An arbitrary continuous function is not determined by its values on an evenly spaced grid. It may change between the grid points while leaving every recorded value unchanged. A frequency restriction changes the problem: if the Fourier transform vanishes outside a bounded interval, the grid values become coordinates in an orthonormal expansion.
The sampling theorem is therefore more than a formal interpolation formula. It identifies a function class, a continuous representative, a permissible spacing and a norm in which the data determine the answer stably. We will prove the theorem through the Fourier-series basis on a finite frequency interval. Poisson summation is not needed for this proof.
Motivation
The cardinal sine
equals one at zero and zero at every other integer. Thus is a natural function for placing one value at the grid point . But infinitely many such interpolation functions do not automatically give a convergent series, and agreeing at grid points does not prove equality between them.
The missing information is the frequency support. On the interval , the functions are a complete Fourier family. After normalizing that frequency basis, its inverse transforms are times the shifted sinc functions. Orthogonality will give convergence and a norm identity, while bounded frequency support will upgrade norm convergence to uniform convergence on the entire real line.
Definitions
We use complex with its usual unnormalized norm and inner product . The symmetric angular-frequency transform is
when the integral exists, and otherwise its unitary extension. The inverse has the same prefactor and exponential .
Definition 2.5.1
(Finite-energy bandlimited functions)
For , write
The number is an angular-frequency bandwidth. This is a closed subspace of , because multiplication by the indicator of is an orthogonal projection in frequency.
An equivalence class cannot be sampled directly: changing values on countably many points leaves the class unchanged. For , however, is integrable by Cauchy–Schwarz on its finite support. Its inverse integral gives a preferred representative
The overlap theorem for the integral and inverse transforms shows that this equals the original class almost everywhere. All samples and pointwise assertions below refer to this representative.
Definition 2.5.2
(The sampling grid and sinc functions)
For spacing , the samples are , . At the critical relation , set
A series indexed by is initially interpreted through finite symmetric sums. We will also prove absolute convergence at every point for the reconstruction series.
Theorem/proposition
Theorem 2.5.3
(Bounded evaluation and smooth representatives)
The preferred representative of every is smooth, and for every nonnegative integer ,
In particular, point evaluation is a bounded linear functional on , with norm exactly . At , .
Theorem 2.5.4
(Exact critical sampling and reconstruction)
Let and . The functions form an orthonormal basis of . For every ,
The reconstruction converges in and uniformly for all real ; at each it is absolutely convergent. Conversely, every sequence is the sample sequence of exactly one preferred representative in , reconstructed by the displayed formula.
For a nonnegative integer , let retain only . Then
Theorem 2.5.5
(What the spacing threshold means)
For , the same reconstruction and norm identity hold whenever , by viewing in the larger space . If , not every sequence is compatible with the smaller bandwidth.
If , sampling on is not injective on : two different finite-energy bandlimited functions can have all the same samples. Thus this spacing cannot reconstruct every member of uniquely.
The equality case is included. A frequency value assigned at either endpoint of the support does not affect an transform. Pure tones with spectral point masses belong to a different setting and must not be substituted for finite-energy functions.
Proof sketch or proof idea
Bounded support makes evaluation meaningful
Proof
(Differentiate under the integral, then test the kernel)
Differentiating the inverse integral gives
Each integrand is absolutely integrable. Difference quotients are dominated by a constant times or on the bounded interval, so dominated convergence justifies all derivatives and their continuity. Cauchy–Schwarz then gives
The power integral and Plancherel yield the claimed constant. The representative is also uniformly continuous, for example by its bounded first derivative. Two continuous representatives of the same class agree everywhere, because a nonzero continuous difference would be nonzero on an interval of positive measure.
To see that the evaluation norm is sharp, fix and choose
Its inverse transform is the function
Here and , so the ratio of the point value to the norm is . In fact the inverse formula says .
The frequency basis becomes the sinc basis
Proof
(Orthonormal basis, uniform tails and the converse)
On with , the functions
are an orthonormal basis of . Orthogonality follows by integrating the exponentials; completeness is the Fourier-series completeness theorem after the change of variable . Extend by zero outside the interval.
Their inverse transforms can be computed by an ordinary integral:
At , the integral equals directly, matching . Unitarity carries this orthonormal basis onto an orthonormal basis of . For ,
The Hilbert-space expansion and Parseval therefore give the reconstruction in and its sample-energy identity. Orthogonality gives the exact squared error after truncation. This argument proves convergence rather than assuming that an interpolating series must equal the function.
Every residual still lies in . Apply the evaluation bound uniformly in and then the exact error:
The tail tends to zero, proving uniform convergence on the whole line.
Conversely, for , the series defines an element of by orthonormal completeness. Bounded evaluation lets us pass to the limit at each grid point. Since , the samples are exactly . The sample norm identity proves uniqueness.
Proof X-ray
Samples are coefficients only after the bandwidth is fixed
Three inputs carry the proof, each used once. Completeness of Fourier series on , after , makes a basis; unitarity of the transform moves that basis to ; and the inverse integral turns the coefficient into the sample . The critical relation is used exactly here: for a larger band the -periodic exponentials cannot span , and for a smaller one they are no longer orthogonal. Uniform convergence is a separate step. It applies the bounded-evaluation theorem to the residual , which stays in , and the same bound returns in the converse, where it lets the series be evaluated at each grid point. Sinc interpolation without the frequency-support hypothesis has none of these steps, even when all its grid values are correct.
A cardinal identity controls absolute convergence
Proof
(Parseval applied to the evaluation kernel)
At critical spacing the evaluation function above is
Apply Parseval in the sinc basis to . Its coefficients are , where sinc is real and even. Hence
Cauchy–Schwarz now gives
The same inequality for any tail gives a bound by the corresponding tail. Thus pointwise absolute convergence and the uniform tail estimate are consistent. We have not needed .
Concept lensStructural
Reconstruction is an isometry after one scale factor
The map identifies the bandlimited Hilbert space with . Choosing samples is choosing orthonormal coordinates. Pointwise values are meaningful because frequency confinement makes evaluation continuous; arbitrary point values have no such status.
Narrower spacing and genuine finite-energy aliasing
Proof
(Nested bandlimited spaces and a modulated pair)
If , then , so the theorem for applies directly. In the strict case, arbitrary data reconstruct a member of , which need not belong to . The narrower support imposes additional compatibility relations on the samples. For example, the data and all other reconstruct in . Its transform is the nonzero constant on , so it cannot belong to when . Uniqueness in rules out any other interpolant from the smaller space.
If instead , choose and let
The inverse-indicator calculation already made for the sinc basis, now with spacing , gives . Thus its transform is supported in . Modulation shifts support, so the transforms of and lie in the disjoint intervals and , both inside . The modulation rule extends from Schwartz functions to by density and unitarity. Thus these are distinct nonzero members of . Yet
This is aliasing with finite energy, rather than a purely formal comparison of nonintegrable sinusoids. Their nonzero difference has every sample equal to zero, proving failure of injectivity.
Worked examples
Worked example 2.5.6
(A single nonzero sample)
At bandwidth , the sequence and for reconstructs
The value and norm are exact, and the evaluation bound is attained at . A sequence supported at one grid point does not reconstruct a function supported near that point: the sinc profile extends along the whole line.
Worked example 2.5.7
(Smoothness and rapid decay do not replace a bandlimit)
Let and . This nonzero Schwartz function has for every integer . Its transform is
For every the two positive Gaussian terms are different, so the transform does not vanish outside any bounded interval. There is no bandwidth for which . Its smoothness, rapid decay and square-summable zero data do not justify sinc reconstruction; the missing support hypothesis is decisive.
Worked example 2.5.8
(No uniform finite-window rate follows from energy alone)
For each , choose
It has norm one, but all retained samples with vanish. Its truncated reconstruction is zero, so the error is one and the uniform error is . Thus no error estimate tending to zero with , depending only on , works for every bandlimited function.
Each fixed function still has a sample tail tending to zero. The example changes the input as changes, placing its energy outside the retained window. A rate requires additional information, such as a weighted sample-tail bound.
Counterexample mode
The critical sine is not a finite-energy counterexample
The function vanishes at every point of . It may look like a counterexample exactly at the critical bandwidth, but it has infinite squared integral on the real line and is not in . A finite-energy transform cannot carry a nonzero point mass at the edge of the band. The equality case of the theorem is valid for the stated class; adding persistent pure tones changes that class.
Common mistakes
Common mistake
(Sampling an arbitrary representative)
Point samples must come from the preferred continuous representative. Redefining a general function at every grid point can create arbitrary data without changing its equivalence class; such redefinitions are not the bandlimited representative used by the theorem.
Common mistake
(Confusing angular and ordinary frequency)
Angular frequency and ordinary frequency give different numerical cutoffs. Here the critical relation is . With ordinary frequency , the same condition is a sampling frequency at least twice the largest ordinary frequency. These are the same condition after converting units.
Common mistake
(Calling small sample errors stable)
Stable recovery is quantified in the sample norm. If is added to the samples, its reconstruction error has norm exactly and uniform norm at most . Merely saying that every error is small does not imply that infinitely many errors form an sequence.
The Weierstrass test based on does not apply to data. The sampling series still converges uniformly by the Hilbert-space and cardinal estimates.
Summary
Frequency confinement turns grid values into orthonormal coefficients. Plancherel gives the exact sample-energy identity, bounded evaluation gives uniform reconstruction, and the cardinal identity gives pointwise absolute convergence. At a spacing wider than the bandwidth permits, even finite-energy functions can alias. Finite-window accuracy depends on the actual sample tail, not on bandwidth and total energy alone.
Exercises
Quick check
(At critical spacing, suppose the only sample errors are and . Find the reconstructed error function, its exact norm and its exact uniform norm.)
Solution · Answer
The error is . The sample-energy identity gives . Write and , both real. Then by the cardinal identity. At , and , so equality holds. Thus ; the general bound is valid but not sharp for these data.
Quick check
(Use the cardinal identity to prove, for ,)
Deduce .
Solution · Answer
For nonintegral , . Substitute its square in and divide by the nonzero factor. At , the left side is . Since , pairing the equal positive and negative half-integer terms gives . The sums are nonnegative and convergent, so this pairing is legitimate.
Quick check
(Let and . Suppose and the samples satisfy . For a nonnegative integer , give explicit bounds for both truncation errors after retaining .)
Solution · Answer
If , then . Hence . The theorem gives and . This extra weighted condition is not implied by the bandlimit alone: arbitrary sequences occur at critical bandwidth.
Quick check
(Take , , and . Prove that is a nonzero member of with all grid samples zero. Compute by its two shifted frequency bands.)
Solution · Answer
The bandwidth of is , and , since its critical spacing is and its samples there are one at zero and zero elsewhere. The functions and have disjoint frequency supports and , so both lie in and are orthogonal. Their difference divided by is , giving . In particular it is nonzero. At , the sine factor vanishes, so every sample is zero.
Quick check
(Suppose is sampled on . Prove that)
Explain why the sequence , for is incompatible with this narrower bandwidth.
Solution · Answer
Apply the critical sampling theorem to the same with spacing , then evaluate its uniformly and pointwise absolutely convergent reconstruction at . This gives the displayed formula. If the proposed sequence were the samples of such an , its even-index samples would force . In particular , contradicting the proposed value zero. Although every sequence occurs in , the narrower space imposes relations between its even and odd samples.