A Fourier coefficient measures how much a periodic function correlates with one oscillation. Computing those coefficients is a finite integral calculation. Recovering the function from infinitely many of them is a convergence problem, and its answer depends on the summation method and the regularity of the function. Keeping those questions separate prevents several common misuses of a formal Fourier expansion.
Motivation
The functions provide increasingly fine oscillatory patterns on a circle. Orthogonality makes a finite best approximation easy to compute. It does not, by itself, guarantee that the resulting partial sums approach the original function at every point or uniformly.
There are three different notions in play. Pointwise convergence controls each fixed input with a potentially different index. Uniform convergence controls all inputs at once. Mean-square convergence controls an integral of the squared error and can tolerate errors concentrated on small sets. A fourth choice concerns how coefficients are summed: ordinary partial sums and averages of partial sums need not have the same convergence behavior.
This note proves a precise pointwise theorem for piecewise functions and a uniform theorem for averaged sums of continuous functions. It develops elementary consequences in the mean-square norm, without claiming a general theory of spaces or pointwise convergence for arbitrary continuous functions.
For a concrete separation of convergence modes, let on one period and extend it periodically. Its height at zero is always one, while its support shrinks to that point. Integrating the square gives , so it converges to zero in mean square. It does not converge to zero at the origin, and its uniform norm stays one. A statement about mean-square error therefore cannot certify the error at a prescribed point. Conversely, uniform error does control mean-square error for this normalized integral, because integrating a squared pointwise bound preserves the bound. These implications depend on the norm and measure being specified.
Definitions
We work with -periodic functions, identifying the endpoints of . For complex-valued square-integrable functions, use the normalized inner product
For the continuous and piecewise smooth functions considered below, ordinary Riemann integration suffices. When discussing general square-integrable functions, the integral is the Lebesgue integral and functions agreeing almost everywhere represent the same element; the construction of that space is outside this note. Values at finitely many points affect neither the integral nor the inner product.
Definition 4.1.1
(Fourier coefficients and partial sums)
For an integrable periodic function define
The Fejér mean is .
For real functions, the equivalent notation is , where and , with integrals taken over one period. The coefficient has the same integral normalization; its contribution is halved because the constant basis function has a different unnormalized squared norm.
Definition 4.1.2
(Piecewise periodic function)
We call a periodic function piecewise here if a finite partition of a period makes it on each open piece, with finite one-sided limits of both the function and its derivative at the endpoints. Assigned values at the finitely many joins may differ from their one-sided limits. This explicit convention is stronger than merely having a derivative except at finitely many points; the bounded-quotient argument below needs these one-sided controls.
Theorem/proposition
Proposition 4.1.3
(Orthogonal projection and Bessel inequality)
Suppose . The functions are orthonormal for the normalized inner product. The polynomial uniquely minimizes among trigonometric polynomials with frequencies . Moreover,
Theorem 4.1.4
(Fejér's uniform convergence theorem)
If is continuous and -periodic, then uniformly. In particular, trigonometric polynomials are uniformly dense in the continuous functions on the circle.
Theorem 4.1.5
(Pointwise convergence for piecewise smooth functions)
If is periodic and piecewise in the stated sense, then at every ,
At a point of continuity this is . At a jump it is the midpoint of the one-sided limits, regardless of the value assigned at the point itself.
Fejér's theorem concerns averaged sums, not ordinary partial sums. The pointwise theorem does not assert uniform convergence across jumps: a uniform limit of continuous trigonometric polynomials would be continuous. The kernel proof also gives quantitative error control once a modulus of continuity of the chosen function is known. Later theory gives stronger and more delicate results under other hypotheses; none should be read into these statements. Before using a conclusion, state the function class, the summation method and the mode of convergence together, rather than saying only that “the Fourier series converges.”
Proof sketch or proof idea
Integration gives for and one for . Consequently, is orthogonal to every polynomial with these frequencies. For any such ,
This proves optimality and uniqueness in the mean-square space. Taking proves the finite energy identity; increasing gives Bessel's inequality. Equality for the infinite sum requires showing that the residual tends to zero, which orthogonality alone has not yet established.
This distinction is not peculiar to Fourier series. Projection onto any finite orthonormal collection gives the same Pythagorean calculation. If the collection omits a direction, a nonzero residual perpendicular to all its members remains invisible to every coefficient. In the Fourier setting, the density argument is what rules out that possibility for the continuous functions considered here. Thus a proof of an infinite expansion has two separate tasks: compute the projection correctly, and show that the available subspaces become sufficiently rich. Doing only the first task proves Bessel's bound but does not prove Parseval equality or recovery of the original function.
Finite summation inside an integral gives the convolution formula
At the last quotient is defined by its limit . Summing a finite geometric series proves the formula away from zero. The kernel has normalized integral one, but changes sign. Thus it does not act like a simple positive weighted average.
Averaging the kernels yields
The identity follows either by summing the geometric series and its conjugate, or by counting the pairs producing each frequency in the square. Thus and . For ,
Proof
(A positive approximate identity gives uniform convergence)
Write
Uniform continuity on the circle supplies such that the bracket has magnitude less than for , uniformly in . Nonnegativity and normalized mass one bound this part of the integral by .
For the rest, the bracket is bounded by , and the displayed tail bound makes the normalized integral of outside tend to zero. For sufficiently large , this contribution is also less than . The bound is uniform in . The proof uses both positive mass control and concentration near zero; knowing the integral is one would not suffice for a signed kernel.
Proof X-ray
Order of choices in the Fejér proof
The proof fixes , then chooses from uniform continuity, and only then lets grow; depends on neither nor . The near part uses positivity and mass one, so it is an average of quantities smaller than . The far part uses only the size bound and the tail estimate: the normalized integral of over is at most , which tends to zero. Uniformity in comes from uniform continuity on the compact circle. For a function with a jump, no single controls the bracket near the jump, and the near-part estimate is lost.
For ordinary sums, symmetry of gives, with ,
The normalization follows from . Piecewise regularity makes the quotient bounded near zero: the numerator is by the finite one-sided derivative bounds, and is comparable to . Away from zero it is piecewise continuous and bounded, hence Riemann integrable.
For any Riemann integrable on a compact interval, as . Approximate in integral absolute error by a step function. Each step's oscillatory integral has magnitude at most times the step height; the error contributes at most its integral absolute value. First choose the approximation error, then let grow. Applying this elementary oscillatory-integral lemma proves the pointwise theorem.
There is also a useful mean-square consequence for continuous . Fejér gives a trigonometric polynomial uniformly close to , hence close in normalized mean square. For large enough to include its frequencies, projection optimality gives . Thus ordinary sums converge in mean square for continuous , and the finite energy identity yields Parseval equality for this class. This argument does not prove uniform convergence of ordinary sums.
Counterexample mode
Fejér's theorem needs continuity
Claim: the Fejér means of every bounded piecewise continuous periodic function converge uniformly to it. Take the sawtooth for , extended periodically. Its series contains only terms, so every and hence every vanishes at . Each is continuous, so as , while . Therefore for every . The repaired statement is the theorem as stated: uniform convergence requires to be continuous on the whole circle, including across the identified endpoints.
A reliable calculation therefore ends with a short interpretation. Identify the periodic extension, including any endpoint jump. State whether the theorem concerns the original partial sums or their averages. Check the regularity needed by that theorem, and only then substitute a convenient input into the resulting identity. Substituting into a formal series before justifying convergence can produce a correct-looking number without a valid argument. At a jump, the recovered midpoint is part of the theorem and must be used even if the function was assigned a different isolated value there.
Worked examples
Worked example 4.1.6
(1. A finite expansion checks the normalization)
For , the real coefficients are , , , and all remaining ones vanish. Therefore for . Orthogonality gives . The nonconstant sine and cosine terms have normalized squared norm , while the constant has norm one.
Worked example 4.1.7
(2. The sawtooth and its endpoint jump)
Take for and extend periodically. Oddness gives . Integration by parts gives . Thus the series is . It converges to for interior points and to zero at the identified endpoints, the midpoint of and . A value assigned at either endpoint cannot alter the coefficients. Uniform convergence to the sawtooth is impossible because of the jump.
Worked example 4.1.8
(3. A continuous corner and absolute convergence)
For on , periodic extension is continuous. Evenness gives , and . Hence
The coefficient absolute values are summable, so the Weierstrass comparison test gives uniform convergence of these ordinary sums. The pointwise theorem identifies their limit as . At zero the formula gives . A corner obstructs differentiability but does not prevent this uniform convergence.
Half-range sine and cosine expansions
Sometimes data are given only on . The Fourier machinery above expects a periodic function on a full period, so first decide how to extend the data to . Reflecting evenly produces a cosine series; reflecting oddly produces a sine series. These extensions can be different periodic functions even though they agree on the original interior. Their endpoint behavior need not agree.
Suppose is piecewise on , with finite one-sided endpoint limits. Define the even extension by for , and the odd extension by for and for . Set and , and extend each function -periodically. Other values assigned at the reflection points do not affect the coefficients, although the odd extension would no longer be literally odd. By symmetry, the odd integrands integrate to zero, while the even integrands contribute twice their integral over . Thus
and
The symbol records coefficients; it does not by itself establish equality. The piecewise smooth convergence theorem above says that each series converges to the midpoint of the corresponding extension's one-sided limits. At an interior continuity point this is . At and , the cosine series tends to the respective one-sided limits of , whereas the sine series is identically zero at both endpoints. The odd extension has a jump there when the corresponding one-sided limit is nonzero. Changing a single assigned endpoint value cannot change the coefficients or repair that jump.
Worked example 4.1.9
(Two expansions of the same function)
Take on . Integration by parts gives
Also . Consequently, for ,
The sine formula gives zero at , not : the odd periodic extension jumps from to there. It cannot converge uniformly to on the closed interval, since its partial sums all vanish at . The cosine coefficients are absolutely summable, so the cosine series converges uniformly; the pointwise theorem identifies its sum with throughout . The faster coefficient decay here belongs to this particular extension and example, not to every cosine series.
Quick check
(Expand on using each reflection. What does each series give at the endpoints?)
Solution · Answer
The even extension is the constant function, so its cosine series is , including at the endpoints. For the odd extension, , giving . The pointwise theorem gives on and at both endpoints, where the odd extension jumps between and . In particular, the sine series does not converge uniformly to on the closed interval.
Common mistakes
Margin Note
Finish a coefficient calculation with a convergence argument
Never infer pointwise recovery solely from a coefficient calculation. State the theorem and verify its hypotheses. In particular, the pointwise theorem uses the periodic one-sided limits at an endpoint, not two copies of the same interior limit.
Common mistake
(Coefficients cannot see isolated values)
Changing finitely many function values changes no coefficient. Therefore Fourier coefficients cannot determine those assigned values without a convention or a continuity assumption. For continuous functions they do determine the whole function, because Fejér means recover it uniformly.
Common mistake
(Decay like is not absolute convergence)
The sawtooth coefficients are not absolutely summable, since diverges; the series converges pointwise only through cancellation. Summable coefficient magnitudes do give uniform convergence, but that sufficient condition is stronger than necessary and fails for every function with a jump.
Common mistake
(Averaging changes the summation method)
Fejér's theorem proves uniform convergence of the averages . It is not a proof that the original partial sums converge uniformly, or even pointwise, for a continuous ; ordinary sums need separate hypotheses such as the piecewise condition.
No: some continuous periodic functions have Fourier series that diverge at a point, so Fejér averaging is not a technicality.
A natural next question
Do the ordinary partial sums of every continuous function converge?
No: some continuous periodic functions have Fourier series that diverge at a point, so Fejér averaging is not a technicality.
The signed kernel is the obstruction. On continuous functions with the uniform norm, the functional has norm , and . Since these norms are unbounded, the uniform boundedness principle of functional analysis yields a continuous with . The positive kernel has normalized norm exactly one, so the averaged sums escape this mechanism.
Summary
Orthogonality identifies finite Fourier sums as mean-square projections. Kernel formulas turn convergence into an integral estimate. The positive Fejér kernel gives uniform recovery of continuous periodic functions. The signed Dirichlet kernel requires extra cancellation, which piecewise regularity supplies for pointwise recovery. Always record both the convergence mode and the summation method.
Exercises
Quick check
(Compute the Fourier coefficients of and its normalized squared norm.)
Solution · Answer
The real coefficients are , , and all other coefficients, including , are zero. In complex notation, , and . The squared norm is .
Quick check
(Suppose a continuous periodic function has all Fourier coefficients zero. Prove that it vanishes everywhere.)
Solution · Answer
Every partial sum and hence every Fejér mean is identically zero. Fejér's theorem says these means converge uniformly to , so . Continuity is essential for an everywhere statement: a function supported at one point has all coefficients zero without vanishing at that point.
Quick check
(For a trigonometric polynomial , show that need not equal even when . Find its coefficients.)
Solution · Answer
For , frequency appears in exactly when , hence in of the averaged sums. Therefore . Nonzero frequencies are damped, so equality at finite generally fails. Their weights tend to one as .
Quick check
(Let be integrable and suppose . Prove uniform convergence of its Fourier series to a continuous function . If is also piecewise , identify and explain the constraint on jumps.)
Solution · Answer
The bound gives uniform absolute convergence, and the uniform limit is continuous. For piecewise , the pointwise theorem gives . At nearby continuity points on either side of a proposed jump, . Continuity of then forces the two one-sided limits to agree. Thus genuine jumps are impossible under both hypotheses; isolated assigned values may still differ from the common limit.