and boundary components in Sobolev-space representations.

3. Legendre structure, tensor products, and block organization

The constructive Sobolev paper gives explicit Legendre formulas. On $x(s)$8, the classical Legendre polynomials $x(s)$9 satisfy

$y(s)$0

and

$y(s)$1

The normalized basis functions are

$y(s)$2

so that

$y(s)$3

The corresponding $y(s)$4-dimensional tensor-product basis is

$y(s)$5

(Jagt et al., 2023).

For a block indexed by $y(s)$6, the associated trace data

$y(s)$7

depends only on the free coordinates in

$y(s)$8

That block is expanded in a lower-dimensional Legendre tensor basis: $y(s)$9 The graded basis is then the union of all such blocks. Because the underlying inner product splits over the family of subdomains $s$0, the blocks are mutually orthogonal in the paper’s discrete-continuous Sobolev norm (Jagt et al., 2023).

This block-diagonal decomposition differs from the handwriting paper’s one-dimensional Gram–Schmidt basis, but the two viewpoints are compatible. Both build orthogonal polynomial representations under Sobolev-type norms, and both use derivative information to stabilize approximation. A plausible implication is that the handwriting Legendre–Sobolev basis may be viewed as the one-dimensional analogue of the more general derivative-and-trace decomposition formalized in higher dimensions.

4. Coefficient computation and reconstruction

In the handwriting-recognition setting, the parameterized curve is projected onto the chosen basis by computing inner products and using numerical integration to calculate coefficients. The paper does not explicitly write

$s$1

but states that such projection is implied by orthogonality (Corless et al., 13 Sep 2025). The representation remains

$s$2

In the constructive Sobolev framework, coefficient computation is stated explicitly for both interior and boundary blocks. For the interior highest-order block $s$3,

$s$4

For a boundary block $s$5,

$s$6

with

$s$7

The reconstructed approximation is

$s$8

(Jagt et al., 2023).

The distinction between the two papers is methodological rather than conceptual. The handwriting paper emphasizes coefficient vectors as features for classification, while the constructive Sobolev paper emphasizes invertible representation of Sobolev functions through derivative and boundary blocks. In both cases, the basis coefficients are meaningful because orthogonality suppresses interference among components, and reconstruction is expressed as a superposition of basis functions or reconstructed blocks.

5. Conditioning, stability, and convergence properties

The handwriting paper motivates basis selection through conditioning of polynomial evaluation. It notes that different bases can have dramatically different condition numbers and that change-of-basis matrices are often ill-conditioned, with condition numbers that can be exponential in degree. Using the standard definition

$s$9

the paper argues qualitatively that ill-conditioned bases amplify both computational rounding errors and data errors (Corless et al., 13 Sep 2025).

For Sobolev-based representations, the paper defines the norm

$ds^2 = dx^2 + dy^2,$0

and proves two bounds linking coefficient perturbations to differences in function value and derivative. The first is

$ds^2 = dx^2 + dy^2,$1

where $ds^2 = dx^2 + dy^2,$2 is the differentiation matrix and $ds^2 = dx^2 + dy^2,$3 is the step size. The second is

$ds^2 = dx^2 + dy^2,$4

The derivative coefficients satisfy

$ds^2 = dx^2 + dy^2,$5

These results justify the interpretation that small coefficient changes imply bounded changes in both function shape and derivative, which is directly relevant when symbol comparison depends on geometry as well as local variation (Corless et al., 13 Sep 2025).

The constructive Sobolev paper gives a complementary perspective on stability. Since the inner product splits over blocks and each block uses an orthonormal Legendre basis, the global Gram matrix is block-diagonal with orthonormal blocks. The paper states that this organization improves convergence and stability, particularly because refinement can be targeted anisotropically by codimension and degree (Jagt et al., 2023).

It also records a standard $ds^2 = dx^2 + dy^2,$6 projection estimate: $ds^2 = dx^2 + dy^2,$7 but notes that the Sobolev norm of the error of direct $ds^2 = dx^2 + dy^2,$8 projection can grow: $ds^2 = dx^2 + dy^2,$9 By contrast, if each block is projected separately and reconstruction is performed with $x(s), y(s), \quad s \in [-1,1].$0, the approximation is optimal in the discrete-continuous Sobolev norm: $x(s), y(s), \quad s \in [-1,1].$1 This sharpens the conceptual meaning of a Legendre–Sobolev graded basis: it is not merely an orthogonal polynomial system, but an organizational scheme that aligns approximation with Sobolev regularity and trace structure (Jagt et al., 2023).

6. Use in mathematical handwriting recognition

The most concrete application in the supplied material is mathematical handwriting recognition. The handwriting paper models digital ink trajectories as arc-length-parameterized curves, projects them onto Legendre, Chebyshev, Legendre–Sobolev, and Chebyshev–Sobolev bases, and uses the resulting coefficient vectors as features for an SVM pipeline (Corless et al., 13 Sep 2025).

For the UNIPEN digits experiment, trajectories are uniformly resampled to exactly 8 points. Classification is performed with a one-vs-one strategy for 10 classes, yielding 45 binary SVMs, with an 80%/20% training/testing split and 100 random splits. Accuracy is aggregated through min/max/mean across runs for each degree and basis (Corless et al., 13 Sep 2025).

The empirical findings reported for basis trade-offs are summarized below.

Basis	Accuracy behavior	Computational/stability characterization
Legendre	Plateaus closer to ~96%	Faster to compute, but coefficients grow rapidly with degree
Chebyshev	Close to Legendre–Sobolev at higher degrees	Slightly slower due to weighted inner product
Legendre–Sobolev	Peaks close to ~97%	Improved stability and accuracy, with modestly higher cost
Chebyshev–Sobolev	Around 97.5–98% at degree ~12	Most controlled coefficient growth, highest computational cost

Accuracy increases with degree up to approximately 10–12 and then stabilizes. Legendre and Chebyshev show steeper growth of cumulative coefficient norms with degree, with Legendre consistently the largest. The Sobolev variants maintain more controlled cumulative norm growth, and Chebyshev–Sobolev is reported as most stable in terms of cumulative norms at high degrees, followed by Legendre–Sobolev. Computation time grows almost linearly with degree; Sobolev bases are more expensive because of derivative terms, and Chebyshev–Sobolev is the most expensive because it combines derivative terms with the Chebyshev weight (Corless et al., 13 Sep 2025).

The paper’s practical recommendation is therefore specific: for highest recognition accuracy and stability, use Chebyshev–Sobolev around degree 10–12; for lower computational cost with acceptable accuracy, use Legendre or Chebyshev at moderate degree; and avoid change-of-basis later in the pipeline because of numerical instability (Corless et al., 13 Sep 2025). In this application, Legendre–Sobolev graded bases occupy the middle ground between plain Legendre efficiency and Chebyshev–Sobolev robustness.

7. Interpretation, limitations, and related misconceptions

A common misconception is to treat “Legendre–Sobolev graded basis” as a closed-form family analogous to classical Legendre polynomials. The handwriting paper does not provide explicit recurrence relations or closed-form expressions for the Legendre–Sobolev or Chebyshev–Sobolev basis functions; practical construction is numerical, via Gram–Schmidt under the chosen Sobolev inner product (Corless et al., 13 Sep 2025). By contrast, the constructive Sobolev paper gives explicit classical Legendre recurrences, but its “Legendre–Sobolev graded basis” is a blockwise assembly over derivative and boundary components rather than a single modified recurrence family (Jagt et al., 2023).

Another misconception is that Sobolev augmentation automatically solves all conditioning problems. The handwriting paper explicitly warns that change-of-basis matrices are often ill-conditioned, with potentially exponential condition numbers in degree, and that sensitivity to data errors remains inherent to ill-conditioned bases. It notes that bidiagonal factorization of totally nonnegative matrices can mitigate numerical instability for certain least-squares problems, but does not claim that this resolves the broader sensitivity issue (Corless et al., 13 Sep 2025).

The two papers also leave different open directions. The handwriting paper states that it does not provide explicit quantitative bounds for the evaluation operators of Legendre, Legendre–Sobolev, Chebyshev, and Chebyshev–Sobolev bases, and does not derive detailed geometric error bounds for plane-curve proximity under coefficient perturbation. It identifies extension of such geometric error analysis as an open direction. It also notes that numerical mitigation strategies such as bidiagonal factorization are mentioned but not integrated into the handwriting pipeline (Corless et al., 13 Sep 2025).

The constructive Sobolev paper, in turn, shows how approximation in Sobolev space can be reduced to approximation in a product of $x(s), y(s), \quad s \in [-1,1].$2 spaces through the bijection $x(s), y(s), \quad s \in [-1,1].$3 and the reconstruction operator $x(s), y(s), \quad s \in [-1,1].$4. This suggests a wider significance for graded Legendre–Sobolev constructions beyond handwriting: they provide a principled way to encode both interior derivative information and boundary data while preserving invertibility and blockwise orthogonality (Jagt et al., 2023). A plausible implication is that the one-dimensional handwriting representation can be interpreted as a specialized instance of the broader strategy of matching basis design to the Sobolev structure of the target object.

Taken together, these works characterize the Legendre–Sobolev graded basis as an orthogonality-based representation framework in which derivative information is incorporated into basis construction or into the organization of coefficient blocks. In one dimension, it yields compact and comparatively stable curve descriptors for digital ink; in higher dimensions, it supports constructive Sobolev representations with blockwise Legendre expansions, explicit reconstruction operators, and norm-optimal approximation in a discrete-continuous Sobolev sense (Corless et al., 13 Sep 2025, Jagt et al., 2023).