Why coefficient accuracy matters
A spectral method replaces a function by finitely many waves. Even when the correct modes are known, their coefficients are integrals; a computer often sees only measured values at finitely many positions. If those positions are uneven, an unweighted average gives each measurement the same influence even though the measurements represent intervals of different lengths. One must say what integral the data approximate and how much each sample contributes.
We will use a simple left-endpoint rule, not because it is always the best rule, but because its error can be proved directly and its weights reveal the geometry. It applies to any finite partition, with no assumption that adjacent intervals have comparable lengths. It is a weighted generalization of the discrete Fourier coefficient: on an equally spaced grid, its weights all equal and it reduces to the usual equally weighted discrete Fourier transform. The estimates below control an approximation; they do not automatically control a maximum error or a derivative.
The continuous coefficients and a weighted rule
On let be complex-valued, with matching endpoints, and write
The integer fixes the retained space.
Definition 2.3.1
(Weighted left-endpoint Fourier rule)
Choose a finite partition with and . For , suppose the observed value is . Define
Thus uses exact samples but numerical quadrature, while also includes measurement error. The weights are positive and sum to one.
A sample at represents the interval immediately to its right; changing to midpoint or trapezoidal sampling changes the rule and requires its own error analysis.
Theorem 2.3.2
(A finite-mesh coefficient bound)
Suppose and on . For every integer ,
If for every , then . More generally it is at most .
Proof
(Freeze the integrand at each left endpoint)
Put . Since and , the triangle inequality gives
Subtract the integral from the rule, split into cells and bound each contribution:
Because and , the second bound follows. The noise contribution is . Its absolute value is at most , which is at most under the uniform bound. Weighted Cauchy–Schwarz gives the root-mean-square bound because . These are deterministic bounds: no independence or zero-mean noise is assumed.
Proof X-ray
Where each hypothesis enters the coefficient bound
The one move is to compare with its frozen value on the cell to the right of . The Lipschitz constant of contributes ; the bound is needed only because the phase also moves, at speed , and it multiplies . Integrating the distance over one cell gives , so the estimate is a sum of squared cell lengths: one wide cell is penalized quadratically, but nothing requires neighboring cells to be comparable. The step closes the argument only by giving up that information. The noise part uses none of the regularity: it needs only that the weights are positive and sum to one. Hence refinement shrinks the quadrature term for each fixed , while a bound on is passed through unchanged.
An exact decomposition before an upper bound
The exponentials are orthonormal for the normalized inner product, and our bounded lies in .
Theorem 2.3.3
(Exact error decomposition)
For every trigonometric polynomial ,
Proof
(Orthogonality of the residual)
Write . For , the inner product of with is , so the first term is orthogonal to the span of the retained exponentials, which contains the second term. The Pythagorean theorem gives , and orthonormality evaluates the last norm as .
Taking gives
Completeness of the trigonometric system additionally identifies the first term as . This is Parseval's tail, and requires exact coefficients. The finite identity itself needs only orthogonality, not completeness. The second term can be split coefficient by coefficient as
These two complex errors may reinforce or cancel.
Common mistake
(Adding the squared quadrature and noise errors)
It is tempting to write , as in the exact decomposition. That identity came from orthogonality of functions; two complex numbers have no such property. If both errors equal and in the same direction, the true square is , not . Bound the modulus of the sum by the triangle inequality first, and square afterwards.
With and , the triangle inequality yields the usable certificate
For fixed the quadrature part tends to zero as , while a fixed noise bound need not. For a sequence with both and , the right-hand side proves convergence only if the whole coefficient-error sum tends to zero too. For example, under fixed and the uniform noise bound , the sufficient conditions and make our coefficient bound vanish. They are sufficient, not necessary; in practice errors can cancel or a better quadrature rule can improve the estimate. An absolute pointwise error in the finite polynomial is bounded by the sum of the coefficient errors, but the tail alone gives no pointwise tail bound.
Worked example: three unequal cells
Worked example 2.3.4
(A spurious low-frequency coefficient)
Let and use , , , . The cell weights are , and the samples are . For the exact coefficients are and . The rule instead gives
and, since takes the values ,
There is no truncation error: . With noiseless samples, orthogonality gives the exact squared error
The coefficient at happens to be exact because its integrand is constant. That does not make the other computed coefficients exact. Here , , and ; the theorem bounds the zero-mode error by , comfortably above its actual value . The estimate is valid despite the large gap, but a large upper bound is not an accuracy certificate for a small tolerance.
Aliasing is not cured by a coefficient label
Counterexample mode
Two different functions with the same samples
Take equally spaced nodes and . Every sample equals , just as it does for the constant function. For , the exact projection is zero, yet the sampled rule produces . Hence it invents a constant component even with no noise. Under , its other retained discrete coefficients vanish as well, so the reconstructed polynomial is and its squared normalized error is : one unit of exact truncation error plus one unit of coefficient error.
This family changes as changes: its Lipschitz constant is . It therefore does not contradict convergence of the quadrature for a fixed Lipschitz function and a fixed frequency as . Unequal nodes may remove this particular exact equality of samples, but finite measurements still require assumptions on the unknown function; nonuniformity alone is not an anti-aliasing theorem.
Common mistake
(Averaging uneven samples with equal weights)
Replacing by approximates an integral against the density of the nodes, not against . Put equal cells on and equal cells on , and take , so . As , the mesh size tends to zero and the weighted rule converges to , but the equal-weight average is dominated by the samples on the left half and converges to the mean of there, . Refinement does not repair the wrong weights.
Noise, differentiation, and interpretation
Worked example 2.3.5
(A constant offset survives refinement)
Suppose every measurement carries the same additive error . Then
For the sum of the weights is one, so the zero-mode error is exactly , however fine the partition. On an equally spaced grid, and when does not divide , so the offset stays in the constant mode. On the three-cell partition it leaks:
each of squared modulus . For the noise alone contributes to the coefficient error, below the crude bound from . A probabilistic noise model may lead to a smaller typical error, but this requires additional assumptions; it does not follow from a deterministic bound.
Worked example 2.3.6
(Reading the certificate for a smooth signal)
Take , , the uniform grid with , and . Here and the tail vanishes, because . On a uniform grid , so , giving and . The certificate reads
that is, an error of at most about . The bound is far from sharp here. Each integrand with is a combination of with , and the uniform left-endpoint rule integrates these exactly, so . Only the noise remains, and its contribution is at most , an error below . The certificate is valid; the sharper value uses exactness of this particular rule.
Coefficient errors can be more consequential when the computed polynomial is differentiated. For a nonnegative integer , orthogonality gives
Here also for . This is an identity between finite polynomials. It is not a bound on the derivative of minus that of the reconstruction unless has the required weak derivative and its derivative tail is separately controlled. The factors explain why tiny high-frequency coefficient errors can be magnified by differentiation.
Exercises
Quick check
(For and the partition , calculate and and compare them with the exact coefficients.)
The cell weights are .
Solution · Answer
. Also , while . Unequal interval weights are essential; exact integration of constants does not imply exact integration of the oscillatory integrand.
Quick check
(Suppose , , and . Give a guaranteed bound for using the maximum-mesh estimate.)
Combine the quadrature and noise parts by the triangle inequality.
Solution · Answer
The quadrature bound is . The noise contribution is at most , so the triangle inequality gives . This upper bound need not be attained.
Quick check
(For , retain and suppose the only nonzero computed coefficients are and . Find the exact normalized squared error.)
Use the exact error decomposition.
Solution · Answer
The exact projection is zero and . Orthogonality gives .
Quick check
(The only coefficient errors in a degree- polynomial are and . Compute the squared normalized differences between the computed polynomial and , and between their first derivatives. Why is the latter not necessarily the error in approximating ?)
Differentiation multiplies the coefficient of by .
Solution · Answer
The polynomial error squared is . Differentiation kills the constant error and multiplies the second-mode error by , giving . To conclude anything about one must also account for the derivative of the discarded tail and justify that derivative.
Further study
This unit gives a deterministic first-order rule on a one-dimensional periodic interval. Higher-order and adaptive quadrature, multidimensional sampling, floating-point roundoff, statistical noise models, and stable nonuniform fast transforms need separate analysis.