Fourier coefficients are defined by integrating over an entire period, yet convergence at one point can depend only on the function near that point. The reason is cancellation: away from the point, the Dirichlet kernel oscillates faster and faster. Near the point, its singular-looking denominator must be controlled by a local hypothesis.
This chapter proves localization, measures the size of the partial-sum operator, and shows why continuity alone cannot guarantee pointwise convergence. Fejér means behave better, and the distinction between the two summation methods is essential.
The Dirichlet kernel
Definition 1.4.1
(Partial sums and the Dirichlet kernel)
For , extended periodically, write
The finite geometric sum gives
The quotient has its continuous limiting value at multiples of . Finite interchange of sum and integral yields
The kernel is even but changes sign. Its integral is one; its absolute integral is not uniformly bounded. Thus it is not a positive approximate identity like the Fejér kernel.
Oscillatory cancellation
Theorem 1.4.2
(Riemann–Lebesgue lemma on a finite interval)
If , then
The frequency may be real, not only an integer. The same conclusion holds with sine or cosine in place of the exponential.
Proof
(Approximate first, then oscillate)
For an interval indicator , direct integration gives an absolute bound . The same holds, with a constant depending on the function, for any finite linear combination of interval indicators. Such functions are dense in : approximate measurable sets by finite unions of intervals in measure using Lebesgue regularity, then approximate simple functions and finally .
Given , choose an interval-step function with . Since ,
For this fixed the remaining integral tends to zero. First send to infinity, then to zero. The identities and give the sine and cosine statements even for complex .
Localization at a point
Theorem 1.4.3
(Localization of partial-sum limits)
If two periodic functions agree almost everywhere in a neighborhood of on the circle, then
In particular, if either sequence of partial sums converges at , so does the other, to the same limit. Their values away from the neighborhood may be completely different.
Proof
(Discard the remote part by oscillation)
Set . Choose such that almost everywhere for . The partial-sum integral becomes
On this set is bounded, so the coefficient of the sine is integrable. Apply the Riemann–Lebesgue lemma separately on the two intervals. This is a statement about a fixed point and fixed functions; it does not by itself give a uniform rate over a family.
Common mistake
(Reading localization as a statement about each partial sum)
Localization concerns limits, not individual partial sums. If on , extended periodically, then vanishes near , yet . Every finite partial sum at depends on all of ; only the difference between two functions that agree near tends to zero.
Local agreement and a jump
Worked example 1.4.4
(A remote modification cannot change the limit)
Suppose equals almost everywhere near zero but is otherwise any periodic integrable function. Compare it with the continuous periodic extension of from . The latter is piecewise smooth, so its partial sums at zero converge to zero by the pointwise convergence theorem in the earlier notes. Localization gives the same limit for , regardless of its values elsewhere. In particular, changing does not alter any coefficient or the limit.
Worked example 1.4.5
(The midpoint is a local statement)
For the square wave on and on , for every by oddness. Any periodic integrable function agreeing almost everywhere with near zero has the same partial-sum limit at zero, even if it differs far away. The assigned value at the jump is irrelevant.
The Gibbs chapter shows that nearby peaks can nevertheless persist. A pointwise midpoint limit does not imply uniform convergence on a neighborhood of the jump.
Measuring the size of the partial-sum operator
Definition 1.4.6
(Lebesgue constant and best approximation error)
Define the Lebesgue constant
Let be the trigonometric polynomials of degree at most , and set for continuous periodic .
Theorem 1.4.7
(Logarithmic growth and a best-approximation bound)
There are positive constants , independent of , such that for every ,
Moreover, for every continuous periodic ,
Consequently, is sufficient for uniform convergence of to . Density of trigonometric polynomials alone gives only and does not ensure that stronger rate.
Proof
(Norm growth and approximation error)
The lower bound follows from the oscillation-interval calculation below. For the upper bound, by its finite-sum definition, and for ,
Split the positive half-period at : the first part has integral at most , and the second at most . Evenness gives the upper logarithmic estimate after the normalizing factor.
For any , orthogonality gives . The convolution formula yields , hence
Taking the infimum over proves the bound. This is a sufficient estimate, not a necessary convergence criterion: cancellation in a particular function may be much better than the worst-case operator norm.
Proof X-ray
Approximation error and amplification
The approximation bound uses two facts about of different kinds. The algebraic one, for by orthogonality, is the key move: it rewrites the error as the remainder plus applied to , so the only source of error is the remainder and its amplification. The analytic one, , bounds that amplification using nothing but the absolute integral of the kernel. Taking the infimum over closes the argument. Nothing about is used beyond , which is why the estimate is sufficient but not necessary.
The upper estimate for splits at , where the trivial bound and the decay bound cross up to constants. The inner region contributes a bounded amount; the logarithm comes entirely from , because the kernel decays only like . The matching lower bound in the next section shows that this amplification is real, which is what the uniform boundedness argument exploits.
Worked example 1.4.8
(The first Lebesgue constant)
For , , which is positive for and negative for . With the antiderivative ,
By evenness, . The signed integral of is one, so the excess is caused entirely by the negative lobes. The value is approached by for continuous with that approximate the sign of , which is the norm computation used in the counterexample below; already for the partial-sum operator can enlarge the sup norm.
Why continuity alone is insufficient
Counterexample mode
A continuous function with unbounded partial sums at one point
Work in the real Banach space of continuous real -periodic functions with the uniform norm. For each , define the bounded linear functional . Changing variables and using evenness of the real kernel gives . Thus
The upper bound follows immediately from the integral formula. For the reverse bound, the continuous functions have sup norm at most one, and pointwise and is dominated by . Dominated convergence gives equality of the norm.
These norms grow without bound. Put . For , . On each interval with ,
These intervals lie in . Combining with evenness and normalization gives the explicit lower bound . Since the sum is at least , it dominates a positive multiple of for . Thus . The uniform boundedness principle now gives a continuous periodic real function for which . This is an existence proof, not an explicit formula for .
For the same function, Fejér means converge uniformly to by Fejér's theorem. Averaging the partial sums can restore convergence even when the unaveraged values at a point are unbounded.
Common mistake
(Concluding that every continuous function diverges somewhere)
Unbounded norms give the existence of some continuous function with ; they say nothing about a given function. Every trigonometric polynomial has for large , and every function with converges uniformly by the theorem above. Divergence is a property of particular functions that the operator-norm argument shows to exist.
Exercises
Quick check
(Let be periodic and integrable and equal to almost everywhere near . Prove that even if is assigned a different value.)
Solution · Answer
Compare with the constant function using localization. Every partial sum of that constant equals . A single value does not change any coefficient.
Quick check
(Let be the periodic extension of for . Find the partial-sum limits at and at the seam by comparison with locally identical odd functions.)
Solution · Answer
At zero, is odd and its partial sums are identically zero there. Periodicity also makes almost everywhere, so it is odd about the seam; its partial sums at are zero for every . This agrees with the midpoint from the earlier piecewise-smooth convergence theorem.
Quick check
(If a continuous periodic function satisfies for all sufficiently large , prove uniform convergence of to . Does the estimate alone prove this through the displayed bound?)
Solution · Answer
The theorem gives . The weaker assertion is false: every continuous periodic function has by trigonometric approximation, yet the counterexample above has unbounded partial sums at zero. The displayed error bound also cannot force convergence without a suitable rate.
Quick check
(Suppose is integrable and odd on the circle. Find for every . Does this settle convergence at other points?)
Solution · Answer
The even Dirichlet kernel times an odd function has integral zero over a symmetric period, so for every . This says nothing by itself about convergence at other points.