Motivation
A periodic signal can contain oscillations at many integer multiples of a fundamental frequency. Fourier analysis asks how these oscillations combine and, more subtly, in what sense the resulting infinite sum recovers the original function. An integral formula for coefficients is only the beginning: a sequence of partial sums can converge in mean square without converging at every point, and a jump changes what one should expect at that point.
We work with period and use complex exponentials to make orthogonality transparent. All integrals in the first part may be read as Lebesgue integrals; piecewise continuous functions are sufficient for the explicit calculations. The notation identifies functions equal almost everywhere.
Definitions
Definition 1.1.1
(Coefficients and partial sums)
For its Fourier coefficient and symmetric partial sum are
The integral defines a coefficient for each integer . The symbol records these coefficients; it does not, by itself, assert pointwise convergence of the sum.
Equip with the normalized inner product and norm . Since , the coefficients are inner products with an orthonormal family. For real-valued , it is also useful to define and for ; then . The factor is part of the convention and changes corresponding energy formulas if a different normalization is chosen.
From complex coefficients to sines and cosines
For real-valued define the constant coefficient as well:
Conjugating the coefficient integral gives . Pairing the two modes and using Euler's formulas therefore gives, for every finite ,
The factor on the constant term compensates for the convention : the average of is . This is an identity between finite sums; writing the corresponding infinite series with still makes no convergence claim. Evenness makes and oddness makes , provided the equalities hold almost everywhere. These observations shorten the integrations, but neither parity nor a coefficient formula alone proves convergence.
Theorem/Proposition
Theorem 1.1.2
(Best approximation and Bessel's inequality)
For , uniquely minimizes among trigonometric polynomials . Moreover,
Theorem 1.1.3
(Fejér's theorem and completeness)
If is continuous and -periodic, then the arithmetic means converge uniformly to . Consequently trigonometric polynomials are dense in and, for every ,
The last equality is Parseval's identity for series. It is an statement, not a claim that every pointwise partial sum converges.
Proof sketch or proof idea
Proof
(Orthogonal projection and the Fejér kernel)
Orthogonality gives for . For any polynomial of degree at most , the Pythagorean identity gives
Orthogonality also gives . This proves the finite energy identity and the minimization; equality in the displayed Pythagorean identity requires in , hence as trigonometric polynomials. Letting increase gives Bessel's inequality.
For the second result, extend periodically. To see where the averaging kernel comes from, first define the Dirichlet kernel and use the finite coefficient sum:
Indeed, in each coefficient integral substitute and use periodicity to move the resulting interval of length back to . The sum is finite, so no interchange of a limit and an integral is involved. A geometric sum also gives
Now count frequencies in the average. For , the term occurs in exactly when , hence times. In the squared geometric sum it occurs once for each pair with , again times. Thus
Dividing by shows that averaging gives weight to frequency for , and weight zero outside. For example, takes the value at , whereas . Positivity belongs to the average, not to every partial-sum kernel. Substituting the identity into the finite average gives:
The kernel is nonnegative and has normalized integral one, by orthogonality. Fix and choose by uniform continuity so that whenever , uniformly in . Positivity and normalization bound the near part of by . On , the geometric-sum formula gives . Its mass there tends to zero, so boundedness of makes the far part less than for all sufficiently large . The order is to choose first, then .
Continuous periodic functions are dense in : Lebesgue-measure regularity gives density of continuous compactly supported functions in on the open interval, and each such function extends continuously and periodically with equal endpoint values. Fejér's theorem therefore makes trigonometric polynomials dense in . Given , choose such a polynomial within of . For all , the best-approximation property yields . In the finite energy identity the left-hand side tends to zero, so the symmetric partial sums of tend to : this is Parseval.
Proof X-ray
How Fejér closes the energy argument
The first half uses nothing but orthogonality, so by itself it can only give Bessel's inequality. The decisive move is to average the Dirichlet kernels: takes negative values, and the near-part estimate needs to turn into a bound on the whole near integral. Uniform continuity is used once, to choose one for all ; the far-part bound then decides how large must be, which is why comes first. Parseval closes through monotonicity: once a polynomial is within of , best approximation keeps for every . No step claims that the ordinary partial sums converge uniformly.
Why continuous periodic functions are dense
The density used above deserves a closer look: a discontinuous function can be close in mean even though no continuous function follows its jump uniformly. Write for the circle of circumference and use normalized measure . For every , continuous functions on are dense in . Besides standard facts including dominated convergence and translation invariance, the approximation argument uses regularity of Lebesgue measure on the circle.
First let be measurable. Regularity and compactness let us approximate in measure by a finite union of arcs: can be made arbitrarily small. One way is to choose compact with small, then cover by finitely many arcs contained in . The indicator error obeys
After merging overlapping arcs, replace each arc indicator by a continuous periodic ramp which equals it outside arbitrarily short neighborhoods of its endpoints and stays between zero and one. The th power of the error is at most the measure of those neighborhoods. Thus every indicator can be approximated in by continuous periodic functions.
Next approximate a bounded measurable function by a finite-valued simple function, rounding its real and imaginary parts on a fine grid; replace each indicator in that finite sum by a ramp approximation. For a general , first truncate its magnitude: satisfies by dominated convergence applied to . Combining the three approximation errors proves density. No Fourier completeness has been assumed in this argument.
This also explains a useful consequence. With , translation preserves the norm. Given , choose continuous periodic with . Uniform continuity on the compact circle then gives , and
where one first fixes and then lets tend to zero. This conclusion is for finite : if periodically, then for . Small movement changes the function by one on a short interval, however short. For , its exact error is because the symmetric difference has length .
Quick check
(For , let be the continuous periodic ramp described below. Compute and explain why this does not give uniform convergence.)
The function is zero outside , linear from to on , equal to on , and linear down to zero on .
Solution · Answer
Only the two ramps contribute, each with integral . Including the normalized measure gives . Yet : the error approaches one on sets of positive measure near an endpoint. This is precisely the difference between mean and uniform approximation.
Parseval for two functions
Energy is the inner product of a function with itself. To compare two different signals, the corresponding identity retains phase information.
Corollary 1.1.4
(Parseval's identity for two functions)
For , the series below converges absolutely and
Proof
(Pass to the limit in the finite identity)
Cauchy–Schwarz and the energy form of Parseval bound the sum of absolute values by . For the identity, orthogonality gives the finite equality
Both projections converge in . In particular, , so Cauchy–Schwarz applied to
shows that the left side tends to zero. This proves the identity without multiplying two infinite pointwise series or interchanging their sum and an integral.
Worked example 1.1.5
(Comparing a signal with a shifted copy)
Let and , with real . Direct integration gives . In frequency language , while and . Thus the identity gives . The conjugate in the identity reverses the phase of the second signal. At the signals have the same energy, yet their inner product is zero: energy alone cannot describe their alignment.
Quick check
(Replace by . What is ?)
Expand the shifted sine before integrating.
Solution · Answer
Expand . Orthogonality gives . This also checks the sign of the phase convention.
Worked examples
Worked example 1.1.6
(An even function and its constant term)
Let on and extend it periodically. It is even, so , and
For the last equality, integrate twice by parts: the first boundary term vanishes because , and . Thus the constant in the series is , not . The coefficients are absolutely summable, so the series converges uniformly by the Weierstrass test. Its sum has the same coefficients as by termwise integration; Fejér uniqueness for continuous periodic functions identifies the sum with . Hence, at every real under the periodic extension,
As a check, the value at gives . This conclusion uses the justified equality at , not the coefficient calculation alone.
Worked example 1.1.7
(How many modes are needed?)
The preceding calculation also tells us how to measure the error when we stop after finitely many modes. Let on , extended periodically, and let contain the frequencies . For , pairing the real coefficients above gives . The finite projection identity therefore gives the exact squared error
Here , and the constant mode contributes . Each positive frequency has a matching negative frequency, which accounts for the factor . This finite formula needs orthogonality but does not yet need completeness.
Parseval supplies the missing step from the finite identity to the tail: letting yields , and hence
For , comparison with the decreasing function gives
Taking square roots, the root-mean-square error is of order in this example. For instance, guarantees . The norm uses the normalized average over one period, so this is a root-mean-square bound, not a bound on the largest error at each point. A separate uniform estimate follows here from absolute convergence: for every . A small error alone would not have implied that pointwise bound.
Worked example 1.1.8
(A shifted interval and a different period)
Suppose a measurement is specified on , where , and is then repeated with period . Taking converts one period into . Thus the coefficients relative to the shifted origin are
The factor is forced by the change of variables, and Parseval becomes for . Changing the origin changes coefficient phases, not their magnitudes.
For a concrete calculation, take on the open interval. With and , integration by parts gives, for ,
The boundary term is essential: , but has different endpoint values. Pairing positive and negative frequencies gives
The rescaled piecewise- pointwise theorem gives the interior value . At the periodic seam the one-sided limits are and , so the sum is , irrespective of the value assigned at the seam. The normalized energy is ; in the coefficient sum the constant contributes , and the positive and negative frequencies together contribute . This checks both the scaling and the factor of two.
Worked example 1.1.9
(The periodic sawtooth)
Let for , set , and extend periodically. Oddness makes every cosine coefficient zero. For , integration by parts gives
Thus . The function has a jump from to at the identified endpoints, so cannot converge uniformly to : a uniform limit of continuous functions is continuous. At the jump each sine term vanishes, and the partial sums equal zero, the midpoint of the one-sided limits.
At an interior point , we can also justify the ordinary series and identify its sum without assuming that mean-square convergence is pointwise convergence. Put . Since and , the geometric-sum formula bounds every partial sum of by . Summation by parts, using that decreases to zero, therefore proves convergence of , and hence of the displayed sine series at this fixed .
To identify its sum, use the Fejér kernel already derived above. Although the sawtooth is not continuous on the whole circle, it is bounded and continuous at this particular . The same near-and-far split proves : near use continuity at , and away from zero use the vanishing kernel mass and the bound on . Since the arithmetic means of a convergent sequence have the same limit as that sequence, the ordinary partial sums also tend to . Thus
For instance, setting leaves only odd terms and gives . This evaluation is justified by convergence at that point; it is not a termwise substitution into an unproved formal identity. At the geometric bound above fails because , but every sine term is zero there, so the endpoint still gives the midpoint zero rather than the left limit .
Worked example 1.1.10
(A pulse with adjustable width and position)
A pulse need not occupy half a period or be centered at the origin. Fix and . Let agree away from the endpoints with the -periodic indicator of the arc , interpreted modulo . Its endpoint values may be assigned arbitrarily. The fraction of a period for which the pulse is on is ; this is its duty cycle.
Since the coefficient integrand is periodic, integrate over any full period and substitute on the pulse interval. For ,
The mean must be calculated separately: the apparent division by is not a formula at . Pairing the positive and negative modes gives the real series
The midpoint theorem applies because is piecewise constant. Thus (2) converges to inside the pulse, to outside it, and to at either endpoint, regardless of the assigned endpoint value. It also converges in . These conclusions are distinct: the midpoint theorem does not claim uniform convergence. Indeed, no sequence of continuous partial sums can converge uniformly to this discontinuous pulse.
Parseval gives another check. Since almost everywhere, its normalized energy is . The mean mode contributes , while each positive mode has a negative partner of the same magnitude. Consequently
For example, leaves precisely the odd terms and recovers . The center changes the phases in (1), but not the magnitudes or energy. Width also matters: a mode vanishes exactly when is an integer multiple of ; there is no general rule that all even modes vanish.
Quick check
(Let and . Find , and , the series value at an endpoint, and the total energy outside the constant mode.)
Use (1) for the coefficients and for the energy.
Solution · Answer
Equation (1) gives , and . The endpoint value is . The normalized total energy is and the constant contributes , so the remaining modes contribute . This is the combined energy of positive and negative frequencies.
Worked example 1.1.11
(An energy calculation)
Return to the periodic sawtooth for from the earlier example, rather than the pulse just discussed. Its oddness gives ; pairing its sine coefficients gives for . Its normalized energy is . Its complex coefficients satisfy for . Parseval therefore gives , and hence . The equality uses completeness, not Bessel's inequality alone.
Worked example 1.1.12
(Changing one point changes no coefficient)
Set for except at , where , and extend periodically. Thus precisely at the points , . Every coefficient is zero, so every partial sum is zero. The series converges to zero at the exceptional point, not to the chosen value one. This does not contradict convergence: and zero represent the same element.
Why the midpoint appears
Theorem 1.1.13
(Dirichlet's pointwise theorem)
Let be -periodic and piecewise on a finite partition, with finite one-sided limits of both and at partition points. Then at every . In particular the limit equals at continuity points.
Proof
(Dirichlet kernel and the Riemann–Lebesgue lemma)
Fix and put . The Dirichlet kernel is even and has integral , so its convolution formula gives
The subtraction is legitimate because convolution with the kernel leaves a constant unchanged. Write . The constant terms in the two one-sided expansions cancel. Since the one-sided derivatives are bounded near , integrating on each side and taking the one-sided limit gives as (also for complex-valued ). This remains true when is a periodic seam: use the periodic extension on each side. Away from zero, is integrable and is bounded away from zero. Consequently belongs to .
The Riemann–Lebesgue lemma now makes the oscillatory integral tend to zero. To see the lemma needed here, approximate the integrable quotient in by step functions. Each interval integral of tends to zero by its explicit antiderivative, while the approximation error is at most the error because . This proves the limit without assuming that the original Fourier series can be evaluated term by term at the jump.
At a jump the limit is the midpoint of the one-sided limits. Near the jump the partial sums can overshoot even as grows. For a jump of magnitude under these hypotheses, the limiting overshoot is approximately ; its spatial width shrinks, but its relative height does not. This is the Gibbs phenomenon. It concerns behavior near a jump, whereas the midpoint theorem concerns the value at the jump itself.
Worked example 1.1.14
(A jump and an assigned value)
Let on and on , assign and give the seam any common value, then extend all these values periodically. Changing a finite set of values in one period does not change the coefficients. At the one-sided limits are and , so the partial sums tend to , not . At the limit is ; at the periodic seam the one-sided limits are and , so the limit is again .
Quick check
(A -periodic function satisfying the hypotheses of Dirichlet's theorem has limits from the left and from the right at , while its assigned value there is . Find .)
Decide first whether the assigned value can affect any coefficient.
Solution · Answer
The theorem gives . The assigned value does not enter the coefficient integrals.
A nonlinear operation creates new frequencies
A frequency description of a signal does not remain unchanged when we alter the signal pointwise. For example, clipping the negative half of a sine wave is a nonlinear operation: it produces a nonzero average and infinitely many cosine frequencies, even though the original sine has only one frequency.
Worked example 1.1.15
(Half-wave rectification)
Let , extended with period . The useful decomposition
separates an odd function from an even one. Put . Since is even, every sine coefficient vanishes. It also has period : translating its coefficient integral by shows that , so odd frequencies vanish. For , the product-to-sum identity gives
The identity used in the second line is . Thus
This equality holds at every , not just in mean square. Indeed the cosine coefficients are absolutely summable, so the series converges uniformly. Termwise integration identifies its coefficients with those of continuous ; completeness in makes the difference zero almost everywhere, and continuity then makes it zero everywhere. The corners of do not themselves make the Fourier series discontinuous.
At , the equality gives . This can also be checked without Fourier convergence: telescopes. The nonzero constant is the average height; forgetting the factor doubles it. The new frequencies arise because clipping is nonlinear, not because taking a Fourier series changes the function.
Quick check
(For , find its real Fourier series and normalized energy. Compare the new frequency with those of .)
A double-angle identity gives a finite series.
Solution · Answer
The identity is already a finite Fourier series. Its average is and orthogonality gives . Squaring introduces a constant and frequency ; it is another nonlinear operation, though this particular one produces only finitely many modes.
Common mistakes
Common mistake
(Reading the symbol as a convergence theorem)
Computing all coefficients does not establish pointwise, uniform, or absolute convergence. State the function class and the mode of convergence before invoking an appropriate theorem. In particular, Fejér means converge uniformly for continuous periodic functions; continuity alone does not even ensure that the unaveraged partial sums converge at every point: continuous periodic functions with a divergent Fourier series at a point exist.
Common mistake
(Doubling the constant term)
With , the constant term of the real series is , the average of . Writing instead doubles the mean: for it would give instead of , and for the half-wave rectifier instead of . The complex form has no such factor, because is already the average.
Counterexample mode
Fejér means of a discontinuous function
Claim: since Fejér means repair the partial sums, they converge uniformly for every bounded piecewise continuous function. Test the square wave with on and on . It is odd, so and every is zero; hence for every . Each is a trigonometric polynomial, hence continuous, so there is with while . Thus for every : the convergence is not uniform. The repair is the hypothesis of Fejér's theorem, continuity on the whole circle. Without it, the near-and-far argument of the sawtooth example still gives at each point where is continuous.
Summary
Orthogonality produces the coefficients and the best finite approximation. Positive Fejér kernels prove density, and density upgrades Bessel's inequality to Parseval and convergence. Pointwise convergence and jump behavior require additional hypotheses.
Exercises
Quick check
(Find the Fourier coefficients of and compute .)
Solution · Answer
The nonzero coefficients are , , and . Orthogonality gives .
Quick check
(For on and on , extended periodically, find the sine coefficients and the value predicted at by the midpoint theorem.)
Solution · Answer
Oddness makes the cosine coefficients zero and . The one-sided limits at zero are and , so the midpoint is zero.
Quick check
(If is continuous and periodic and all its Fourier coefficients vanish, prove everywhere.)
Solution · Answer
Every Fejér mean is zero because every coefficient is zero. Their uniform limit is by Fejér's theorem, so vanishes at every point. For a general representative one can conclude only that almost everywhere.
Quick check
(For the periodic extension of , compute and its exact squared error. Which theorem says no other degree-one trigonometric polynomial has smaller error?)
Solution · Answer
The coefficients above give . The finite energy identity gives . The orthogonal projection theorem gives the least error, with equality only for .
Quick check
(Let on , extended with period . Find and using origin as above. What does the series give at ? If the origin is changed to , how does change?)
Solution · Answer
With , the formulas give and . At the periodic one-sided limits are and , so the sum is . The coefficient with origin is ; its magnitude is unchanged.