Motivation
A Fourier series contains infinitely many modes, while a computer stores finitely many numbers. Two operations are therefore unavoidable: truncate the series and estimate its coefficients from samples. They are not the same operation. Truncation discards high frequencies when exact coefficients are known; sampling can fold those high frequencies into lower ones before any truncation takes place. The distinction becomes especially visible near a jump, where a partial sum overshoots even though it converges in mean square.
We use the normalization and . A function on an interval is treated as periodic only after its endpoints have been identified. Smoothness on the open interval alone does not guarantee a smooth periodic extension.
Definitions
Definition 2.7.1
(Truncation, samples, and aliasing)
For an integer , sample at , , and define the discrete coefficient
Indices are integers modulo . Distinct continuous frequencies that give identical values at every sample are aliases: for every integer . A spectral truncation is instead , formed from the exact continuous coefficients.
For every , frequencies differing by a multiple of agree at the samples; increasing changes which frequencies alias but does not eliminate aliasing altogether. Thus no deterministic procedure using only these samples can distinguish two functions whose sample values agree. A band-limit or regularity assumption supplies the information that the data alone do not contain.
Main results
Theorem 2.7.2
(Smoothness and truncation error)
Let be an integer. Suppose is absolutely continuous on , , and the endpoint traces of match. Then for ,
The norm here is the ordinary, unnormalized integral norm.
Proof
(Integrate by parts, then weight the tail)
Integrate by parts times. Matching endpoint traces remove every boundary term, yielding the coefficient identity; the zero Fourier coefficient of is also zero, since it is the endpoint difference of divided by . Parseval for the derivative gives . Parseval for gives the squared truncation error as . For these terms ; substituting the derivative coefficients and taking square roots proves the bound. This is an upper bound, not a claim that every smooth function has exactly this rate. The matching traces are essential to the integration by parts.
Theorem 2.7.3
(Aliasing formula for absolutely convergent series)
If and at the sample points, then
In particular, if all coefficients outside a set of representatives modulo vanish, the discrete coefficients recover those retained coefficients exactly.
Proof
(Collapse the finite geometric sums)
Absolute convergence permits interchange of the finite sample sum and the Fourier series. The inner sum is , equal to one if is divisible by and zero otherwise, by the finite geometric sum. Only the displayed aliases survive. With at most one nonzero coefficient in each residue class, each sum contains the desired coefficient alone. If , the frequencies are distinct representatives; for even the two frequencies and coincide on the samples and cannot both be independently recovered.
Proof X-ray
What the aliasing proof actually uses
The proof has one computation and one justification. The computation is the discrete orthogonality of the sampled exponentials: averaging over gives one or zero according to whether divides , and this is what sorts every continuous frequency into a residue class. The justification is the exchange of the finite sum over samples with the infinite Fourier series. It needs two things: the series must equal the stored value at every sample, and it must converge absolutely so that the rearranged sum over a residue class is meaningful. For the square wave the first point fails at : its Fourier series sums to there, so a stored value is not what the formula sees. Exact recovery is a separate step, using a band-limit so that each class contains at most one nonzero coefficient. By contrast, the decay theorem used matching periodic traces; none of these hypotheses follows merely from having measurements.
Worked examples
Worked example 2.7.4
(A frequency hidden by sampling)
Take and . At , . The samples therefore produce discrete coefficients at residue classes and , exactly as for . The continuous coefficient at of is zero: the apparent low mode is an alias, not its Fourier projection. Sampling faster or imposing a valid band-limit can resolve the ambiguity.
Worked example 2.7.5
(A smooth interior with an endpoint jump)
On let and extend periodically. Integration gives for . Although in the interior, the endpoint values and do not match; the periodic extension jumps from to . Hence the derivative theorem's trace hypothesis fails. Parseval gives
The norm bound is consistent with a jump and is much slower than an arbitrarily high power. Integrating from to infinity proves the displayed tail bound.
Worked example 2.7.6
(The persistent overshoot of a square wave)
Let for and for , periodically extended. Its odd-frequency partial sum is
Differentiating the finite sum gives . The first positive local maximum occurs at . Setting and passing to a midpoint Riemann sum gives
At this is approximately : the overshoot above the right-hand value one tends to about , or of the jump of height two. Its location tends to the jump, so this does not contradict convergence at each fixed continuity point or convergence in .
Methods and reasoning
Cesàro averaging and the Fejér kernel
A filter changes how retained modes are weighted. The Fejér mean averages successive partial sums:
It equals convolution with . This kernel is nonnegative and has normalized integral one, because expanding the square leaves only its diagonal terms after integration. For away from zero modulo , the geometric-series formula bounds by a constant divided by . Thus its mass outside any fixed neighborhood of zero tends to zero.
If is continuous and periodic, it is uniformly continuous. Given , choose a small neighborhood in which for every . Split the convolution integral into this neighborhood and its complement. The first contribution is at most because has unit normalized mass; the second tends uniformly to zero because is bounded and the kernel mass there tends to zero. Hence uniformly. For a bounded integrable periodic function with finite one-sided limits at a jump, the same split and the evenness of give their midpoint as the limit; the two half-kernels carry exactly half the mass each.
For a bounded real function, convolution with this nonnegative unit-mass kernel stays between its essential lower and upper bounds. In particular, Fejér means of the square wave cannot overshoot beyond . This comes at a cost: high retained modes are damped, broadening transitions and changing the approximation. A generic filter may have a sign-changing kernel and need not preserve this range. Fejér averaging is a specific, proved remedy, not a claim that all filters remove Gibbs oscillations.
What the FFT computes
An FFT is an efficient algorithm for a finite discrete Fourier transform; it returns , not automatically . The alias relation holds for each ; the alias itself changes when changes, so it is not one fixed high-frequency function in a refinement experiment. For a fixed unbandlimited function whose Fourier series is absolutely convergent to the sampled values, its coefficients contribute according to the alias formula at each . For a trigonometric polynomial of degree sampled at equispaced points, the aliasing theorem recovers each coefficient exactly in exact arithmetic. For an unbandlimited function, each discrete coefficient includes its whole alias class. With and the theorem's smoothness assumptions, integration by parts and Cauchy–Schwarz yield , so for a fixed retained the alias terms decrease as sampling is refined. The size of that error depends on the derivative and on how close is to the edge of the represented band; it is distinct from finite-precision FFT error.
If a function is smooth inside but its periodic endpoint traces mismatch, an FFT sees the periodic jump. Artificially joining incompatible endpoints can therefore produce slow coefficient decay and ringing. One may change the model, subtract a known boundary-matching function, or choose a basis adapted to the boundary conditions. None of these repairs is equivalent to claiming that the original periodically extended data were smooth.
Residue classes become a periodization of the Fourier transform, and a band limit permits exact reconstruction from equally spaced samples.
A path to generalization
What changes for functions on the whole line?
Residue classes become a periodization of the Fourier transform, and a band limit permits exact reconstruction from equally spaced samples.
Let have transform vanishing for , and sample at spacing . Then nothing is lost:
Without the band limit, the samples determine only the periodized transform on , the analogue of the alias classes here. The finite residue classes become a continuum of folded frequencies, and reconstruction is an infinite interpolation series rather than a finite sum.
An error budget for sampled coefficients
Suppose we retain frequencies from to and have at least samples, so these frequencies occupy different residue classes. Let use the coefficients computed from those samples. Add and subtract the exact projection before applying the triangle inequality. The total distance from the periodic model to the computed polynomial is bounded by the sum of two terms: the distance from the model to its exact projection, and the distance between that projection and the sampled polynomial. In fact the two pieces are orthogonal, since the first is orthogonal to every retained exponential, so their squares add exactly; the triangle inequality is the convenient weaker form. Orthogonality also makes the square of the second distance equal to times the sum of the squared differences between corresponding discrete and continuous coefficients.
This identity gives the terms distinct meanings. The first is truncation error and remains even if the integrals for the coefficients are evaluated exactly. The second contains aliasing, and may also contain measurement and floating-point errors. For instance, if every sample is perturbed by at most a fixed amount, the discrete average defining each coefficient changes by at most that amount: every complex exponential has modulus one. Combining the number of retained coefficients with this bound gives a conservative estimate for the total coefficient error. More sampling can reduce aliasing of a fixed smooth function, but it cannot by itself remove an error already present in each measurement.
There may be a third difference before either numerical step. If the physical problem is defined on an interval but we replace it by a periodic model with mismatched endpoint traces, we must compare that model and its boundary behavior with the original problem. Taking more Fourier modes solves the chosen periodic approximation more closely; it does not make incompatible physical boundary conditions agree. In practice one should state the model, the sampling procedure and the retained frequencies before interpreting a graph or quoting a convergence rate.
Why a persistent peak does not prevent mean-square convergence
Return to the square wave. Parseval expresses the squared integral error as a tail of squared odd-frequency coefficients. Those coefficients decay like the reciprocal of the frequency, so their squared tail tends to zero. The peak described in the example occurs at a point whose distance from the jump decreases as the number of retained modes increases. Thus the location at which the large error is observed moves with the approximation; it is not a fixed continuity point. The integral error measures both height and width, while the maximum error can remain large on a shrinking neighborhood.
The contrast between projection and averaging is equally instructive. The exact partial sum is the best approximation in the mean-square norm among polynomials of its degree, but its convolution kernel takes both signs. A signed combination is not a weighted average and can exceed the range of the data. The Fejér kernel is nonnegative and has total mass one, so its result stays in that range. It does not follow that Fejér averaging gives a smaller mean-square error than the optimal partial sum of the same degree; its benefit here is range preservation and more controlled behavior near the jump, at the price of damping high-frequency information.
Common misconceptions
Each of the following confusions mixes two statements that were proved under different hypotheses.
Common mistake
(An invisible frequency is not an absent one)
A high frequency can be invisible in the samples without being absent from the function. For the samples of and coincide, so a discrete coefficient at residue cannot tell which was present. Only a band-limit or a decay assumption decides between them.
Common mistake
(A persistent peak does not contradict convergence)
A narrow Gibbs region can have a persistent maximum even while its total squared contribution shrinks. The overshoot of about sits at , which moves toward the jump; at each fixed continuity point and in mean square the partial sums still converge.
Weighting frequencies also changes the approximant: filtered and unfiltered partial sums should not be compared as if they used identical coefficients.
Counterexample mode
Interior smoothness without matching endpoints
Claim: if is smooth on the open interval and , then . Take on . Then and , so the claimed bound is . The worked example gives
which exceeds for every , because . The claim fails for every . The repair is the trace hypothesis: integration by parts produces the boundary term , and only matching periodic traces remove it.
Quick check
Quick check
(For , which frequency between and aliases with ?)
Reduce modulo .
Solution · Answer
Since , at every sample. No calculation from those samples alone distinguishes them.
Exercises
Quick check
(Sample at equispaced points. Identify a lower-frequency cosine with the same samples and explain the corresponding alias classes.)
Reduce modulo .
Solution · Answer
, so . The nonzero continuous coefficients at and enter the discrete residue classes and , respectively, just as the coefficients of do.
Quick check
(A periodic function has , with matching at the endpoints and absolutely continuous. Bound in terms of .)
Check the hypotheses of the smoothness theorem for .
Solution · Answer
Apply the smoothness theorem with : . The hypotheses make both integration-by-parts boundary terms vanish.
Quick check
(For the square wave above, show that is a local maximum of when .)
Use the closed form of the derivative.
Solution · Answer
On the denominator is positive. The numerator changes from positive to negative at , which lies in that interval for . Thus the derivative changes sign from positive to negative.
Quick check
(Explain why a Fejér mean of any real function satisfying almost everywhere also lies between zero and two.)
Use the two properties of the Fejér kernel.
Solution · Answer
The Fejér kernel is nonnegative and has normalized integral one. Its convolution with is a weighted average, so integrating the inequalities against the kernel gives .
Further study
Higher-dimensional aliasing, nonuniform samples, rigorous optimal filters, roundoff analysis and spectral approximations on nonperiodic domains require additional tools. These results cover a finite periodic model and do not imply a universal convergence rate.