Are the Chebyshev nodes equally spaced?

No. The common cosine grid clusters near the endpoints of the standard interval.

Is differentiation done with only neighboring nodes?

A global polynomial differentiation matrix generally couples a node to many or all other nodes.

Does the word spectral refer to sound or light here?

No. It refers to the mathematical approximation framework, not an application involving physical spectra.

AI Glossary / Mathematics and optimization / Advanced

Chebyshev pseudospectral method

Definition

The Chebyshev pseudospectral method approximates a continuous function by a polynomial represented through values at Chebyshev nodes, then enforces equations at selected nodes. Derivatives are obtained from the interpolating polynomial. It is used for differential equations and, in optimal control, to turn continuous trajectory problems into finite-dimensional optimization problems.

What it means

A continuous trajectory contains values at infinitely many times. A computer must use a finite representation. This method places sample nodes in a deliberate nonuniform pattern and connects the nodal values with a global polynomial.

The word pseudospectral refers to working with values at collocation points while retaining the global approximation associated with spectral methods. It does not mean the method is fake or only approximately numerical. Chebyshev polynomials and their associated nodes provide the underlying approximation structure; this differs from Chebyshev iteration for linear systems.

How it works

A common endpoint-including grid uses zⱼ=cos(jπ/N), for j=0,…,N, on [−1,1]. Nodes cluster near both ends. Physical time is mapped onto this standard interval, with the necessary derivative scaling.

Represent the unknown function with the degree-N polynomial interpolating its nodal values. A differentiation matrix maps those values to the derivative of that polynomial at the nodes. Differential equations become algebraic conditions at the chosen collocation nodes, with boundary conditions imposed consistently.

For optimal control, state and control nodal values become decision variables. Discretized dynamics, endpoint conditions and path restrictions constrain them; a suitable quadrature approximates integral costs. A nonlinear optimizer solves the resulting finite problem. Node choices, quadrature and boundary treatment belong to a particular formulation and should not be mixed casually across variants.

Three practical examples

1. Boundary-value modeling: approximate a smooth temperature profile subject to specified endpoint temperatures and a differential equation describing the interior.

2. Trajectory planning: represent a vehicle's states and controls at Chebyshev nodes, then optimize a discretized performance criterion under dynamic and endpoint constraints.

3. Scientific computation: calculate derivatives or eigenvalue approximations for smooth differential models where global polynomial approximations are appropriate.

Think of it this way

Think of describing a smooth curved rail using carefully placed reference points and one continuous mathematical curve through them. The shape between points is supplied by the curve, not by drawing unrelated short straight segments.

The analogy has limits. Real measurements may contain noise, while high-degree interpolation and differentiation can amplify numerical difficulties. A smooth-looking curve also does not prove that it obeys the underlying equation between the reference points.

Worked example

Consider an original example with N=2 and nodes ordered as z=(1,0,−1). The first-derivative matrix is D=[[1.5,−2,0.5],[0.5,0,−0.5],[−0.5,2,−1.5]]. Let the nodal values come from p(z)=z², giving v=(1,0,1).

Multiplying gives Dv=(2,0,−2), exactly matching p′(z)=2z at the three nodes. Applying D again gives D²v=(2,2,2), matching the constant second derivative.

Now interpret z as a rescaled time on [0,4], with t=2(z+1). The same nodal vector represents x(t)=(t/2−1)² at times (4,2,0). Because dz/dt=1/2, the time-derivative matrix is D/2. It returns (1,0,−1), the correct derivatives with respect to time. Forgetting the scaling would double each nonzero derivative.

This calculation is exact because the interpolated function is already a degree-two polynomial. General functions introduce approximation error. The example checks differentiation and time mapping by hand; it does not execute an optimal-control solver or establish a trajectory's feasibility.

Illustrative case study

Fictional case: a robotics team prototypes a smooth arm-motion trajectory with a Chebyshev discretization. The optimizer reports small dynamic residuals at the collocation nodes. An independent simulation, however, reveals a short constraint violation between nodes.

The team refines the representation and inspects the reconstructed trajectory on additional points. It also investigates whether the motion contains a sharp transition that is poorly represented by one global polynomial. The final report distinguishes the discrete optimizer's success from the separate evidence supporting the continuous trajectory.

Common misconceptions and mistakes

Do not infer continuous feasibility from nodal feasibility alone. Path constraints can be violated between collocation points.

Do not omit the time-interval scaling of derivative matrices or integral weights. Also keep node ordering consistent with every state vector, operator and boundary condition.

Do not assume that increasing polynomial degree always improves the answer. Smoothness, rounding error, conditioning and discontinuities affect performance. Splitting a domain or using a different discretization can be more appropriate than one very high-degree polynomial. A local nonlinear-programming result is not automatically a globally optimal continuous-control solution.

Business use

This method can reduce the number of unknowns needed for accurate representations of sufficiently smooth problems. Its business relevance is indirect: it supports engineering simulations and planning systems on which operational choices may depend.

Evaluate the complete workflow, including model validity, solve time, interpolation error and independent trajectory checks. Compare results after increasing resolution or altering the domain partition. Keep algorithmic tolerances separate from physical requirements; a numerically small residual must be interpreted in the units and scale of the actual system.

Check your understanding

Why is D/2 used when mapping the example from z∈[−1,1] to t∈[0,4]?

Answer: The coordinate relation t=2(z+1) gives dz/dt=1/2. The chain rule therefore multiplies derivatives with respect to z by one half.

Related terms

Collocation (remote sensing)
collocation
Chebyshev polynomial
differentiation matrix
spectral method
optimal control
nonlinear programming
quadrature
mesh refinement

Learn next

derive the three quadratic interpolation basis functions for nodes (1,0,−1) and differentiate them to reconstruct D. Then compare nodal residuals with residuals at additional points for a function that is not polynomial.

Sources

Sources: Weideman and Reddy's primary differentiation-suite description was read: https://appliedmaths.sun.ac.za/~weideman/research/differ.html . Fahroo and Ross, “Direct Trajectory Optimization by a Chebyshev Pseudospectral Method,” indexed primary abstract consulted: https://doi.org/10.2514/2.4862 . Full trajectory-paper retrieval failed; its abstract establishes transcription scope. Exact optimal-control variant details require publication-level source review. The differentiation example is original and independently calculated.

Previous: Chauvenet's criterion Next: Checkpoint