Boundary-Adapted Chebyshev Basis Functions

• Boundary-adapted Chebyshev basis functions are modified trial spaces that inherently enforce prescribed endpoint conditions in spectral approximations.
• They use techniques like algebraic recombination, multiplicative edge factors, and weighted constructions to tailor the basis for various boundary and operator challenges.
• These adaptations enhance computational efficiency, improve matrix conditioning, and ensure uniform error distribution in applications like beam models and degenerate elliptic problems.

Boundary-adapted Chebyshev basis functions are modified Chebyshev trial spaces constructed so that an approximation satisfies prescribed endpoint constraints by design rather than by post hoc correction. In the cited literature, this adaptation takes several forms: algebraic recombination such as $T_{n+2}-T_n$, multiplicative edge factors such as $(1-x^2)T_n(x)$ or $(1-\xi^2)^2T_n(\xi)$, weighted constructions matched to degenerate operators, mapped rational bases for infinite intervals, and smooth partition-of-unity blends of local Chebyshev interpolants. Their common purpose is to encode boundary behavior into the approximation space itself, although some closely related Chebyshev methods instead retain the standard basis and impose boundary conditions by collocation constraints or row replacement (Jalili et al., 15 Sep 2025, Yuan et al., 29 May 2026, Zhang et al., 2021, Miquel et al., 2017).

1. Algebraic foundation and endpoint structure

The standard Chebyshev polynomials of the first kind are defined on $[-1,1]$ by

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

with the trigonometric representation

$T_n(x)=\cos(n\arccos x).$

Their derivatives are expressed through Chebyshev polynomials of the second kind by

$T_n'(x)=nU_{n-1}(x).$

These identities are central because endpoint values are explicit: $T_n(1)=1,\qquad T_n(-1)=(-1)^n.$ On a physical interval $[a,b]$ or $[0,L]$, the standard affine maps

$(1-x^2)T_n(x)$0

transfer the problem to $(1-x^2)T_n(x)$1, with chain-rule factors converting derivatives between physical and reference coordinates (Jalili et al., 15 Sep 2025, Bhowmik, 2014).

Boundary adaptation exploits this endpoint algebra. If a basis element is formed so that its value, slope, or another required trace vanishes at $(1-x^2)T_n(x)$2, then any finite expansion in that basis inherits the same property. In this sense, the endpoint identities of $(1-x^2)T_n(x)$3 are not merely convenient formulas; they are the mechanism by which homogeneous boundary conditions are encoded exactly at the level of the approximation space. For operators with weighted or singular endpoint structure, the same principle is combined with a natural weight, such as $(1-x^2)T_n(x)$4 for first-kind Chebyshev orthogonality or $(1-x^2)T_n(x)$5 for certain endpoint-degenerate elliptic problems (Yuan et al., 29 May 2026).

2. Canonical constructions on bounded intervals

Two of the most explicit homogeneous Dirichlet constructions are the Chebyshev difference basis

$(1-x^2)T_n(x)$6

and the quadratic-factor basis

$(1-x^2)T_n(x)$7

Both satisfy $(1-x^2)T_n(x)$8 and $(1-x^2)T_n(x)$9 exactly, and both are treated in detail for homogeneous Dirichlet problems and for asymptotic coefficient analysis under weak endpoint singularities (Zhang et al., 2021). A nearly identical difference construction appears in the degenerate elliptic setting, where

$(1-\xi^2)^2T_n(\xi)$0

defines a trial space $(1-\xi^2)^2T_n(\xi)$1 in which every $(1-\xi^2)^2T_n(\xi)$2 satisfies $(1-\xi^2)^2T_n(\xi)$3 by construction (Yuan et al., 29 May 2026).

For Euler–Bernoulli beam models, the adaptation may target higher-order endpoint traces. In the nonlinear CNTRC beam framework, the clamped–clamped basis is

$(1-\xi^2)^2T_n(\xi)$4

At $(1-\xi^2)^2T_n(\xi)$5, both the function and its first derivative vanish: $(1-\xi^2)^2T_n(\xi)$6 so any Ritz expansion $(1-\xi^2)^2T_n(\xi)$7 satisfies the clamped conditions pointwise. For simply supported ends, the basis

$(1-\xi^2)^2T_n(\xi)$8

enforces $(1-\xi^2)^2T_n(\xi)$9 exactly, while the bending-moment condition is treated differently, as discussed below (Jalili et al., 15 Sep 2025).

These constructions illustrate two distinct design rules. The difference basis uses cancellation between neighboring Chebyshev modes, whereas edge-factor bases multiply by a polynomial vanishing at the boundary. Both are algebraically simple, but they lead to different coefficient asymptotics, endpoint derivative behavior, and matrix structures (Zhang et al., 2021).

3. Strong enforcement, weak enforcement, and constraint-based alternatives

Boundary-adapted Chebyshev basis functions are most naturally associated with strong enforcement: the approximation space is chosen so that the essential boundary conditions hold for every admissible expansion coefficient vector. This is the case for homogeneous Dirichlet constraints under $[-1,1]$0, $[-1,1]$1, or $[-1,1]$2, and for clamped beam constraints under $[-1,1]$3 (Zhang et al., 2021, Yuan et al., 29 May 2026, Jalili et al., 15 Sep 2025).

The simply supported Euler–Bernoulli beam example shows that boundary adaptation need not imply strong enforcement of every endpoint condition. In the CNTRC beam formulation, $[-1,1]$4 gives exact endpoint displacement cancellation, but in general

$[-1,1]$5

which is not identically zero for arbitrary $[-1,1]$6. The paper therefore enforces $[-1,1]$7 in the weak, natural sense through the Ritz/Galerkin projection. This is described as standard within the Euler–Bernoulli energy formulation, and the resulting discrete solution satisfies the moment-free condition increasingly accurately as $[-1,1]$8 grows. The same source explicitly remarks that strong enforcement of $[-1,1]$9 would require higher-order edge factors and/or boundary lifting, but that this was not the adopted construction (Jalili et al., 15 Sep 2025).

A common misconception is that Chebyshev discretizations must use boundary-adapted basis functions whenever nontrivial boundary data are present. The literature includes direct counterexamples. One collocation method for linear $$T_0(x)=1,\qquad T_1(x)=x,\qquad T_{n+1}(x)=2xT_n(x)-T_{n-1}(x),$$0-order BVPs converts the equation to a first-order system, expands each component in an unmodified Chebyshev basis, and imposes boundary data by tau-style row replacement: $$T_0(x)=1,\qquad T_1(x)=x,\qquad T_{n+1}(x)=2xT_n(x)-T_{n-1}(x),$$1 with analogous row combinations for Robin conditions (Bhowmik, 2014). Likewise, an FFT-based Chebyshev collocation solver for diffusion problems with general Dirichlet and Neumann conditions enforces BCs by constraint elimination in collocation space rather than by modifying the polynomial basis (Quecedo et al., 19 Jun 2026). These methods are not boundary-adapted in the basis-function sense, but they are relevant because they define the principal alternative strategy: boundary conditions can be encoded either in the approximation space or in the discrete constraint system.

4. Approximation theory, coefficient asymptotics, and endpoint behavior

For analytic functions, Chebyshev approximations exhibit spectral convergence. In one collocation framework this is stated as truncation error $$T_0(x)=1,\qquad T_1(x)=x,\qquad T_{n+1}(x)=2xT_n(x)-T_{n-1}(x),$$2 for some $$T_0(x)=1,\qquad T_1(x)=x,\qquad T_{n+1}(x)=2xT_n(x)-T_{n-1}(x),$$3, while the weighted modal scheme for a degenerate elliptic problem gives

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

for $$T_0(x)=1,\qquad T_1(x)=x,\qquad T_{n+1}(x)=2xT_n(x)-T_{n-1}(x),$$5 when the solution is analytic (Bhowmik, 2014, Yuan et al., 29 May 2026). In the nonlinear CNTRC beam application, the Chebyshev–Ritz reduced model exhibits exponential convergence with the representative fit

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

giving $$T_0(x)=1,\qquad T_1(x)=x,\qquad T_{n+1}(x)=2xT_n(x)-T_{n-1}(x),$$7 and $$T_0(x)=1,\qquad T_1(x)=x,\qquad T_{n+1}(x)=2xT_n(x)-T_{n-1}(x),$$8 for $$T_0(x)=1,\qquad T_1(x)=x,\qquad T_{n+1}(x)=2xT_n(x)-T_{n-1}(x),$$9; the abstract summarizes this as fundamental-frequency error below $T_n(x)=\cos(n\arccos x).$0 for $T_n(x)=\cos(n\arccos x).$1 (Jalili et al., 15 Sep 2025).

For weak endpoint singularities, the basis choice alters coefficient decay in a precise way. If the Chebyshev coefficients of the unconstrained expansion satisfy

$T_n(x)=\cos(n\arccos x).$2

for

$T_n(x)=\cos(n\arccos x).$3

with $T_n(x)=\cos(n\arccos x).$4, then the difference-basis and quadratic-factor coefficients satisfy

$T_n(x)=\cos(n\arccos x).$5

Thus the difference coefficients decay more slowly than $T_n(x)=\cos(n\arccos x).$6 by one power of $T_n(x)=\cos(n\arccos x).$7, and the quadratic-factor coefficients more slowly still by two powers relative to $T_n(x)=\cos(n\arccos x).$8 (Zhang et al., 2021).

The same source distinguishes error distribution from coefficient decay. For truncated unconstrained Chebyshev expansions, the interior pointwise error is $T_n(x)=\cos(n\arccos x).$9, but in narrow boundary layers near the endpoints it deteriorates to $T_n'(x)=nU_{n-1}(x).$0. By contrast, the error in the difference and quadratic-factor bases is described as nearly uniform oscillation over the entire interval, rather than concentration near the endpoints. Under interpolation or least-squares constructions that enforce the same endpoint constraints, the difference and quadratic-factor bases produce identical approximants of the form

$T_n'(x)=nU_{n-1}(x).$1

and therefore identical error norms, even though their coefficient sequences behave differently (Zhang et al., 2021).

Endpoint derivative scaling also depends strongly on the adapted basis. Because $T_n'(x)=nU_{n-1}(x).$2, both the standard Chebyshev basis and the quadratic-factor basis lead to endpoint derivatives scaling like $T_n'(x)=nU_{n-1}(x).$3 in truncated sums. For the difference basis,

$T_n'(x)=nU_{n-1}(x).$4

and the large endpoint contributions cancel to leave $T_n'(x)=nU_{n-1}(x).$5 behavior, so endpoint derivative magnitudes scale like $T_n'(x)=nU_{n-1}(x).$6. This is one reason the difference basis is highlighted as advantageous for conditioning and derivative-sensitive discretizations (Zhang et al., 2021).

5. Matrix structure, quadrature, and computational consequences

Boundary adaptation is not only a device for satisfying endpoint conditions; it also shapes the discrete operator. In the CNTRC beam Ritz formulation, after inserting

$T_n'(x)=nU_{n-1}(x).$7

the stiffness and mass matrices become

$T_n'(x)=nU_{n-1}(x).$8

The integrals are evaluated by Gauss–Chebyshev quadrature, and the basis is orthonormalized with respect to the mass inner product

$T_n'(x)=nU_{n-1}(x).$9

to improve conditioning. The same formulation assembles the von Kármán nonlinear term through derivative-based quantities

$T_n(1)=1,\qquad T_n(-1)=(-1)^n.$0

The paper attributes stabilization and acceleration of this evaluation to the smooth, well-resolved derivatives furnished by the boundary-adapted basis (Jalili et al., 15 Sep 2025).

In the endpoint-degenerate elliptic problem with $T_n(1)=1,\qquad T_n(-1)=(-1)^n.$1, the weighted Sturm–Liouville identity

$T_n(1)=1,\qquad T_n(-1)=(-1)^n.$2

and the adapted basis $T_n(1)=1,\qquad T_n(-1)=(-1)^n.$3 yield a stiffness matrix with entries

$T_n(1)=1,\qquad T_n(-1)=(-1)^n.$4

where $T_n(1)=1,\qquad T_n(-1)=(-1)^n.$5 and $T_n(1)=1,\qquad T_n(-1)=(-1)^n.$6 for $T_n(1)=1,\qquad T_n(-1)=(-1)^n.$7. This is a banded matrix with bandwidth $T_n(1)=1,\qquad T_n(-1)=(-1)^n.$8 in the modal index, and even and odd modes decouple into two independent tridiagonal blocks. The implementation guidance given there states $T_n(1)=1,\qquad T_n(-1)=(-1)^n.$9 memory, $[a,b]$0 assembly, and $[a,b]$1 factorization for each tridiagonal block (Yuan et al., 29 May 2026).

The computational impact can be substantial. For the CNTRC beam model, the spectral reduced system provides roughly $[a,b]$2 speedup and $[a,b]$3 lower memory than high-fidelity FEM at comparable accuracy, with exact satisfaction of essential boundary conditions for the clamped case reported to remove spurious boundary-layer artifacts, improve stiffness-matrix conditioning, and accelerate convergence (Jalili et al., 15 Sep 2025). In the partition-of-unity setting, global differentiation matrices assembled from local Chebyshev operators and smooth weights are reported to be block-sparse, with about $[a,b]$4 sparsity in a representative case (Aiton et al., 2017).

6. Broader variants: weighted, partitioned, and infinite-domain adaptations

Boundary adaptation extends beyond polynomial edge factors on a single finite interval. One variant is weight matching for degenerate operators. For coefficients behaving like $[a,b]$5, the degenerate elliptic analysis identifies the natural weight

$[a,b]$6

and notes that the difference basis $[a,b]$7 still enforces $[a,b]$8, while orthogonality and sparsity arguments adapt after replacing $[a,b]$9 by $[0,L]$0 (Yuan et al., 29 May 2026).

A second variant uses overlapping subdomains and smooth partitions of unity. On a cover $[0,L]$1, local Chebyshev interpolants $[0,L]$2 are blended by compactly supported $[0,L]$3 weights $[0,L]$4 satisfying $[0,L]$5: $[0,L]$6 For the two-patch case, the weights are built from the bump

$[0,L]$7

scaled to the overlapping intervals. The resulting global approximation is $[0,L]$8 but not analytic, preserves local approximation accuracy through

$[0,L]$9

and avoids explicit matching conditions at breakpoints. The paper also shows the derivative-accuracy tradeoff $(1-x^2)T_n(x)$00 as the overlap parameter $(1-x^2)T_n(x)$01, and suggests balancing this with Chebyshev differentiation growth by choosing $(1-x^2)T_n(x)$02 (Aiton et al., 2017).

A third variant is asymptotic adaptation on the infinite line. Under the map

$(1-x^2)T_n(x)$03

the rational Chebyshev families

$(1-x^2)T_n(x)$04

have distinct boundary behavior as $(1-x^2)T_n(x)$05: $(1-x^2)T_n(x)$06 remain bounded, while $(1-x^2)T_n(x)$07. Interleaved hybrid bases $(1-x^2)T_n(x)$08 and $(1-x^2)T_n(x)$09 are then used for parity-mixed operators on $(1-x^2)T_n(x)$10, preserving sparse banded discretizations and $(1-x^2)T_n(x)$11 transform complexity (Miquel et al., 2017).

The principal limitations are also basis-dependent. The CNTRC beam framework assumes Euler–Bernoulli kinematics and omits transverse shear deformation, rotary inertia, and damping; the same source states that Timoshenko theory would require adapting both $(1-x^2)T_n(x)$12 and the rotation field $(1-x^2)T_n(x)$13 (Jalili et al., 15 Sep 2025). The degenerate elliptic setting requires an integrable weighted inverse, expressed through the weighted Poincaré and Lax–Milgram framework; non-homogeneous Dirichlet data require lifting functions or augmented bases, and discontinuous or sign-changing coefficients break the weighted Sturm–Liouville identity used to obtain sparsity (Yuan et al., 29 May 2026). The partition-of-unity construction gains global smoothness, but because the blended approximation is $(1-x^2)T_n(x)$14 rather than analytic, it is not an analytic global basis in the strict spectral-function sense (Aiton et al., 2017).

Boundary-adapted Chebyshev basis functions therefore comprise a family of related constructions rather than a single formula. Their defining feature is the deliberate encoding of endpoint or asymptotic behavior into the approximation space, with the practical consequences appearing in exact or weak boundary satisfaction, altered coefficient asymptotics, different endpoint derivative scaling, and often markedly improved matrix structure for spectral discretizations (Jalili et al., 15 Sep 2025, Zhang et al., 2021).